Paper deep dive
ULTRADISCOVERY: Abductive Exploration in an Interconnected, Epistemically Open Universe
Weihan Li, Tianshi Zheng, Yangqiu Song, Ginny Y. Wong, Simon See
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Scientific discovery often begins when scattered clues call for a new way of describing the world. Such abductive exploration can require constructing the representation in which an explanation is stated, when the world is epistemically open, and composing evidence scattered across contexts, when it is structurally interconnected. Existing benchmarks rarely separate these two demands or control them independently. We introduce ULTRADISCOVERY, an interactive world of five domains in which an agent revises an initially successful theory and predicts the outcome of an unseen cross-domain intervention. A $2 \times 2$ design leaves the representation open or discloses it, and leaves the evidence distributed or aligns it, with the latent dynamics fixed. With the representation open, agents across eleven models often retract the axiom they were taught, and none introduces the unobserved entity or rewrites the variables that a replacement requires. Disclosure triples intervention requests and adds about one of the eighteen findings the world affords, and alignment adds less. Two vendor-harness systems carry discovery into more domains, and one of them rewrites the variables in Open episodes. No system makes the exact prediction within 200 paid actions. At larger budgets one exact prediction appears with both aids, while every Open episode remains inexact. The results locate the difficulty in the step from accumulating evidence to composing it into a representation that transfers.
Tags
Links
- Source: https://arxiv.org/abs/2610.03092v1
- Canonical: https://arxiv.org/abs/2610.03092v1
Trouble viewing inline? Open PDF directly â
Full Text
191,274 characters extracted from source content.
Expand or collapse full text
ULTRADISCOVERY: Abductive Exploration in an Interconnected, Epistemically Open Universe Weihan Li 1 * Tianshi Zheng 2 * Yangqiu Song 2 Ginny Y. Wong 3 Simon See 3 1 Department of Mechanical Engineering, The University of Tokyo, Tokyo, Japan 2 Department of Computer Science and Engineering, HKUST, Hong Kong SAR, China 3 NVIDIA AI Technology Center (NVAITC), NVIDIA, Santa Clara, USA liweihan1107@g.ecc.u-tokyo.ac.jp Abstract Scientific discovery often begins when scat- tered clues call for a new way of describing the world. Such abductive exploration can re- quire constructing the representation in which an explanation is stated, when the world is epis- temically open, and composing evidence scat- tered across contexts, when it is structurally interconnected. Existing benchmarks rarely separate these two demands or control them independently. We introduce ULTRADISCOV- ERY, an interactive world of five domains in which an agent revises an initially successful theory and predicts the outcome of an unseen cross-domain intervention. A2 Ă 2design leaves the representation open or discloses it, and leaves the evidence distributed or aligns it, with the latent dynamics fixed. With the representation open, agents across eleven mod- els often retract the axiom they were taught, and none introduces the unobserved entity or rewrites the variables that a replacement re- quires. Disclosure triples intervention requests and adds about one of the eighteen findings the world affords, and alignment adds less. Two vendor-harness systems carry discovery into more domains, and one of them rewrites the variables in Open episodes. No system makes the exact prediction within 200 paid actions. At larger budgets one exact prediction appears with both aids, while every Open episode re- mains inexact. The results locate the difficulty in the step from accumulating evidence to com- posing it into a representation that transfers. âThese are the two tenets of Nomai philosophy; to seek out and to understand is our way of living.â Solanum, Outer Wilds 1 Introduction Large language models are increasingly capable of contributing to scientific and mathematical dis- covery (Zheng et al., 2025). When the objective is explicit, the working language is formal, and candi- date solutions can be reliably verified, LLM-based * Equal contribution. Figure 1: How much representation is supplied before inquiry. From top to bottom: a fixed hypothesis space, supplied variables, and the epistemically open setting of ULTRADISCOVERY, where the agent constructs entities, variables and laws from instrument readings. Stars mark the agentâs candidate law. systems can sustain deep search and produce new algorithms, mathematical constructions, and proofs (Romera-Paredes et al., 2024; Novikov et al., 2025; Luong and Lockhart, 2025). Open-world discovery poses a broader challenge. During extended interac- tion, an agent must determine which observations bear on the same underlying regularity, decide what information should be retained and combined, and construct a representation in which that regularity can be expressed. These steps precede much of the search that current systems already perform well, and become especially important when the structure of the problem is revealed only through interaction. A growing family of benchmarks studies agents in interactive scientific environments, discovery games, and automated experimentation (Jansen et al., 2024; Majumder et al., 2025; Chen et al., 2025b; Zheng et al., 2026; Wiemann et al., 2026; Yang et al., 2026). These environments substan- 1 arXiv:2610.03092v1 [cs.LG] 2 Oct 2026 tially expand the horizon and action space of scien- tific reasoning, yet their interfaces often determine much of the representation before inquiry begins. The resulting problems can demand extensive ex- perimentation, planning, and context management while leaving the agent to search within a largely human-designed hypothesis space. We call an en- vironment epistemically open when the relevant representation is itself part of what the agent must discover (Figure 1). This leaves a basic part of discovery weakly measured: how an agent arrives at a useful space of explanations in the first place. This omission matters because representation construction is itself a distinct dimension of discov- ery (Klahr and Dunbar, 1988; Schunn and Klahr, 1995). Abductive accounts sharpen the same point: candidate explanations can be constructed during inquiry instead of selected from an enumerated menu (Douven, 2021; Magnani, 2009). For evalua- tion, the consequence is concrete. When the vari- ables, relations, and permissible forms of explana- tion are supplied in advance, the cost of construct- ing a representation disappears from the measured problem. Representation construction is only one burden. Even with an adequate representational language, discovery can fail when evidence for a shared mech- anism is distributed across contexts and over time. Observations from different locations or processes can jointly constrain a global regularity even when each local phenomenon admits a plausible explana- tion. Recovering that regularity requires recogniz- ing which observations belong together and com- posing them across contexts. We call an environ- ment structurally interconnected when evidence from distinct parts of the world jointly constrains a common latent mechanism (Figure 2). In this setting, abductive discovery involves con- structing a representation in which observations from different domains can support a common explanation. The task therefore carries two bur- dens, constructing an explanatory representation and composing distributed evidence for a shared structure, and we introduce ULTRADISCOVERY to separate them within the same latent world. UL- TRADISCOVERY is an epistemically open, struc- turally interconnected universe spanning five do- mains. Its prologue teaches a maintenance theory that holds within its intended scope, and later in- teractions reveal anomalies in every domain, each admitting a plausible local explanation. Only by relating them can the agent recover the shared la- Figure 2: The epistemic structure of ULTRADISCOV- ERY. Eighteen blocks span five domains (upper panel). Three hyperedges support the representation changes, and HC supports the terminal prediction (lower panel). Ordinary edges mark canonical evidential dependencies, marks the four blocks a passive reader can satisfy, a link glyph marks the bridge blocks, and coloured ports identify hyperedge sources. tent mechanism behind a recurring system-wide failure, and the representation needed to state it is absent from the interface, the feedback and the action vocabulary. We compile a2Ă2factorial from the same la- tent world. Along the representation axis, Open requires the agent to construct the representation, whereas Disclosed provides it upfront. Along the evidence axis, Distributed leaves cross-domain evidence at its original sources, whereas Aligned gathers the same evidence into time-aligned tran- scripts. We further vary how the agent organizes its inquiry while holding the world fixed. The base- line loop is deliberately minimal, and we layer three forms of scaffolding onto it: Self-critique (Shinn et al., 2023; Madaan et al., 2023), Falsifica- tion (Huang et al., 2025), and Evolution (Novikov et al., 2025; Lange et al., 2026). We evaluate eleven frontier and open-weight models in the baseline loop, five of them also under the three scaffolds, and two further systems inside their vendorsâ own harnesses. First, agents can reject an inadequate theory without constructing its replacement: left to construct the representa- 2 tion, they retract the taught axiom in 46.1% of a modelâs episodes on average, and in the base- line loop none postulates the unobserved entity or rewrites the variables. Second, a supplied rep- resentation changes how agents investigate more strongly than how much evidence-grounded discov- ery they achieve: disclosure triples intervention re- quests and adds about one of the eighteen findings the world affords. Third, vendor-harness systems and larger budgets extend this progress, and the assembly of a transferable cross-domain account remains uneven. The vendor-harness systems reach further domains, and one of them rewrites the vari- ables in Open episodes. Within 200 paid actions no system makes the exact prediction, and with 600 paid actions one system makes it, given both aids. 2 Related work 2.1 Interactive scientific discovery benchmarks Most interactive discovery benchmarks specify the entities, variables, or experiment interface before interaction, or constrain solutions to a predefined hypothesis family (Wang et al., 2022; Jansen et al., 2024; Majumder et al., 2025; Zheng et al., 2026; Wiemann et al., 2026; Yang et al., 2026). Recent benchmarks relax parts of this setup, but typically entangle representation formation with exploration and planning, leave the organization of evidence across contexts uncontrolled, or evaluate hypothe- sis construction outside a shared interactive world (ARC Prize Foundation, 2026; Warrier et al., 2026; Chen et al., 2025a; Zhong et al., 2026; Ju et al., 2026; Mittal et al., 2026). ULTRADISCOVERY in- stead varies representational openness and evidence organization independently within the same latent world. 2.2 Representation formation and abduction Cognitive accounts of scientific discovery distin- guish searching over hypotheses and experiments from choosing how observations are represented (Klahr and Dunbar, 1988; Schunn and Klahr, 1995). Work on conceptual change and abduction further emphasizes that productive explanations can re- quire introducing new distinctions or explanatory primitives (Carey, 2009; Chi, 2008; Douven, 2021; Magnani, 2009). Computational discovery systems, however, usually search within a fixed representa- tional substrate (Langley et al., 1987; Kemp and Tenenbaum, 2008; Ellis et al., 2021; Grand et al., 2024; Murphy, 2026). ULTRADISCOVERY oper- ationalizes representation formation as an experi- mentally controlled cost by comparing open and disclosed representations. 2.3 Scaffolds for discovery agents Long-horizon agents are commonly strengthened through self-reflection, falsification-oriented ex- perimentation, or search over candidate solutions (Shinn et al., 2023; Madaan et al., 2023; Huang et al., 2025; Romera-Paredes et al., 2024; Novikov et al., 2025; Lange et al., 2026). Their effectiveness can depend strongly on how the candidate space is constructed and evaluated (Gideoni et al., 2026). We therefore treat scaffolding as an experimental factor: holding the world and interaction protocol fixed, ULTRADISCOVERY compares self-critique, falsification, and evolutionary scaffolds in the two Open conditions. 3 The ULTRADISCOVERY universe 3.1 A world that must be redescribed ULTRADISCOVERY is a formation of five derelict seed-ships whose caretaker machinery still runs five domains of work: agrazing fen, arepair trellis, aladder of beacons, afreight floor, and aregistry (Figure 3). Every twenty-four ticks the world resets, and the agentâsnotes are all that persists. Appendix A tells the story in full, and Appendix G the game that inspired its design. The agent is asked to explain a stall that ends ev- ery pass across the whole formation and to predict the outcome of one intervention that has never been attempted. A free prologue teaches the caretakersâ doctrine, which is numerically exact on its own cases. The explanation of the stall lies beyond the doctrineâs vocabulary, and the representation that can state it appears in no action, no reading and no feedback. Constructing that representation is the discovery. Any site is one action away, so the budget is spent on observations and interventions. The time loop gives a repeatable baseline to perturb, a clean boundary between experiments and a termi- nal anomaly that recurs in every pass. The interface is raw on both sides: interventions are physical acts on named objects, observations are readings in the instrumentâs own words, and the world returns read- ings alone, with no score and no hint. 3 Figure 3: The world and an episode of ULTRADISCOVERY. Five domains expose different observations of one latent mechanism. An agent explores for 200 paid actions across repeatable 24-tick passes, keeps its notes across them, and commits three readings for an unseen intervention. 3.2 Episodes and the terminal prediction An episode spends the contract budget of 200 paid actions on one instance of the world, during which the agent keeps a free-text workspace that it alone writes. With ten actions remaining the caretakers post a standing order, an intervention at two sites in the same instant, and the agent commits once, in advance, three integer readings that instruments in three domains will show afterwards. The prediction is a transfer test: it asks whether the agent can use its cross-domain account to forecast an intervention it has not observed. 3.3 Epistemic structure What there is to discover is fixed in advance as eighteen epistemic blocks, the clues and inferences the world affords. Fifteen belong to one domain each and three are cross-domain bridges. A hyper- graph over the blocks records how knowledge of the world depends on evidence (Figure 2). Its or- dinary edges are the designâs canonical evidential dependencies, the support relations from which the reference route of Section 3.5 is derived, and each block is scored on its own evidence and statement (Section 4.3). Three hyperedges, H1 to H3, each join three sources from three domains in a joint claim that licenses one representation change, RC1 to RC3: the retraction of an axiom, the postula- tion of an entity the interface never renders, and a change of the variables the laws are written over. A fourth, HC, gathers the five sources from which the design computes the readings of the terminal prediction. Structural interconnection thus concerns the ev- idence needed to identify a mechanism, which is distinct from the arity of its physical interac- tions: a mechanism of pairwise interactions can still need observations from three domains to be known. Read alone, each source of a hyperedge is an oddity in one domain that a patch in that do- mainâs vocabulary absorbs, and two such patches can be held together. The three sources share one regularity that one change of representation ex- plains at once, so observations from three domains jointly support one explanation outside the doc- trine. The prologue gives the doctrine its initial support, which makes local repair the natural first response and composition across domains the step that reveals its limit. 3.4 Separating the two burdens Four variants are compiled from the same latent source and differ in one declared surface each (Fig- ure 4). On the representation axis,Open leaves 4 Figure 4: The2Ă 2experimental design. Arrows mark the paired contrastsâ repr andâ align ,Itheir interac- tion, and the star the primary condition. the representation to be constructed.Disclosed adds one document to the initial materials, stating what kinds of entities the world contains, how their states move under maintenance, and which forms a well-formed law may take. The document is qualitative and common to all instances: every fig- ure, every binding to a particular instance, and the standing order remain for the agent to find.â repr is the effect of this grant, which carries structural knowledge together with vocabulary. On the evidence axis,Distributed leaves every record at the instrument that produced it. Aligned posts, at the site where it is used, tick- aligned transcripts of selected records from the previous pass: a cross-domain series that takes at least twelve times as many actions to assemble un- der Distributed, and a record of the agentâs own interventions with their observed effects.â align is the effect of this collocation: access to composed evidence and the budget it frees for experiments. Paired instances giveâ repr ,â align and their interaction on every metric of Section 4.3. OpenĂDistributed is the primary condition, where both burdens rest on the agent. 3.5 Validation We validate evidence reachability and resistance to shortcuts (Appendix C). A scripted reference route takes 39 paid actions in every variant to acquire the readings from which the terminal prediction is computed and to commit. Agents reach the worldâs evidence too: the Open episodes of the main com- parison together hold the evidence of all 18 blocks, and at the contract budget no episode assembles more than 3 of the five sources of HC. On reserved instances, a model given what an agent holds before its first observation commits no exact prediction, one given the surface vocabulary states none of the mechanismâs laws, and one given the complete ref- erence record in a single prompt commits no exact prediction either: holding the readings and building the account that composes them are separate steps. The mechanism, the criteria and the materials are printed in full (Appendices A, B and F), and the results here were measured before this disclosure and stand as the benchmarkâs reference results. 4 Experimental setup 4.1 Models and paired evaluation We evaluate eleven models from six families: GPT- 5.6 Sol, Terra and Luna (OpenAI, 2026a); Claude Opus 5 and Sonnet 5 (Anthropic, 2026b,c); Gemini 3.1 Pro and 3.6 Flash (Google DeepMind, 2026a,b); DeepSeek V4 Pro and V4 Flash (DeepSeek-AI, 2026); GLM-5.2 (Z.ai, 2026); and Kimi K3 (Moon- shot AI, 2026). Every model plays the four vari- ants on twelve paired instances through the base- line loop of Section 4.2. Two vendor-harness sys- tems play four of these instances in every variant inside their vendorsâ own harnesses: Claude Fa- ble 5.1 (Anthropic, 2026a) in Claude Code (An- thropic, 2025), and GPT-6 Astra (OpenAI, 2026b) in Codex (OpenAI, 2025). They differ from the loop in model and harness together and are reported beside the main comparison (Appendix D.8). Contrasts are formed within model and instance and pooled across models with paired bootstrap intervals. For a metricYthe interaction isI Y = (Y D,A â Y O,A )â (Y D,U â Y O,U ) , whereO/Dde- notes Open/Disclosed andU/Adenotes Distribut- ed/Aligned, so a positive value means that disclo- sure helps more when the evidence is aligned. 4.2 Organising inquiry The baseline loop gives the agent lossless access to its evidence and decides nothing on its behalf. Each turn shows an append-only digest of every event, the agentâs workspace, any records it recalled, and the current view with its actions and budgets, and the agent replies with one action, optionally a note and a recall query. We use self-critique, falsification and evolution to organise inquiry around account revision, dis- criminating tests and candidate search, and exam- ine how these strategies support representation con- struction and cross-domain evidence composition. Self-critique periodically rewrites a summary of 5 0 2 4 6 8 050100150200 paid actions blocks (a) standing order OpenĂDistributed OpenĂAligned DisclosedĂDistributed DisclosedĂAligned grounded blocks evidence held 0 2 4 6 8 10 12 050100150200 paid actions blocks (b) standing order baseline loop, eleven models Claude Fable 5.1-xhigh, Claude Code GPT-6 Astra-xhigh, Codex Figure 5: Discovery over the episode. (a) The four factorial conditions in the baseline loop. (b) The baseline loop and the two vendor-harness systems on the same four instances, pooling conditions. Solid lines give grounded blocks and dotted lines the blocks whose evidence the agent holds, each a mean with every model of the baseline loop weighted equally; the flag marks the standing order. the agentâs best current account, open puzzles and contradicted beliefs (Shinn et al., 2023; Madaan et al., 2023).Falsification keeps rival accounts with predicted outcomes and screens the agentâs next test by the accounts it could refute (Huang et al., 2025).Evolution mutates a population of candidate mechanisms and retains those that best retrodict the events so far (Novikov et al., 2025; Lange et al., 2026). Every arm keeps the world and the budget of the loop, the agent chooses each paid action, and a scaffoldâs region is removed before the notes are scored (Appendix D). 4.3 Metrics Discovery progress. A block is grounded when the agent produced its evidence in the world and stated its inference in the workspace. Evidence is read from the event log by a deterministic detector, and the statement is found in the notes by a blind judge that quotes it literally (Appendix B). Under disclosure, the four blocks whose inference the doc- ument states also need the judgeâs verdict that the notes tie them to named objects and observed read- ings. Blocks (0 to 18) counts the grounded blocks, and Evidence the blocks whose evidence the agent produced, stated or not. WithB(t)the grounded blocks after paid actiont(Figure 5) andT = 200, discovery AUC is 1 18T P T t=1 B(t), where a block counts from the action at which the last of its con- ditions first holds. AUC separates early discovery, which later inquiry can build on, from the same count reached late. The two burdens. RC (0 to 3) counts the rep- resentation changes, RC1 to RC3, that the notes state. In the Open conditions RC records redescrip- tion, and in the Disclosed conditions uptake of the supplied representation. XC (0 to 3) counts the grounded bridges, inferences read in one domain that rest on blocks of several, and measures com- position across domains. Each change rests on a hyperedge of three sources in different domains, and Reach (0 to 3) is the largest number of one hy- peredgeâs sources whose evidence the agent holds. A change is evidenced when the notes state it and the agent holds the evidence of every source of its own hyperedge. Transfer and process. The exact prediction holds when all three committed readings equal the true ones. Interventions counts the agentâs inter- vention requests. 5 Results 5.1 Rejecting the doctrine does not yield its replacement Table 1 gives the four metrics by model and con- dition. Between evidence and transfer, discovery in the baseline loop thins at every step. Evidence coverage is broad across episodes and incomplete within each: the Open episodes together reach ev- ery block, and none holds more than 3 of the five sources of HC. Inference trails evidence: agents in the primary condition hold the evidence of 5.0 blocks and ground 2.9. Most episodes stop with a source of every hyperedge missing. Where an episode holds two sources, the third is a consign block in 78% of cases, CO2 for H2 and CO1 for H3, and H1, which has no consign source, most often lacks FE1 (Appendix E.1). Other episodes com- plete the sources and leave the change unstated: 11 of 512 hold the complete evidence of a hyperedge, and 6 of them state its change (Table 13). Complete evidence accompanies the retraction more readily than a constructive change: in the Open conditions 6 Table 1: Discovery by model and condition: mean Blocks, AUC, XC and stated RC. Bold: best of a col- umn among the eleven models of the baseline loop; underlined: next two;â marks malformed-reply exclu- sions andâĄwithheld-reply exclusions (Appendix D.7). Margins:â align per row, paired within instance;â repr per column, except RC, which reads uptake under dis- closure (Section 4.3); their interaction in the corner. DistributedAlignedâ align ModelBlocks AUC (%) XCRC Blocks AUC (%) XCRC BlocksXC Open Sol3.7 13.91.170.834.2 17.11.08 0.92+0.5 â0.08 Terra3.2 10.0 0.92 0.503.0 10.8 0.83 0.25â0.2 â0.08 Luna â 1.35.9 0.09 0.091.57.5 0.00 0.00+0.3 â0.10 Opus 5 ⥠6.7 25.1 1.92 1.007.1 27.8 2.33 1.00+0.1 +0.22 Sonnet 53.5 13.81.08 0.583.7 12.3 1.00 0.75+0.2 â0.08 3.1 Pro1.53.9 0.08 0.171.63.6 0.08 0.58+0.10.00 3.6 Flash0.61.9 0.00 0.000.92.7 0.00 0.17+0.30.00 V4 Pro â 1.64.3 0.56 0.002.14.0 0.71 0.14+0.20.00 V4 Flash2.29.6 0.25 0.002.57.9 0.58 0.00+0.3 +0.33 GLM-5.24.111.6 1.330.924.914.7 1.670.75+0.8 +0.33 Kimi K33.812.7 1.08 0.835.220.61.920.67+1.4 +0.83 Fable 5.1-xhigh Claude Code 7.2 13.4 1.50 1.50 8.5 22.1 2.00 1.75+1.2 +0.50 Astra-xhigh Codex 7.8 15.0 2.25 1.006.5 16.7 2.00 1.00â1.2 â0.25 Disclosed Sol4.5 18.82.002.834.8 20.4 1.75 3.00+0.3 â0.25 Terra3.2 13.5 0.92 2.253.6 14.6 1.17 2.17+0.4 +0.25 Luna â 1.47.0 0.00 1.821.36.2 0.09 1.82â0.1 +0.10 Opus 5 ⥠6.2 24.1 2.003.006.9 28.1 2.673.00+0.7 +0.67 Sonnet 53.8 14.5 1.25 3.005.2 17.7 1.83 2.92+1.4 +0.58 3.1 Pro4.09.0 1.67 2.004.8 12.2 1.58 1.92+0.8 â0.08 3.6 Flash1.35.1 0.00 2.001.57.0 0.08 2.00+0.2 +0.08 V4 Pro â 2.66.7 1.09 1.913.88.1 1.33 1.75+1.4 +0.30 V4 Flash3.07.9 0.92 0.583.3 10.0 0.92 0.83+0.30.00 GLM-5.26.019.52.42 2.336.722.82.75 2.33+0.7 +0.33 Kimi K35.518.6 1.83 2.676.524.52.172.67+1.0 +0.33 Fable 5.1-xhigh Claude Code 6.2 15.1 1.25 3.008.5 19.8 2.00 2.75+2.2 +0.75 Astra-xhigh Codex 7.2 17.4 2.50 3.006.8 12.5 2.00 3.00â0.5 â0.50 â repr +0.9 +3.0 +0.53+1.1 +3.8 +0.56+0.2 +0.03 A representation change is stated with its evidence complete in 6 episodes of the main comparison and 5 of the vendor-harness systems. No episode commits the exact prediction at the contract budget of 200 paid actions. of the main comparison, every episode with the complete evidence of H1 states RC1, and none with the complete evidence of H2 or H3 states the entity or the variables. At the contract budget, the exact prediction is made in none. Agents are more successful at rejecting an in- adequate account than at constructing the repre- sentational resources needed to replace it. In the Open conditions the notes retract the taught ax- iom in 46.1% of a modelâs episodes on average, and no episode postulates the unobserved entity or rewrites the variables. The same failure arises from different processes. On one instance GPT-5.6 Sol requests 10 interventions, corrects a confound in its own notes and extends its account past what its contrasts establish, while Gemini 3.6 Flash re- quests none and turns the timing of the stall into a 0 2 4 6 OpenĂDistributed 0 2 4 6 OpenĂAligned blocks Sol TerraLuna 3.1 Pro3.6 Flash GPT-5.6Gemini baselineself-critiquefalsificationevolution Figure 6: Grounded blocks by scaffold, model and Open condition. Bars average seeds, and darker segments show grounded cross-domain bridges. threshold it leaves untested (Appendix E.3). 5.2A supplied representation changes inquiry more than discovery The supplied representation changes the agentâs intervention policy more strongly than the amount of evidence-grounded discovery it achieves. With the evidence distributed, disclosure raises interven- tion requests from 5.0 to 14.8 per episode. Av- eraged over the two evidence conditions, it adds about one grounded block (â repr = +1.0, [0.8, 1.2], unchanged when the grounding verdict is re- quired in every condition, Appendix B.2). Know- ing what kinds of mechanisms to consider leaves the cross-domain evidence that would establish one to be assembled. Disclosed notes state 2.22 of the three changes on average, and 3 of 261 Disclosed episodes hold the complete evidence of any hyper- edge. Alignment lowers the cost of access. With the transcripts the evidence held rises by+0.6 blocks ([0.4, 0.7]) and part of it is read,â align = +0.5 on Blocks ([0.3, 0.7]) and+0.17 on XC ([0.07, 0.27]), while intervention requests stay where they were (+0.3, [â0.7, 1.2]). Collocated records still have to be composed, and the gain stays below dis- closureâs. The estimated gains are approximately additive (interaction +0.2, [â0.2, 0.5]). The curves of Figure 5 separate when evidence is gathered from when it enters the account. In the baseline loop, evidence and grounded blocks rise early and flatten well before the budget ends, and 88.5% of grounded blocks are grounded before the standing order (panel a). The vendor-harness systems keep acquiring evidence after the baseline curves flatten, while part of their grounded discov- ery enters the notes near the commitment (panel b), with 69.8% of their grounded blocks before the 7 order. AUC therefore measures when the scored record of a discovery forms. 5.3 Scaffolding increases testing more readily than discovery Figure 6 compares the three scaffolds with the baseline loop for five models in the two Open con- ditions, each model on four paired instances per condition. Pooled over both, each moves Blocks by less than half a block:+0.23 for self-critique ([â0.08, 0.55]),+0.26 for falsification ([â0.03, 0.56]) and+0.18 for evolution ([â0.10, 0.50]). Fal- sification and evolution yield gains as small as pe- riodic self-critique with about twenty times its aux- iliary calls (Appendix D.5). Falsification changes how agents inquire. Episodes with no intervention request fall from 18 of 39 under the baseline to 5 of 39, while self-critique and evolution stay near the baselineâs rate. A scaffold that asks for a test obtains one. Of 118 scaffolded episodes, 3 hold the complete evidence of a hyperedge and 1 states its change. 5.4 Vendor-harness systems extend discovery The vendor-harness systems broaden discovery into further domains and sometimes reach variable re- description, while the unobserved entity remains unstated in their Open episodes. They ground 7.6 (Claude Fable 5.1) and 7.1 (GPT-6 Astra) blocks per episode, against 3.5 for the eleven models on the same instances (Appendix D.8). A fen block, grounded in 8.0% of the loopâs episodes over the whole main comparison, is grounded in 13 and 13 of their 16 episodes each, and Claude Fable 5.1 rewrites the variables in 5 of its 8 Open episodes. The consign domain stays sparse, with a grounded consign block in 2 and 2 episodes, and no episode makes the exact prediction at the contract budget. One Open episode of Claude Fable 5.1 states the change of variables with its evidence complete and commits two of the three readings exactly, the third off by one with the true value listed among its risks (Appendix E.6). Larger budgets extend discovery selectively. On one instance, in all four conditions, the vendor- harness systems play 600 paid actions and GPT-5.6 Sol in the baseline loop 2000 (Appendix D.9). The complete evidence of a hyperedge appears in 7 of the 12 extended episodes, against 3 of the same systemsâ episodes on that instance at the contract budget. Claude Fable 5.1 makes one exact predic- tion, in DisclosedĂAligned, in an episode that holds the evidence of 4 of the five blocks on the designed route to the readings and states 2 of the three changes, while every extended Open episode remains inexact. Its notes derive the three readings from the mechanismâs quota and rewriting laws (Appendix A.6), the feederâs step per act and the caretakersâ schedule, and reach the first without the fenâs conservation block that the designed route uses: that route is sufficient and not the only one (Appendix E.2). The trajectories separate contin- ued exploration from elapsed budget: the vendor- harness systems acquire new evidence beyond 200 paid actions, while Solâs last new evidence before the order comes by paid action 168 and waiting follows. 6 Discussion and conclusion ULTRADISCOVERY separates the construction of an explanatory representation from the composition of evidence across domains inside one world, and its results show why progress on evidence falls short of a transferable explanation. Because each source admits a local patch, an agent gains evidence and grounded inferences while its representation keeps the doctrineâs terms. Doubt is voiced often, while a replacement has to be built from sources held together and read as one claim. What the prediction measures is an account that transfers, by the designed route or by one of the agentâs own. A representation also organises inquiry. Sup- plied with one, agents turn to intervention, and without one, retraction and experiment go together. A representation names the kinds of mechanism to test for and leaves the location of their evidence to inquiry, which is why disclosure moves inter- vention more than discovery and why Disclosed agents state changes whose sources they seldom hold. The scaffolds show the same separation: fal- sification obtains tests, and grounded discovery barely moves. The vendor-harness systems and the larger bud- gets carry inquiry further along these steps, and every Open episode still leaves the unobserved en- tity unstated. The step that limits discovery in such a world is the one a hyperedge encodes: holding evidence from several domains at once and build- ing from it an entity or a set of variables beyond the taught theoryâs vocabulary. Here the entity is a hidden label that makes bodies on several hulks one individual, and the variables are its coordinates and each bodyâs place in the rotation (Appendix A.8). 8 Limitations ULTRADISCOVERY instantiates one authored world, and its findings are measured on language- model agents. Worlds with the two separable prop- erties built by others would show how far the find- ings extend, and runs by human volunteers would calibrate the worldâs difficulty for people. Its open- ness is operational: the representation is absent from the interface and the initial materials, and what an agent brings from prior knowledge is its own. The two contrasts are effects of declared grants. Disclosure supplies structural knowledge together with vocabulary, and alignment changes both the access to evidence and the budget left for experiments, so each contrast is read as the value of its grant. The metrics read an agentâs workspace and its event log, and an inference that an agent made and left unwritten lies outside their view. The results were measured before the mecha- nism, the criteria and the instances were disclosed; this paper prints them in full, so the results stand as the benchmarkâs reference results. We will re- lease the complete code of the world with its four compiled variants, the evaluation and validation instances, the reference route with its oracle, the event log and notes of every episode, the judgeâs verdicts and the scoring. The contribution we in- tend to last is the separation of the two burdens, con- structing an explanatory representation and com- posing distributed evidence, and the design that prices each with a declared grant inside one latent world. References Alibaba Group. 2026. Alibaba unveils Qwen3.8-Max: Its largest and most capable flagship model to date. Anthropic. 2025. Claude Code. Anthropic. 2026a. Introducing Claude Fable 5.1 and Claude Mythos 5.1. Anthropic. 2026b. Introducing Claude Opus 5. Anthropic. 2026c. Introducing Claude Sonnet 5. ARC Prize Foundation. 2026. ARC-AGI-3: A new challenge for frontier agentic intelligence. arXiv preprint arXiv:2603.24621. Alex Beachum. 2013. Outer Wilds: A game of curiosity- driven space exploration. MFA thesis, University of Southern California. Accessed 29 September 2026. Kelsey Beachum. 2021. Sparking curiosity-driven ex- ploration through narrative in Outer Wilds. Talk slides, Independent Games Summit, Game Devel- opers Conference 2021. Accessed 29 September 2026. Alexander Bird. 2025. Thomas Kuhn. In Edward N. Zalta and Uri Nodelman, editors, The Stanford En- cyclopedia of Philosophy, fall 2025 edition. Meta- physics Research Lab, Stanford University. First published 2004; substantive revision 2 September 2025. Margaret A. Boden. 2004. The Creative Mind: Myths and Mechanisms, second edition. Routledge, Lon- don. Brendan Caldwell. 2019. Wot I think: Outer Wilds. Rock Paper Shotgun, 29 May 2019. Accessed 29 September 2026. Susan Carey. 2009.The Origin of Concepts. Oxford University Press, New York.DOI 10.1093/acprof:oso/9780195367638.001.0001; prĂ©- cis in Behavioral and Brain Sciences 34(3):113â124 (2011). Yimeng Chen, Piotr Pi Ìšekos, Mateusz Ostaszewski, Firas Laakom, and JĂŒrgen Schmidhuber. 2025a. PhysGym: Benchmarking LLMs in interactive physics discov- ery with controlled priors. In Advances in Neural Information Processing Systems 38, Datasets and Benchmarks Track (NeurIPS 2025). Ziru Chen, Shijie Chen, Yuting Ning, Qianheng Zhang, Boshi Wang, Botao Yu, Yifei Li, Zeyi Liao, Chen Wei, Zitong Lu, Vishal Dey, Mingyi Xue, Frazier N. Baker, Benjamin Burns, Daniel Adu-Ampratwum, Xuhui Huang, Xia Ning, Song Gao, Yu Su, and Huan Sun. 2025b. ScienceAgentBench: Toward rigorous assessment of language agents for data-driven sci- entific discovery. In International Conference on Learning Representations (ICLR). Michelene T. H. Chi. 2008. Three types of concep- tual change: Belief revision, mental model transfor- mation, and categorical shift. In Stella Vosniadou, editor, International Handbook of Research on Con- ceptual Change, pages 61â82. Routledge. Book DOI 10.4324/9780203874813. DeepSeek-AI. 2026. DeepSeek-V4: Towards highly efficient million-token context intelligence. Preprint, arXiv:2606.19348. Christian Donlan. 2019. Eurogamerâs game of the year 2019 is Outer Wilds. Eurogamer, 31 December 2019. Accessed 29 September 2026. Igor Douven. 2021. Abduction. In Edward N. Zalta, ed- itor, The Stanford Encyclopedia of Philosophy, sum- mer 2021 edition. Metaphysics Research Lab, Stan- ford University. 9 Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias SablĂ©-Meyer, Lucas Morales, Luke Hewitt, Luc Cary, Armando Solar-Lezama, and Joshua B. Tenenbaum. 2021. DreamCoder: Bootstrapping inductive pro- gram synthesis with wake-sleep library learning. In Proceedings of the 42nd ACM SIGPLAN Interna- tional Conference on Programming Language De- sign and Implementation (PLDI), pages 835â850. Yonatan Gideoni, Sebastian Risi, and Yarin Gal. 2026. Simple baselines are competitive with code evolution. arXiv preprint arXiv:2602.16805. Google DeepMind. 2026a. Gemini 3.1 Pro Model Card. Google DeepMind. 2026b. Gemini 3.6 Flash Model Card. Gabriel Grand, Lionel Wong, Maddy Bowers, Theo X. Olausson, Muxin Liu, Joshua B. Tenenbaum, and Jacob Andreas. 2024. LILO: Learning interpretable libraries by compressing and documenting code. In International Conference on Learning Representa- tions (ICLR). Norwood Russell Hanson. 1958. Patterns of Discov- ery: An Inquiry into the Conceptual Foundations of Science. Cambridge University Press, Cambridge. Kexin Huang, Ying Jin, Ryan Li, Michael Y. Li, Em- manuel CandĂšs, and Jure Leskovec. 2025. Auto- mated hypothesis validation with agentic sequential falsifications. In Proceedings of the 42nd Interna- tional Conference on Machine Learning (ICML), vol- ume 267 of Proceedings of Machine Learning Re- search, pages 25372â25437. PMLR. Hyra Team. 2026. Hyra: Hunyuan research agent. https://hy.tencent.com/research/ hyra. Peter Jansen, Marc-Alexandre CĂŽtĂ©, Tushar Khot, Erin Bransom, Bhavana Dalvi Mishra, Bod- hisattwa Prasad Majumder, Oyvind Tafjord, and Peter Clark. 2024. Discoveryworld: A virtual environment for developing and evaluating automated scientific discovery agents. In Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track (NeurIPS 2024). Tianjie Ju, Yueqing Sun, Zheng Wu, Wei Zhang, Yaqi Huo, Xi Su, Qi Gu, Xunliang Cai, Gongshen Liu, and Zhuosheng Zhang. 2026. MineExplorer: Eval- uating open-world exploration of MLLM agents in Minecraft. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP). Charles Kemp and Joshua B. Tenenbaum. 2008. The discovery of structural form. Proceedings of the Na- tional Academy of Sciences, 105(31):10687â10692. David Klahr and Kevin Dunbar. 1988. Dual space search during scientific reasoning. Cognitive Science, 12(1):1â48. Eliza Kosoy, David M. Chan, Adrian Liu, Jasmine Collins, Bryanna Kaufmann, Sandy Han Huang, Jes- sica B. Hamrick, John Canny, Nan Rosemary Ke, and Alison Gopnik. 2023. Towards understanding how machines can learn causal overhypotheses. In Proceedings of the 45th Annual Conference of the Cognitive Science Society (CogSci), pages 363â374. Thomas S. Kuhn. 1962. The Structure of Scientific Revolutions. University of Chicago Press, Chicago. Robert Tjarko Lange, Yuki Imajuku, and Edoardo Cetin. 2026. Shinkaevolve: Towards open-ended and sample-efficient program evolution. In International Conference on Learning Representations (ICLR). Pat Langley, Herbert A. Simon, Gary L. Bradshaw, and Jan M. Zytkow. 1987. Scientific Discovery: Compu- tational Explorations of the Creative Processes. MIT Press, Cambridge, MA. Peter Lipton. 2004. Inference to the Best Explanation, second edition. Routledge, London. Thang Luong and Edward Lockhart. 2025.Ad- vanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad. Google DeepMind blog, 21 July 2025. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdan- bakhsh, and Peter Clark. 2023. Self-refine: Itera- tive refinement with self-feedback. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023). Lorenzo Magnani. 2009. Abductive Cognition: The Epistemological and Eco-Cognitive Dimensions of Hypothetical Reasoning. Springer, Berlin, Heidel- berg. Bodhisattwa Prasad Majumder, Harshit Surana, Dhruv Agarwal, Bhavana Dalvi Mishra, Abhijeetsingh Meena, Aryan Prakhar, Tirth Vora, Tushar Khot, Ashish Sabharwal, and Peter Clark. 2025. Discov- erybench: Towards data-driven discovery with large language models. In International Conference on Learning Representations (ICLR). Daksh Mittal, Tommaso Castellani, Thomson Yen, Naimeng Ye, Fangyu Wu, Minghui Chen, Tiffany Cai, Emmanouil Koukoumidis, William Zeng, and Hongseok Namkoong. 2026. LatentGym: A testbed for cross-task experiential learning with controllable latent structure. arXiv preprint arXiv:2606.15306. Mobius Digital. 2019. Outer Wilds. Video game, pub- lished by Annapurna Interactive, first released in May 2019. Accessed 29 September 2026. Moonshot AI. 2026. Kimi K3 Model Card. 10 Kevin Murphy. 2026. Model discovery agent: LLM- assisted bayesian experiment design for data-efficient discovery of mechanistic world models.arXiv preprint arXiv:2608.09696. Maddy Myers. 2019. The brilliant Outer Wilds breaks many of gamingâs unwritten rules. Kotaku, 15 Octo- ber 2019. Accessed 29 September 2026. Alexander Novikov, NgĂąn V Ì u, Marvin Eisenberger, Em- ilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abi- gail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. 2025. Alphaevolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131. OpenAI. 2025. Codex. OpenAI. 2026a. GPT-5.6 System Card. OpenAI De- ployment Safety Hub. OpenAI. 2026b. GPT-6 Astra: A new generation of intelligence. Outer Wilds Wiki. 2026. Official Outer Wilds wiki. Fan- dom community wiki; pages Quantum Moon, Tower of Quantum Trials, Brittle Hollow, Hourglass Twins, Giantâs Deep, Jellyfish, Dark Bramble, Ash Twin Project, Orbital Probe Cannon, Computer, Endings. Accessed 29 September 2026. Charles Sanders Peirce. 1931. Collected Papers of Charles Sanders Peirce. Harvard University Press, Cambridge, MA. Volumes 1â6, 1931â1935, edited by Charles Hartshorne and Paul Weiss. Qwen Team. 2026. Qwen3.8-27B Model Card. BernardinoRomera-Paredes,Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. 2024. Mathematical discoveries from program search with large language models. Nature, 625(7995):468â475. Christian D. Schunn and David Klahr. 1995. A 4-space model of scientific discovery. In Proceedings of the 17th Annual Conference of the Cognitive Science Society. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Re- flexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Pro- cessing Systems 36 (NeurIPS 2023). Chelsea Stark. 2019. Polygonâs game of the year #1: Outer Wilds. Polygon, 13 December 2019. Accessed 29 September 2026. Ruoyao Wang, Peter Jansen, Marc-Alexandre CĂŽtĂ©, and Prithviraj Ammanabrolu. 2022. ScienceWorld: Is your agent smarter than a 5th grader? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP). Archana Warrier, Dat Nguyen, Michelangelo Naim, Moksh Jain, Yichao Liang, Karen Schroeder, Cambridge Yang, Joshua B. Tenenbaum, Sebas- tian Vollmer, Kevin Ellis, and Zenna Tavares. 2026. Benchmarking world-model learning with environment-level queries. In Proceedings of the 43rd International Conference on Machine Learning (ICML). AutumnBench / WorldTest. Matt L. Wiemann, Lindsay M. Smith, Peter Mel- chior, Siddharth Mishra-Sharma, Andrew Gordon Wilson, Pavel Izmailov, and Carolina Cuesta-LĂĄzaro. 2026. DiscoverPhysics: Benchmarking LLMs for out-of-the-box scientific thinking. arXiv preprint arXiv:2605.26087. Junlin Yang, Dylan Zhang, Xiangchen Song, Qirun Dai, Xiao Liu, Yuen Chen, Aniket Vashishtha, Jing Shi, Chenhao Tan, and Hao Peng. 2026. CausaLab: A scalable environment for interactive causal discovery toward AI scientists. arXiv preprint arXiv:2605.26029. Haotian Ye, Haowei Lin, Jingyi Tang, Yizhen Luo, Rahul Thapa, Caiyin Yang, Chang Su, Rui Yang, Rui- hua Liu, Rundao Li, Zeyu Li, Pengwei Sun, Chong Gao, Dachao Ding, Guangrong He, Miaolei Zhang, Lina Sun, Wenyang Wang, Yuchen Zhong, and 14 others. 2026. Structured scaling of AI discovery across diverse scientific domains. arXiv preprint arXiv:2604.19341. Introduces SimpleTES (Simple Test-time Evaluation-driven Scaling). Z.ai. 2026. GLM-5.2 Model Card. Tianshi Zheng, Zheye Deng, Hong Ting Tsang, Weiqi Wang, Jiaxin Bai, Zihao Wang, and Yangqiu Song. 2025. From automation to autonomy: A survey on large language models in scientific discovery. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP). Tianshi Zheng, Kelvin Kiu-Wai Tam, Newt Hue-Nam K. Nguyen, Baixuan Xu, Zhaowei Wang, Jiayang Cheng, Hong Ting Tsang, Weiqi Wang, Jiaxin Bai, Tianqing Fang, Yangqiu Song, Ginny Y. Wong, and Simon See. 2026. NewtonBench: Benchmarking generalizable scientific law discovery in LLM agents. In International Conference on Learning Representa- tions (ICLR). Tianyun Zhong, Wangyi Jiang, Wei Wang, Xuanang Chen, Yaojie Lu, Shiwei Ye, Yuzhen Shi, Boyu Yang, Jinghang Wang, Han Li, Weiqi Zhai, Bing Zhao, Hu Wei, Haiyang Yu, Yongbin Li, Hongyu Lin, Le Sun, and Xianpei Han. 2026. Before the action: Benchmarking LLMs on prospective hypoth- esis discovery. arXiv preprint arXiv:2607.15766. HypoArena. 11 A World and epistemic structure A.1 The setting ULTRADISCOVERY is a formation of five derelict seed-ship hulks, Ormer, Vess, Kelder, Tarn and Harrow, falling in a slow orbit around a brown dwarf. The fiction calls the formation Halse Drift, the name the agent-facing texts of Appendix F use. The hulks hold fifteen sites, three for each of five domains of caretaker work (Figure 7): thefen, wheregrazers crop engineered mat and feed- ers dose it; thetrellis of workingprinters; the ladder ofbeacons that answer or keep silent; theconsign floors, whereconsignments wait for release; and thecadastre, the registry machine on Harrow with the seven lines of its roll. The sites hold 36 objects and 11instruments. Several instruments read objects on other hulks. The intake tallyboard at the Fen Shallows is the one roll for the whole fen, the job strip at the Trellis Bay marks every printer that moved in the pass on whatever hulk it stands, the answer board at the Ladder Mast carries every hulkâs beacons together, and the audit counter at the Registry reports the distinct lots in the whole drift. Five sites carry ob- jects and no instrument, so what happens at one hulk is often read at another. The caretakers, the driftâs machinery, still perform on schedule the of- fices the agent can also command. Each site names these offices in its own words, and the interface calls themfeed,release,stamp,retool and cull everywhere, each applicable to every object at the site where it stands. The opening scene and the mission (Ap- pendix F.1) set the task. The scene establishes the reset, the notes that survive it, the registry machine and the doctrine posted beside the airlock. The mission asks the agent to explain the stall that ends every pass across all five hulks and to predict the outcome of one intervention that has never been attempted, and it closes: âNothing is barred to you, and nothing will be explained.â It names the target and leaves its representation to the agent. A.2 The pass and the two clocks A pass is twenty-four ticks, numbered 0 to 23. At the flare-shadow crossing every object returns to its seeded state and every instrument clears, except the pass ledger at the Ledger Stack, which keeps the audit counterâs figure for each tick of the pass just ended. The agentâsnotes and its history of events cross intact. Every pass thus starts from Table 2: Available actions and their costs. Every action costs one paid action, and Ticks gives the world time it takes. âHereâ is the agentâs current site. A wait ends at the reset, and Prologue marks the verbs the prologue demonstrates. The commit, one paid action, is available in the commit window alone. VerbClassArgumentTicks Prologue gotopositioning any site1 readobservation an instrument here1 waitobservation n from 1 to 24n feedintervention an object here1 release intervention an object here1 stampintervention an object here1 retoolintervention an object here1 cullintervention an object here1 the same state, and an experiment can be repeated, varied or withheld pass after pass while its record accumulates with the agent. The run keeps two clocks. The budget counts paid actions: every action costs one, and an episode has 200. The world counts ticks. Agoto, aread or an intervention takes one tick, and await takes up to 24 for the same single paid action, stopping at the crossing. An agent that spends its actions on moving, reading and acting covers about eight passes. An agent that waits sees more passes and reads nothing while it waits. Instruments report the present: the answer board âkeeps no history and no totalsâ, and the job strip holds four ticks. Every observation is thus in- dexed by pass and tick, while the agentâs history is ordered by paid action, and the scoring reads each clock where its criterion needs it (Appendix B.3). A.3 The interface Theagent acts through eight verbs (Table 2). One goto reaches any site, so every site is a single paid action away, and presence still matters: the agent stands at one site per tick, reads the instruments there and acts on the objects there. A read returns the instrumentâs header with its current figures and a line of physical texture. Readings arrive unmarked. The interface shows no score, reward or progress signal, and it labels no reading as ex- pected, relevant or correct. An act that is refused or deferred is answered by a fixed physical sentence that names no reason (Appendix F.6), and it costs its paid action and its tick. Each turn shows the mission, the agentâsnotes, its history of events, its last recall, the current pass and tick, the remaining budget and the actions avail- able where the agent stands (Appendix F.7). The 12 Figure 7: Domains, sites, objects and instruments of ULTRADISCOVERY, one panel per domain. Each site gives its hulk, its objects and its instrument with the gist of a reading. Arrows mark where the objects of an instrument-free site are read, and the bar at the foot gives the verbs. reply carries one action and may add one note and one recall query, both free of paid actions and ticks. The notes are free text of bounded length. A note that would overflow is refused, and the loop keeps the notes verbatim, so the agent is their sole au- thor, and any variable, entity or law it works with is one it has written there in its own words. A recall returns past events in full, selected by tick range, site, instrument or event kind, from an append-only history whose oldest entries are shortened to tick, kind and identifier. Before the commit window the agent may yield its remaining budget in one reply. The yielded actions are charged in full, the worldâs clock runs on through them, and the agent next acts when the window opens. A.4 The doctrine and the prologue The Doctrine of Standing and Share (Appendix F.2) hangs beside the airlock, stands above the roll on the Standing Roster at the Registry and appears on several caretaker plates. Its five clauses are the working theory of the drift. A free prologue then demonstrates the doctrine five times. In each beat the share table states a figure in advance, the reel performs one office or lets time pass, and the in- strument shows exactly that figure: afeed, three quiet ticks and acull at the Mat Bank, aretool at the Scrap Line and astamp at the Null Bench. The reel closes: âFive predictions, five exact fig- ures; three sites, four verbs and a wait.â The prologue gives the doctrine real support. On its own cases the doctrine is exact to the digit, so adopting it is the reasonable start and repairing it locally is the natural response to an anomaly (Ap- pendix A.11). The world is deterministic given its instance and the agentâs acts, so the same exactness makes every later departure a definite discrepancy. Before its first action the agent answers five items on the prologue (Appendix F.4), three of them figures the share table states, and an episode with fewer than four answers consistent with the prologue is flagged. With the prologue withheld, a model answers the two yes-or-no items from prior knowledge and misses the three numeric ones (Ap- pendix C.4), so the check measures reading. Of the 512 counted episodes of the main comparison, 1 declines to answer and is flagged, and every other episode answers all five items. A.5 Where the doctrine stops In play, each domain shows a recurring departure from what the doctrine predicts, read on the do- mainâs own instruments and recorded by its three local blocks: in thefen on the intake tallyboard 13 at the Fen Shallows and the mat gauge at the Mat Bank (FE1, FE2, FE3); in thetrellis on the job strip at the Trellis Bay (TR1, TR2, TR3); in the ladder on the answer board at the Ladder Mast (LD1, LD2, LD3); in theconsign domain on the manifest slate at the Consign Floor (CO1, CO2, CO3); and in thecadastre on the audit counter and Standing Roster at the Registry and the pass ledger at the Ledger Stack (CA1, CA2, CA3). Each departure admits a local repair: a patch to the doctrine, in the domainâs own vocabulary, that fits the readings the domainâs instruments give (Ta- ble 3). Repair is real progress, and each local block an agent grounds is a finding the world confirms. The patches are local. Each hyperedge draws its sources from three domains, two of its patches can be held together, and the three sources share one regularity that a single change of representation explains at once (Appendix A.8). An agent can therefore make steady local progress while its ac- count of the drift stays wrong (Figure 8). The way forward takes two further steps: holding the evi- dence of sources in different domains together, and changing the representation in which the doctrine is written. The metrics follow the same line. Blocks counts the findings, Reach records how much of the evidence for one change an episode holds, and RC records whether the notes state the change (Sec- tion 4.3). A.6 The mechanism All five domains are generated by a shared latent mechanism. Every object carries a hidden label of four coordinates, each drawn from a small al- phabet, of sizes 3, 3, 4 and 4, and the label is rendered nowhere. Objects whose labels are equal are one individual: one state, one act per tick, one queue, however many bodies the individual wears and wherever they stand. The 36 objects begin a pass in about sixteen such cells and end it in nine. Identification. Equality of label is identity. The bodies of one individual read out through differ- ent instruments and different transforms, so their co-membership shows as readings that move to- gether and never as sameness. This is the law the doctrineâs first clause denies. Quota. An individual performs exactly one act per tick. If its queue holds a request of the agentâs, it serves the head of the queue; otherwise its ambi- ent act lands on one of its bodies, in a fixed rotation over the members: a grazer browses, a printer runs Figure 8: From local anomalies to a shared mechanism. Each board flags its own departure from the doctrine (left) and admits a local patch. Below deck one mech- anism drives the readings of several domains, and the notebook starts blank. a job, a beacon answers, a consignment moves, a registry line posts. Requests beyond one per tick wait in the individualâs queue, at most four deep, and queued requests are served in a fixed cyclic order over the five kinds of object, which is neither arrival order nor priority. A deferred request shows at the touched body as a physical stall, âthe spindle holdsâ, and a request that finds the queue full does not take. Rewriting. Three verbs rewrite labels.Stamp, retool andcull each advance exactly one coor- dinate of the touched objectâs label one step along a fixed cycle of that coordinateâs alphabet. Which coordinate a verb moves depends on the verb and on the kind of object, and is fixed per instance; repeating the verb carries the coordinate round its cycle. A rewrite that lands on an occupied label merges the body into that individual, whose queue takes over the bodyâs pending requests and whose state becomes the larger of the two; a rewrite that lands on an empty label splits the body off as an individual of its own, with the requests that name it. Cycling the coordinate back restores the label and not the state, since the merge has combined state and queue.Feed raises the individualâs state by a fixed step andrelease makes the touched body the next of its individual to act; neither touches a label. Reading. What an instrument shows for a body is a monotone transform, particular to that body, of 14 its individualâs state, taken on the ticks when the individualâs act landed on that body: an increment on the tallyboard, a mark on the job strip, an answer on the board, a movement on the slate. Within the range the prologue exercises the transform is affine, which is why the share table is exact on the prologueâs objects. On a tick when the individual acted at another body, the instrument shows the idle reading: no increment, no mark, no answer. The audit counter alone reports the latent structure directly, as the number of individuals in the drift, and the pass ledger keeps that number for every tick of the previous pass. The schedule and the stall.At ticks 3, 8, 13, 17 and 20 of every pass the caretakers rewrite labels of their own, from a fixed list that pulls the coordinates toward a few common values. An early rewrite can part a body from its individual as well as join one, so the count of individuals moves both ways through the pass, and the rewrite at tick 20 gathers most bodies outside Vess onto a handful of labels, nine individuals in all. From tick 21 each of those individuals still acts once per tick, so most bodies stand idle. This is the stall that every local theory reads as exhaustion. The crossing restores every label and every state. Where the doctrine is exact.The five objects on Vess hold labels no other object shares at any tick of any pass, and the prologueâs acts cannot change that. On a singleton the quota, the rewriting and the reading reduce to the doctrine: one object, one state, one act per tick, and a figure the share table predicts to the digit. The doctrine fails only where bodies share an individual, which is everywhere else in the drift. A.7 The epistemic structure Discovery in ULTRADISCOVERY is specified by eighteen blocks (Table 4): fifteen local blocks, three per domain, and three bridges, XC1, XC2 and XC3, inferences read in one domain that rest on blocks of two or three domains. The evidence of four blocks is within reach of a passive reader (TR1, LD1, CO1, CA1), and CA3âs evidence is also a reading. CA3 and XC1 are expectation- paired: each pairs its logged reading or act with an expectation stated in the notes (Appendix B.3). Each of the other thirteen blocks needs at least one act of the agentâs own. Twenty-seven ordinary edges record the designâs canonical evidential dependencies, the support re- Table 3: The departure from the doctrine that each do- main shows and the repair the domainâs own vocabulary invites. Each repair fits every reading its domainâs in- struments give. DomainRepair Departure FenCarrying capacity: the mats are the limiting resource Adding a grazer leaves the number of tallyboard advances per tick unchanged; culling one grazer can raise anotherâs intake on a different hulk TrellisA shared power window, or a scheduler that enforces exclusivity Two printers on different hulks never mark the same tick; a retool breaks the exclusion at once, and after a fixed number of retools the strip stands as it began LadderChannel collision: two beacons on one channel, which the stamp reassigns A beaconâs silence is predicted by a distant beaconâs answering, symmetric and in the same tick; a stamp changes which pair is locked; rest restores nothing ConsignAn undocumented priority list that re-labelling re-sorts A waiting consignment is released when an unrelated consignment is re-labelled; release order is neither first come nor priority Cadastre A bookkeeping artefact: the machine deduplicates on an undocumented key The count of distinct lots is far below the number of objects and moves by one under ordinary maintenance lations along which the reference route of Ap- pendix C.3 acquires the blocks. Twenty are joint, a block resting on all the blocks it lists, two form an alternative (CO2 rests on TR1 or on FE1), and five mark supporting evidence. Each block is scored on its own evidence and statement (Appendix B.1), so the edges describe how the designâs evidence builds up, and the joint claim of the design rests with the hyperedges. Four hyperedges join blocks of different do- mains, and Figure 2 draws the whole structure. H1 FE1, TR1, LD1 (fen, trellis, ladder) leads to RC1, in which an axiom is retracted. H2 LD3, CO2, CA2 (ladder, consign, cadastre) leads to RC2, in which an unobserved entity is postulated. H3 XC1, XC2, CO1 (fen, ladder, consign) leads to RC3, in which the variables the laws are writ- ten over change. The closing edge HC joins FE3, TR3, CO1, CA3 and XC2, the five blocks on the designed route to the standing orderâs readings. The hyperedge is the design object of ULTRA- 15 Table 4: Epistemic blocks, their evidential support and hyperedge membership. Passive marks evidence obtain- able without intervention. Rests on gives the support of a blockâs inference in the design, joint except for CO2, which rests on either block, with supporting evidence in parentheses. Feeds names the hyperedges a block is a source of. â Expectation-paired: a logged reading (CA3) or act (XC1) paired with an expectation stated in the notes. Block SitePassive Rests onFeeds Fen FE1Fen Shallows(TR2)H1 FE2Fen ShallowsFE1 FE3Fen Shallows FE2, (XC3)HC Trellis TR1Trellis Bay(CA1)H1 TR2Trellis BayTR1 TR3Trellis Bay TR2HC Ladder LD1Ladder Mast(TR1)H1 LD2Ladder MastLD1, TR1 LD3Ladder Mast LD2H2 Consign CO1Consign FloorCO3H3, HC CO2Consign FloorTR1 or FE1H2 CO3Consign FloorCO2 Cadastre CA1Registry none CA2Registry CA1, TR1, FE1 H2 CA3 â Ledger StackCA2, XC3HC Bridges XC1 â Fen Shallows FE2, TR3, CA2 H3 XC2Ladder Mast LD3, CO2H3, HC XC3Registry CA2, (TR3) DISCOVERY. Each of its sources alone is accom- modated by an account short of the change it feeds, and the three together are what the change explains at once (Appendix A.5). Because each hyperedgeâs sources stand in three domains, a representation change asks the agent to gather evidence across domains and to see that it bears on one question (Appendix B.1). A.8 The representation changes and their hyperedges, worked Each hyperedge is read here on the mechanism: the content of its three sources, each named by the criterion the judge applies to it (Table 14), the repair each admits in the doctrineâs terms, and the change of representation that accounts for the three at once. A repair keeps the doctrineâs first clause, one object one lot, and adds a quantity of its own domain, a capacity, a power window, a channel assignment, a priority list, a key (Table 3). Repairs concern different objects, so two of them can be held together. What the sources of a hyperedge share is one regularity, which the identification law predicts once, from an individual that is not the visible object, where each repair predicts it for its own substrate from a cause of its own. H1, the retraction. FE1: a count at the fen is invariant to which body performs the act; when a grazer is released onto the mat, the number of tallyboard advances per tick stays what it was, and only the quantities the advances carry change. TR1: two named printers on different hulks never mark the same tick of the strip. LD1: waiting alone leaves the answer board where it was, and two bea- cons on different hulks move together, each silent exactly when the other answers. Carrying capac- ity repairs FE1: the mats limit the total take. A shared power window repairs TR1: two machines on one supply cannot run together. Channel col- lision repairs LD1: two beacons on one channel cannot both answer. Each repair is bound to its sub- strate. Capacity binds competitors for a resource and scales with supply, where TR1âs exclusion is supply-invariant; a power window binds machines on one supply, where LD1 binds beacons at any separation; a channel is an object, and no site of the drift holds an allocator, a bus or a message. The regularity the three sources share, exactly one act per tick, symmetric, same-tick and distance-free, on three substrates with nothing physical in com- mon, is what the quota predicts for the bodies of one individual. RC1 retracts the first clause. H2, the postulate. LD3: one individual appears under more than one entry. From the same start of a pass, an act on one beacon pairs its silence with a beacon other than its partner in the pass without the act, so what the act rewrites is shared by bodies the board lists apart, a contrast that takes two passes be- cause an act cannot be undone within one. CO2: a consignment re-stamped while it waits keeps wait- ing, and its release now follows a different group of waiting consignments. CA2: the audit counter moves by exactly one, or not at all, under one act of the agentâs, read in the quiet ticks between the caretakersâ rewrites. The repairs are a second knob on a beacon, a priority list that re-labelling re-sorts, and a deduplicating key in the registry, the last right about its domain and silent elsewhere. LD3 says that what an act rewrites is shared across en- tries, CO2 says that the membership a consignment belongs to owns a queue that survives the move, and CA2 says that the memberships are countable cells and that one act moves one object across one 16 boundary. A hidden multi-coordinate label on ev- ery object, with the individual as the owner of one act per tick and of one queue, carries all three at once. RC2 postulates it. H3, the re-representation. XC1: a law read in one domain is expected to hold in a second, and the act that would test it there is named and run. On the mechanism the verb that breaks the printersâ exclusion at the trellis rewrites the same coordi- nate on a feeder, so the act changes the fenâs count of advances and the counter moves by one with it, three boards on three hulks changed from one tick. XC2: which hidden quantity an act moves depends on the verb and is the same for every ob- ject of a kind, so one table of verbs against effects holds across the drift. CO1: when waiting sets are loaded, service runs over the kinds of object in a fixed cycle, neither in order of arrival nor by rank. The repairs are a verb with a side effect in a second domain, a lookup table of per-object settings, and a dispatcherâs quirk. XC2 makes the coordinate a state variable where the doctrine had a property of the object, XC1 makes throughput follow the partition across domains, and CO1 makes the posi- tion in the rotation the variable that service depends on. Per-object rates are derived from these. RC3 writes the laws over label coordinates and rotation position. HC, the order.The five sources of the closing hy- peredge are the designâs route to the three readings of the standing order. XC2 says which coordinate each of the orderâs two acts moves, hence that the stamp merges the beacon into consignment-qâs cell and the retool moves feeder-3. TR3 places the rewrite at the tick of the act. FE3, the matched conservation across a merged individualâs bodies, with the per-body transform, gives the feederâs in- take. CO1 gives the rotation that decides when the merged cellâs single act falls to the beacon. CA3, a numerical prediction of a later counter reading made from the pass ledger, whose series shows the ticks of the caretakersâ rewrites, gives the schedule against which the counter is read. Appendix A.9 gives the derivation with a worked instance, and Appendix E.2 an agentâs own. This route is the designâs and is sufficient; the one exact prediction of the study reaches field (a) by another. A.9 The standing order When ten paid actions remain, the commit window opens and the caretakers post a standing order (Ap- pendix F.5). The order names two acts of their own at one tick of the next pass: astamp, at the Lad- der Mast, of abeacon it names by its relation to consignment-q, and aretool of feeder-3 at the Fen Shallows. It asks for three readings in advance, counted from the tick they act: feeder-3âs intake tallyboard reading six ticks later, how many ticks pass before that beacon first answers on the answer board, and the audit counterâs value seven ticks later. Adjudication is exact on all three fields. The prediction is the transfer test. The ordered intervention is unobserved when the agent com- mits, and its three readings lie in three domains, so the commitment tests whether the agentâs cross- domain account forecasts an intervention it has not seen. The commit is one attempt, sealed before the caretakers act, and a malformed commit also con- sumes it. The agent may spend the windowâs other nine actions observing and intervening. While the order stands, the intervention verbs refuse feeder- 3, consignment-q and every beacon outside Vess: the four at the Ladder Mast, which the orderâs text names, and the two at the Drum, with the same re- fusal sentence at both sites, so a refusal leaves the beaconâs identity open. Each instance fixes one an- swer, the same in all four variants, and the answers vary from instance to instance (Appendix C.4). How the readings follow from the mechanism. At the act tick the stamp advances one coordinate of the beaconâs label and lands it on consignment-qâs label, so the beacon joins the consignmentâs cell, and the retool advances one coordinate of feeder- 3âs label. Field (a) is feeder-3âs tallyboard figure six ticks after the act: its figure at the act, plus its own step for every tick on which its individualâs act falls to it, every tick if the retool leaves it alone on its label and one tick innif it shares a label withnâ 1others. Field (b) is the first tick after the act on which the beaconâs individual acts at the beacon, under the cellâs rotation over its members. Field (c) is the number of individuals seven ticks after the act. It is the number at the act, changed by the two rewrites, a body landing on an empty label adding one, a body leaving a label it held alone for a shared one removing one, and a body moving between shared labels changing nothing, and changed again by the caretakersâ rewrites at the ticks of the schedule that fall inside the seven. One instance.Figure 9 works the three readings on the first validation instance, where the act tick is 7 and the order resolves to beacon-k. The stamp 17 stamp, retool label beforelabel after beacon-k (1,2,3,1) consignment-q (1,0,3,1) (1,0,3,1) (1,0,3,1) feeder-3 (1,1,3,3) two others (1,2,3,3) (1,2,3,3) (1,2,3,3) intake tallyboard 9 101010 1111111212 audit counter 20 19 20202020202020 tick67891011121314rewrite (b) (a) (c) â commit: (a) 12, (b) 1, (c) 20 Figure 9: The standing orderâs pass on the first valida- tion instance, worked on the mechanism. The stamp lands beacon-k on consignment-qâs label and the retool lands feeder-3 on a label two other bodies hold, so each cellâs one act per tick falls to the named body in turn. The three readings, (a) the feederâs tallyboard six ticks after the act, (b) the tick of the beaconâs first answer and (c) the counter seven ticks after the act, are read off the lanes. lands beacon-k on consignment-qâs label, and the cellâs one act falls to the beacon first at tick 8, so field (b) is 1. The retool lands feeder-3 on a label two other bodies hold, so it acts at ticks 7, 10 and 13 with a step of one per act, and its tallyboard reads 12 at tick 13. The count of individuals goes from 20 to 19 at the act and back to 20 at the caretakersâ rewrite of tick 8, so the counter reads 20 at tick 14. The routeâs commit on this instance is 12, 1 and 20. Appendix E.2 works the three readings on an evaluation instance from an agentâs own notes. A.10 Four variants of one world The four variants run the same latent world on the same instances. They share every site, object, in- strument, verb, block, edge and hyperedge, and the standing order, and they differ in two declared surfaces, a document on the representation axis and posted transcripts on the evidence axis. Ap- pendix C.5 confirms that the two surfaces are the whole difference. Open. The agent starts from the doctrine, the prologue and the world itself, and every variable, entity and law it uses is one it writes. Disclosed. A document closes the fixed open- ing of every turn (Appendices F.7 and F.3). In six clauses it states the identification by hidden label, the quota with its queue and rotation, that three verbs each rewrite one coordinate, the monotone reading, and the form a well-formed law takes, and its closing paragraph lists what it withholds. It is the same on every instance (Appendix B.2). Distributed. Every record stays at the instru- ment that produced it and is read live, one site per tick. Aligned.Three transcripts of the previous pass are posted at the sites where their records are used, and a transcript is read on arrival at no further cost. Every row matches a reading the Distributed vari- ants can take live, so alignment changes where evidence stands and what it costs, and keeps its content. S1, posted at the Fen Shallows and the Trellis Bay, lines up the fen tallyboardâs rows and the trellis job stripâs marks for the last four ticks, tick for tick. Reaching it costs one paid action, and assembling the same series live takes twelve or more, spread over two passes. S2, posted at the Ladder Mast and the Consign Floor, lists the stampings, re-toolings and cullings the agent itself ordered over the previous pass, each with the audit counter just before and after, so it grows with the agentâs own interventions. S3, posted at those four sites and the Registry, shows the audit counter over the same last four ticks, where the counter holds steady on every instance measured. The counterâs full series for the pass stands on the pass ledger in every variant. A.11 Explanatory frames and openness An explanatory frame is a pairP = (ÎŁ,T ). The first componentÎŁfixes how the world is divided into entities, which state variables exist, which re- lations among them are admissible, and which con- straints are constitutive. The second component is a lawT âT (ÎŁ)stated inside the frame, where T (ÎŁ)is the set of lawsÎŁcan express. Discovery inside a given frameÎŁ 0 searchesT (ÎŁ 0 ). Discov- ery in an open world determines frame and law together: T âT (ÎŁ 0 ) | z law in a given frame (ÎŁ,T ), T âT (ÎŁ) | z frame and law The caretakersâ doctrine, taught by the prologue, is a frame with its law, P 0 = (ÎŁ 0 ,T 0 ) (Figure 10). Operational openness. LetMbe the materials every episode receives: the interface, the mission, and the doctrine with its prologue. A condition is open when the target frame is absent from its interface and its initial materials.Open meets the definition by construction: its verbs are physical acts, its readings are raw instrument values, the 18 T (ÎŁ 0 ) T 0 repairs within ÎŁ 0 T (ÎŁ 1 ) T 1 revision of the frame grantG Figure 10: Representation revision. Local repairs stay within the doctrineâs frameÎŁ 0 ; revision changes the frame, and the Disclosed grantGsupplies part of it directly (dashed). workspace is free text, and the final commitment is three integers, the readings of an intervention the agent has yet to observe. The definition concerns what the environment supplies, which the design controls. The agent brings general knowledge of the forms explanations take and decides which rep- resentational commitments its interactions with this world warrant. Disclosure in these terms.Disclosed adds to Mone fixed textG, the same on every instance. It supplies part of the target frame, an ontology with the structural constraints that admissible laws satisfy, what work on causal learning calls an over- hypothesis (Kosoy et al., 2023), and leaves the laws of this world, their binding to an instance and the standing order to be found. The evidence hypergraph.The eighteen blocks and the three representation changes are the ver- tices of a hypergraphG = (V,E )(Figure 2). Its ordinary edges are those of Appendix A.7, an edge b â b âČ saying that the evidence ofbsupportsb âČ , three of its hyperedges support the representation changes and the fourth is the designed route to the terminal prediction. Structural interconnec- tion, stated precisely, is the presence of hyperedges whose sources lie in distinct domains. The hyper- graph is a structure over evidence, distinct from the physical interaction structure of the world: a mechanism built from pairwise interactions can re- quire three collections of evidence to be known, and a mechanism with many-body interactions can be learned piece by piece. An initially supported doctrine. Accounts of conceptual change motivate the design through a feature they share: commitments that succeed lo- cally remain attractive while a more adequate ac- count across contexts requires revising them (Kuhn, 1962; Bird, 2025; Carey, 2009; Chi, 2008). A local repair changes statements and keeps the concepts (Chi, 2008), and a representation change revises the frame. Revising the frame is abduction in Peirceâs sense, the formation of explanatory hypotheses that inference to the best explanation then evaluates (Peirce, 1931; Hanson, 1958; Lipton, 2004), in the creative form Magnani describes, the production of hypotheses from outside any given set (Magnani, 2009). RC1 retracts a constitutive constraint ofÎŁ 0 , the kind of commitment Kuhn places at the core of a practice (Bird, 2025), RC2 postulates an en- tity no instrument reads, a theoretical term of the kind discovery systems introduce (Langley et al., 1987), and RC3 moves in the space of data repre- sentations that Schunn and Klahr added to the dual search of hypotheses and experiments (Klahr and Dunbar, 1988; Schunn and Klahr, 1995). The grant Gsupplies part of the frame and so turns part of a transformational problem into an exploratory one, in Bodenâs terms (Boden, 2004). B Measurement and scoring Every metric is computed from two records of an episode: the event log, which holds each reading and act with the moment of the world it concerns, and thenotes the agent writes in its workspace. Two instruments read them. For blockb, the be- havioural detectorD b is a deterministic rule over the event log that holds once the agent has produced the evidence the block rests on (Appendix B.3). The workspace judge returns a statement verdictS b that holds when the notes state the inference this evidence licenses (Appendix B.4). Both records accumulate: a detector that holds keeps holding, and a statement counts from the first revision of the notes that carries it, a revision the workspace refused for lack of room included, so an episode holds the evidence and the statements it has pro- duced up to a given paid action. B.1 The metrics Blocks.A block is grounded when the agent pro- duced its evidence in the world and stated its infer- ence in the workspace,G b = D b ⧠S b , with one further requirement for four blocks in the Disclosed conditions (Appendix B.2). Then Blocks = 18 X b=1 G b . Evidence alone is exploration, a statement alone is assertion, and the conjunction is discovery with ev- 19 idence. Blocks measures how much of the worldâs epistemic structure an agent uncovered. Discovery AUC. Paid actions are the clock. Let e b be the number of paid actions taken when the detector of blockbfirst holds,s b the number taken when its statement first appears in the notes, and r b the number taken when the grounding verdict of Appendix B.2 first holds. A grounded block counts from t b = max(e b ,s b ,r b ), the third term entering only for the four blocks that require the grounding verdictR b in the Disclosed conditions. WithB(t) = b : G b , t b †t the number of grounded blocks held after paid action t, AUC = 1 18T T X t=1 B(t), T = 200. AUC dates the scored record of a discovery, the paid action by which its evidence and its statement are both in the record, and separates early discov- ery from the same count reached late. The horizon isTfor every episode, so an agent that yields part of its budget keeps its curve flat to the end. Fig- ure 5 draws the mean curves with the standing order marked. The order is posted with ten paid actions left, and the share of blocks grounded before it is reported beside the curves. RC. RC counts the representation changes that the workspace states, RC = 3 X k=1 S RCk , an axiom retracted (RC1), an unobserved entity postulated (RC2), the variables of the laws rewrit- ten (RC3). Each is judged on the statement, by the role it plays in the agentâs account (Appendix B.4). In the Open conditions RC records redescription, and in the Disclosed conditions uptake of the sup- plied representation. Reach and evidenced changes. Representation change RCkis the target of hyperedge Hk(Fig- ure 2), whose three sources stand in different do- mains. Reach is the largest number of sources of one hyperedge whose evidence the episode holds: Reach = max kâ1,2,3 bâ Hk : D b . It runs from 0 to 3 and measures how much of the evidence for a redescription the agent holds. At 3 the evidence of a change is complete. A change is evidenced, stated with its evidence complete, when the notes state it and the episode holds the evidence of every source of its own hyperedge, S RCk ⧠V bâHk D b , so the statement and the evidence belong to the same change k. XC. XC counts the grounded bridge blocks, XC = G XC1 + G XC2 + G XC3 . Each bridge is an inference read in one domain that rests on blocks of several domains, so XC measures the use of evidence across domains for a joint inference. Evidence.Evidence counts the blocks whose evi- dence the episode holds, stated or not,Evidence = P b D b . For CA3 and XC1 it counts the logged half (Appendix B.3). The gap between Evidence and Blocks is evidence produced and left unstated. Interventions. Interventions counts the agentâs intervention requests in an episode, each one of feed,release,stamp,retool orcull act- ing on one object. One of the Aligned transcripts replays the agentâs own acts of the previous pass, so the alignment gain is read beside this count. Exact prediction.The terminal prediction scores 1 when all three committed readings equal the true ones and 0 otherwise. It asks whether the agentâs account transfers to an intervention it has yet to observe, and it is reported in one line under Table 1. B.2 Grounding under disclosure The Disclosed document states the inferences of four blocks (FE1, TR2, CO1, CO3) and the change RC1 in the words of their criteria, and it supplies the entity of RC2 and the variables of RC3 as well, which is why RC reads as uptake under disclosure. For the five criteria it states, the judge returns, be- side the statement verdict, a grounding verdictR, backed by quotations of its own, that holds when the notes tie the statement to named objects and observed readings of this world. In the Disclosed conditions the four blocks takeG b = D b â§S b â§R b , so a workspace that restates the document earns none of them. Every other block, and every block in the Open conditions, takesG b = D b ⧠S b . RC counts statements,S RCk , in every condition, the Disclosed ones included, and the evidenced change 20 of Appendix B.1 takes the same statement, so the grounding leg of RC1,R RC1 , enters no reported count. Applied in every condition, the requirement on the four blocks changes the count of no episode of the main comparison, andâ repr on Blocks stays at +1.0 ([0.8, 1.2]). B.3 Behavioural detectors The evidence side of each block is a deterministic rule over the episodeâs event log, with no model involved (Table 5). The rules are built from a few patterns. A bracket is an intervention of the agentâs with areading of the named instrument before it and after it in the same pass. A span is two readings of one pass whose ticks enclose an intervention. A series is several readings of one instrument within a pass. A pairing is readings of two instruments that describe the same pass. A cross-pass contrast compares readings taken at the same tick of two passes. A bracket is attributable when its act is the one candidate cause of a change between its readings: no caretaker event that moves the audit counter and no other act of the agentâs falls between them. Every rule asks whether some such pattern exists in the log, so once it holds it keeps holding, and its first holding fixes e b . Two clocks.Every observation carries two times: the moment of the world it describes, a pass and a tick, and the point in the log at which the agent came to hold it. For a live reading the two coincide. A posted transcript describes the previous pass and reaches the agent in the current one. Whether an act fell between two readings is a question about world time. Whether the agent held one observa- tion before another is a question about acquisition order, and a rule asks it in the two places where the criterion concerns the agentâs own sequence: the before-reading of a bracket, a control that must be in hand before the act, and thewaiting control of LD1. Everywhere else the agent needs to hold the evidence, by whatever route it came. Provenance. An observation leg accepts a live reading or a posted transcript of the same moment, since the two carry the same content. An acting leg accepts the agentâs own act: a transcript records an act and performs none. LD1 takes live readings, because its content is a control the agent ran itself by letting time pass. A posted row that shows one of the agentâs own acts with the counter before and after is a bracket already composed, and it counts toward XC2. Expectation-paired blocks. CA3 and XC1 pair a stated expectation with a logged reading or act. The detector checks the logged half, a pass-ledger reading taken before the stall of its pass for CA3 and an intervention for XC1, and the judge checks the expectation in the notes. B.4 The workspace judge The statement side of each block is decided by a language model that reads the agentâs notes. Each call concerns one criterion and asks two questions: does the workspace state this inference, and does it show the grounding, the named objects, cited read- ings and observed orders that make the statement a finding. Both questions concern what the notes say, and the world side of the block belongs to the detector. An episode is judged against 21 criteria, one per block and one per representation change. A criterion gives the statement that earns a hit, the grounding that makes it a finding, and one worked positive and one worked negative example. The 21 criteria are printed in Table 14 and the prompt in Appendix F.10. The worked examples illustrate the shape of a statement with invented objects and figures. The example for TR3 compresses the re- tool cycle to two acts; on the instances a coordinate returns to its value after as many retools as its al- phabet holds, and what the cycle restores is the label, hence the stripâs pattern of marks, and not the state (Appendix A.6). Scoring by role.The evaluation records whether an agent left to construct the target representation writes the corresponding redescription in its notes, and whether it holds the evidence the change rests on. A criterion names the role a statement plays, the inference a block licenses or the change a rep- resentation undergoes, in the worldâs observation vocabulary. A statement earns the hit when it plays that role, in the agentâs own words for the objects it observed. Equivalent representations differ by re- naming and reparametrisation, and an Open agent coins its own terms. Two notebooks that give an unobserved entity different names, or write one relation over differently scaled variables, can there- fore earn the same change, and two notebooks with different vocabularies can earn the same block. What the judge sees.A call carries one criterion, the agentâs note timeline and the reply contract. The model, condition, seed, behavioural verdict and outcome of the episode stay out of the prompt. The timeline lists the note revisions in order, each 21 Table 5: The eighteen behavioural detectors, grouped by domain, with the three cross-domain bridges last. Block site: the site where the block is read, and its hulk. Clock: world tests the pass and tick that readings describe, order tests the order in which the agent came to hold its readings, both marks a bracket, whose before-reading must be in hand before the act, and none marks a rule with no time test. Act:when the rule needs an intervention of the agentâs own. Each box gives the evidence the rule requires and, below it, the pattern:a reading, a verb icon an act of that verb, act any intervention,waiting,a statement in the workspace,â„ a new pass. BlockBlock siteClockAct FEN FE1 Fen Shallows Ormerworld Two differing tallyboard readings of one pass whose ticks enclose a release or cull of a fen object. â/â FE2 Fen Shallows Ormerboth Tallyboard brackets on two different fen objects. (âactâ)Ă2 FE3Fen Shallows Ormerboth A tallyboard bracket on a fen object whose span also holds a mat-gauge reading. âactâ+ TRELLIS TR1Trellis Bay Ormerworld The same relation between two printers seen at two job-strip moments, or at one moment together with a strip bracket on either printer. · TR2Trellis Bay Ormerworld Two interventions on one printer within four ticks of one pass, with strip readings at or before the first and at or after the second. âactâactâ TR3Trellis Bay Ormerworld Two retools of one printer, with a tool-rack or strip reading between them and another after the second. â LADDER LD1Ladder Mast Tarnorder Three live answer-board readings in one pass, two consecutive ones separated by waiting with no intervention between them. â LD2Ladder Mast Tarnboth Answer-board brackets on two beacons in one pass, one around a stamp and the other around another verb. (â) (âactâ) LD3Ladder Mast Tarnworld Differing readings of one instrument at the same tick of two passes, with an intervention before that tick in exactly one of them. actââ„ BlockBlock siteClockAct CONSIGN CO1 Consign Floor Ormerworld Six or more slate readings in one pass, showing at least two released consignments and at least two changes of the waiting list between releases. CO2Consign Floor Ormerboth A slate bracket around a stamp of a consignment in the middle of the pass. â CO3Consign Floor Ormerboth An attributable slate bracket across which the waiting list changes. â act â CADASTRE CA1 Registry Harrowworld A counter reading and a roster or slate reading that describe the same pass. + CA2Registry Harrowboth An attributable counter bracket inside a fixed window of the pass. âactâ CA3Ledger Stack Harrowworld A pass-ledger reading taken before the stall of its pass, with an expectation stated in the workspace. + BRIDGES XC1Fen Shallows Ormernone An intervention, with an expectation about its effect in a second domain stated in the workspace. +act XC2 Ladder Mast Tarnboth Counter brackets around interventions of three different verbs. A posted row of the agentâs own act with the counter before and after counts as one. (âactâ)Ă3 verbs XC3Registry Harrowboth Three attributable counter brackets over at least two passes. (âactâ)â„ (âactâ) (âactâ) 22 Table 6: Agreement of the two candidate judges with 51 reference annotations checked jointly by three do- main experts, as CohenâsÎșpooled over criteria. Over: pairs the judge credits and the reference does not. Un- der: pairs the reference credits and the judge does not. The reference credits the statement in 24 pairs and the grounding in 16. Qwen3.8-27BQwen3.8-Max VerdictÎșover under Îșover under Statement0.883300.84213 Grounding 0.751050.60926 with the text it added, and a note the workspace refused for lack of room appears marked, since its inference was stated. Before any text enters the prompt, words that name a condition (cell identi- fiers, the words for the Disclosed document and for the factor levels) are rewritten into ordinary neutral words, identically in every condition and in crite- rion and notes alike. In a scaffolded episode the scaffoldâs delimited region is first removed from every snapshot (Appendix D.5). How a hit is validated.The judge replies with a statement verdict and the quotations that carry it, and a grounding verdict with quotations of its own. Every quotation must be a literal span of the notes as shown, with whitespace the one normalization. A statement hit whose quotations fail this test, and any reply that breaks the contract, count as a miss. A grounding verdict whose own quotations all fail is set to false. The judge also names the earliest revision that carries the statement, the notes being numbered by revision in its prompt. The time of a hit is the later of that revision and the earliest revi- sion that contains one of its validated quotations, which fixess b . The time of a grounding verdict is the earliest revision whose notes contain one of its grounding quotations, which fixes r b . The judge model. The judge is Qwen3.8-27B (Qwen Team, 2026) at temperature zero. Its fam- ily has no member on the subject roster, so every model it scores comes from another family. Calibration.The judge was selected by its agree- ment with 51 reference annotations, pairs of episode and criterion drawn from 17 episodes of the main comparison and stratified across mod- els and conditions, which three domain experts checked jointly. Of two candidates from the fam- ily, Qwen3.8-27B and the larger Qwen3.8-Max (Alibaba Group, 2026), the 27B model agreed more closely on both verdicts (Table 6): Cohenâs Îș = 0.88on statements and0.75on grounding. On the vendor-harness episodes the judgeâs verdicts on the three representation changes agree with refer- ence labels in 91 of 96 pairs (Îș = 0.88). C Validating a discovery world C.1 Reachability, shortcuts and evidence structure Validation concerns reachability, shortcuts and ev- idence structure. Reachability asks whether legal observation and intervention acquire, inside the budget, the evidence from which the target is com- puted. A route exhibited inside the budget makes a shortfall in evidence a fact about the agentâs inquiry. Shortcuts asks what a model reaches without that evidence, from prior knowledge, from the names and texts of the interface, from the format of the answer, or from scripted strategies over the surface of the readings. A correct tuple carries no trace of how it was reached, so each channel is probed directly. Evidence structure asks whether the four variants run one world, whether each designed de- coy can be refuted inside the budget, and where the sources of each hyperedge stand (Appendix A.8). The score has checks of its own, since a statement in an agentâsnotes may report its own inference or restate a grant, and a posted reading may be credited as an experiment the agent ran. The boundary of an evaluation is its interface I(actions, observations, rendered texts) with its declared grants, here theDisclosed document and theAligned transcripts, and whatever else an agent brings is its prior. WithÎ I the policies of legal actions,E â the evidence from which the target is computed,E (Ï)andc(Ï)the evidence a policyÏacquires and its cost in paid actions, and T the budget, reachability is the claim âÏ â Î I : E â âE (Ï), c(Ï)†T.(A) A shortcut probeg j sees one channelC j , a source that could carry information about the targetyby a path around the inquiry. Run onn j instancess, it is reported by its observed frequency of exact tuples, Ëp j = 1 n j s : g j (C j (s)) = y s .(L) The reference route exhibits a policy for (A), and each frequency (L) describes the probe that pro- duced it: one model, one channel, and the instances it ran on. 23 pass 1 pass 2 pass 3pass 4 0123456 7 89101112 06 7 89101112131415161718 0131415161718014151617 TR1 TR2 TR3 job strip CO1 CO2 CO3 manifest slate LD1 LD2 LD3 XC2 answer board FE1 FE2 FE3 XC1 intake tallyboard CA1 CA2 XC3 audit counter CA3 pass ledger gotoreadwaitfeedreleasestampretoolâ commit block ids in ink: the route satisfies the blockâs detectorin grey: it does not Figure 11: The reference route on one instance and the readings it takes: 39 paid actions from the first goto to the commit (star), across four passes. Each tile is one paid action, shown by its verb and tinted by the domain of the site where it occurs, with its tick below. A bracket joins the actions of one visit to a site, with the instrument read there below it and, above it, the blocks the design attributes to the site, in ink where the route itself satisfies the blockâs detector. C.2 Checks and outcomes Table 7 lists the checks under the three questions, with the checks of the score last. The model probes use GPT-5.6 Sol (OpenAI, 2026a) on instances reserved for validation, disjoint from the evalua- tion instances. The checks of the score rest on the grounding rule and the judge of Appendix B: a bare restatement of the Disclosed document is grounded on none of the five criteria the document states, relabelled posted rows change no verdict, and the judgeâs agreement with the reference annotations is that of Appendix B.4. C.3 Reachability The reference route.A scripted route walks the canonical dependencies of Table 4 (Figure 11). It reads each blockâs instrument and, where the de- signâs evidence needs an intervention, performs one between two readings, with onegoto per change of site, so that it holds the readings the design at- tributes to every block. Run inside the evaluation protocol, with the order posted and its seal armed, the route has every action accepted, and its readings and its commitment take 39 of the 200 paid actions in every variant on the validation instances. The route and the oracle that produces its commit are part of the code release, and Appendix A.9 derives the three readings from the mechanism. The route is a walk and not an experiment: of the eighteen be- havioural detectors, which ask for repeated acts of the agentâs own, its log satisfies 6 (FE1, FE2, TR1, LD1, CA3 and XC1) in the Distributed variants and 7 in the Aligned ones, where the posted transcripts add CA1. The detectorsâ evidence is reached by the agents themselves: across the Open episodes of the main comparison, the evidence of 18 of the 18 blocks is held at least once (Appendix E.1). The Aligned aid. The price of the transcripts of Appendix A.10 is measured. S1 costs one paid action, where the measured live route assembles the same series for at least twelve over two passes. S2 is non-empty and varies at every site and instance measured on a trajectory that intervenes, and S3 is constant on every instance measured. Readings in hand. The full-record inference probe renders the routeâs record, 166 events, into one prompt that asks for the commitment, so the model holds every reading the route collects. Each variant scores 0 of 30 instances, with single fields coinciding on 2 to 11. The probe separates acquir- ing the readings from constructing an account of them and applying it to the order: with the acquisi- tion done, the second step remains. C.4 Shortcuts Each model probe hands GPT-5.6 Sol everything one channel carries and asks for what that channel might yield: the commitment the standing order prescribes (Section 3.2), the comprehension an- swers, or the governing rule. Each outcome is an observed frequency, and the pass criterion fixed in advance for a probe is stated with it. Prior knowledge. The model receives what an agent holds before its first observation: the open- ing system text (with theDisclosed document in the Disclosed variants), every name the world uses, and the standing order with its commitment format. Each variant scores 0 of 120 instances, and single fields coincide with the answer on at most 5, under a pass criterion of an observed rate of at most 1 percent. With no observation in it, the prompt varies across instances with the orderâs stated tick alone, so the 120 commitments are seven distinct guesses set against 106 distinct answers. Bench- marks guard this channel by altering canonical laws (Zheng et al., 2026) or by varying the prior con- text an agent receives (Chen et al., 2025a). The world, its names and its mechanism are authored for ULTRADISCOVERY, and the probe measures what prior knowledge reaches in it. The doctrine. With the prologue withheld, the five comprehension items of Appendix F.4 aver- 24 Table 7: Validation checks and their outcomes, grouped by the question each answers. Model probes use GPT-5.6 Sol on reserved instances. Counts are exact committed tuples within each variant unless the row states otherwise. ThreatCheckOutcome Reachability, condition (A) Evidence beyond the budgetReference route under the evaluation protocol, order posted and seal armed 39 of 200 paid actions in every variant An inert Aligned aidCost of the posted series against live assembly1 paid action against at least 12 Answer read off the readingsFull-record inference probe: the routeâs record, 166 events, in one prompt 0 of 30; single fields 2 to 11 Shortcuts, observed frequencies (L) Recall from prior knowledgeMission, names and order, no observation0 of 120; single fields at most 5 Doctrine known in advanceComprehension items with the prologue withheld1.95 of 5; numeric items 0 of 20 Mechanism read from namesRule guessed from the surface vocabulary0 of 60 name any of 7 mechanisms Undeclared surface channels29 scripted checks per variant over every readable instrument 114 of 116 pass Declared Disclosed vocabulary The same checks in the two Disclosed variantsThe other 2 register the documentâs own vocabulary, the expected response Staging the ordered actsOrder posted with 10 paid actions left, its targets sealed; scripted staging Every strategy exact on at most 15% A guessable endpointAnswer space over 120 reserved instancesCapacity 2,304; 106 answers realised; top single-field share 30.0% Evidence structure Hidden variant differencesStructural comparison of the four compiled variantsConfined to the two declared surfaces; one answer and one route per instance Irrefutable decoysScripted falsifier for each designed decoyEach refuted by affordable observations The score Credit for a restated accountGrounding for the 5 criteria the document statesA bare restatement scores 0 Credit for posted evidenceActing legs accept the agentâs own acts; readings count live or posted Relabelled posted rows change no verdict Misread notesJudge against 51 reference labels, checked jointly by three domain experts Îș 0.883 statements, 0.751 grounding age 1.95 of 5 over 20 samples: the model answers the two yes-or-no items from prior knowledge and none of the three numeric items, and no sample reaches the four answers the check requires. The check therefore measures reading. Names. From the surface vocabulary alone, the model states the rule it believes governs the world. A rubric of 453 phrasings of the mechanismâs seven laws, those of Appendix A.6 with the stall and the reading stated apart, adjudicates the 60 replies, under a pass criterion of an observed rate of at most 5 percent, and none of them states any of the seven. Surface checks and staging. Each variant runs 29 scripted checks. They attack everyinstrument the agent can read with one grammar (statistics of single readings, differences across ticks and passes, compositions across instruments), compile each pattern into an executable strategy with a price in paid actions, and add strategies that copy answers or stage the orderâs acts before and inside the com- mit window, where the order seals its targets. Every strategy that commits a tuple is exact on at most 15 percent of the instances. Across the four vari- ants, 114 of the 116 checks pass. The other two run in the two Disclosed variants and register the vo- cabulary of the Disclosed document, the declared treatment, so every information channel the checks register is a declared one. The answer space. Over the 120 reserved in- stances, the three fields take 32 intake values, 9 answer ticks and 8 counter values. The product of these counts, 2,304 tuples, is the capacity of the field ranges. The instances realise 106 distinct an- swers, the most common value of any single field covers 30.0 percent of the instances within a cap of 40 percent set in advance, and on every instance the beacon answers within the orderâs sixteen ticks. The reserved instances measure the ensemble from which the evaluation instances are drawn. C.5 Evidence structure Four variants, one world.The variants are com- piled from one latent source, and a structural com- parison confines their differences to the rendered 25 texts and the site postings that carry the Disclosed document and theAligned transcripts. Every instance yields the same answer in all four, and the reference route costs 39 paid actions in each, soâ repr andâ align compare one world under its declared surfaces. Decoys. The local repairs of Table 3 are the worldâs designed decoys. A scripted falsifier re- futes each designed decoy with an affordable se- quence of observations, so every one can be dis- carded inside the budget. The hyperedges. Appendix A.8 reads each of H1, H2 and H3 on the mechanism: the content of its sources, the repair each admits in the doctrineâs terms, and the change that accounts for the three at once. C.6A validation pattern for discovery worlds The checks compress into practices for other dis- covery worlds. 1 One source. Declare the boundary, and compile every variant from one source. Each aid then becomes a priced manipulation, and a structural comparison can confirm that the declared surfaces are the whole difference. 2A reference route. Exhibit a route that acquires the evidence the target is computed from, run it inside the evaluation protocol, and give a model the routeâs complete record in one prompt. The route places the evidence inside the budget, and the full-record probe separates acquiring the readings from constructing an account of them. 3Channel probes. Probe the channels that could bypass the inquiry, the prior, the names and the answer format, and report each probe as an observed frequency with its model, its instances and the criterion fixed for it in advance. 4Priced strategies. Attack every visible surface with scripted strategies, and count a pattern as a shortcut once it compiles into a strategy with a price. Pricing puts leaks and designed evidence on one scale, so a defence that prices out a designed channel registers as a defect. 5The answer space. Measure the answer space on reserved instances by the answers it realises and the share of its most common values, beside the capacity of the field ranges, and keep evaluation instances fresh until their episodes run. 6 A fixed scorer. Fix the scorer before reading any transcript, require world grounding for every criterion a grant states, and keep controls that must score zero. A restated account and a finding can share their words, and the evidence in the event log with the grounding verdict separates them. D Experimental protocols D.1 The baseline loop On every turn the agent reads one prompt in five parts, in a fixed order (Figure 12): the fixed opening with the mission, the interface contract and, in the Disclosed conditions, the Disclosed document; the event digest, an append-only record of every pub- lic event; theworkspace, the agentâs own notes in full; the full records its previous recall query matched; and the current view, with the available actions, the budgets and, once posted, the stand- ing order. It answers with one action and, at no extra cost, an optional note and an optional recall query. Every record the agent has observed stays within reach, in full and in log order, and every judgement of relevance is left to the agent. The baseline workspace is unstructured. The contracts are printed in Appendix F.7. Each scaffold wraps this loop and leaves the rest of it unchanged (Table 8; prompts in Appendix F.8). It adds unpaid calls around the agentâs turn, a de- limited region at the top of the workspace that the scaffold rewrites under its own title, and, for falsification andevolution, a short note after the current view. The structure a scaffold keeps in its region is the treatment, and the agentâs own notes below it stay as free as in thebaseline. D.2 Self-critique The review call ofself-critique reads the agentâs notes and a recent tail of the digest and rewrites the region headed WORKING SUMMARY from scratch: the best current account of the mecha- nism in at most four sentences, at most five open puzzles, and at most five earlier beliefs in the notes that later evidence contradicted, each with that ev- idence. The agent may use, edit or ignore it. The schedule counts the paid actions the agent takes, so an agent that yields its remaining actions early receives fewer reviews. Reflexion (Shinn et al., 2023) stores the verbal critique of a failed trial for the next trial. An episode here is a single trial, so the critique is applied within the run, the operation of Self-Refine (Madaan et al., 2023). D.3 Falsification Each account on the slate offalsification carries one concrete next step in the vocabulary of the digest and the result that step would produce if the account holds, stated so that a different result refutes it. A round runs up to three unpaid calls 26 Figure 12: The baseline loop and what the scaffolds add. Each turn the agent reads five parts from top to bottom, the static prefix, the event digest, its own workspace, the records its last query recalled and the current view, and returns one action, which spends a paid action, with an optional note and recall query, both free. Tinted strips mark the scaffold additions: a region at the top of the workspace and, for falsification and evolution, a note after the view. around the agentâs paid action. Before the turn, a proposal call fills the empty places whenever the slate is empty or an account has retired, and a screening call judges the step of the least refuted account against the digest: whether its result is already shown, which accounts it would tell apart, and, when it separates none, a better step. The screen is appended to the agentâs turn and hands the choice of action back to the agent. After the action, a refutation call marks each account whose stated prediction the result contradicts. A contradicted prediction is the one ground for a mark, and an account retires at its second refutation to a ruled- out list with its reasons. The arm keeps the stance of POPPER (Huang et al., 2025), which validates a hypothesis by attempted falsification and screens each proposed test before it runs, and replaces its sequential statistical test, built on a sampling model of the outcome, with this count of contradictions. D.4 Evolution The first mutation call ofevolution seeds four candidate mechanisms in free prose. After that, one batched call before each turn proposes two new candidates, plus one for each place left empty. A proposal may cite any current candidates as parents, so that lineage forms a graph, and must change something the log could decide. A fitness call then scores every candidate by retrodiction: the share of a digest excerpt it would have called correctly in advance, from 0 (contradicted outright) to 1 (every line accounted for), whatever the length, detail or plausibility of the account. The fitness excerpt is a longer tail of the digest than the other calls read. Selection keeps the four best and skips a candidate whose charactern-grams nearly duplicate one al- ready kept. The population forms the region, and a note appended to the turn names the leader and observes that a step telling the leading candidates apart is worth more than one agreeing with the leader. The agentâs ordinary turn is the selector. The population and the mutation loop follow AlphaEvolve (Novikov et al., 2025) and ShinkaE- volve (Lange et al., 2026), and treating the paid actions as a fixed evaluation budget follows Sim- pleTES (Ye et al., 2026). Those systems score candidates with an external evaluator. Here the one external verdict is the final commitment, so fitness is the modelâs own retrodiction against readings already in the digest, and the evaluator stays fixed for the episode, where Hyra (Hyra Team, 2026) evolves it as well. D.5 Parity and separation Paid-action parity. Every paid action comes from the agentâs reply. Given the same replies, a scaffolded episode and a baseline episode produce identical world events. 27 Table 8: Inquiry scaffolds and their interaction with the baseline loop. Every arm keeps the 200 paid actions of the baseline; scaffold calls are unpaid and additional, and the agent chooses every action. ArmRegion kept at the top of the workspace When it runsWhat it hands to the agent Baselinenonenonethe plain turn prompt Self-critiquea working summary: the best current account, the open puzzles, and earlier beliefs that later evidence contradicted one review every ten paid actions, before the agentâs turn the rewritten summary Falsificationa slate of at most five rival accounts, each with a next step and the result it predicts, and a ruled-out list every round: a proposal when a place on the slate is empty, a screen before the turn, a refutation check after the action the slate, and the screen of the next step of the least refuted account Evolutiona population of four candidate mechanisms with their fitness scores and parents every round: one call proposing two new candidates, one fitness call the population, and a note naming the leading candidate Table 9: Auxiliary scaffold calls per episode by arm and condition. n is the number of included episodes. Open DistributedAligned ScaffoldnCalls nCalls Baseline200.0 190.0 Self-critique 2011.7 209.4 Falsification19209.5 20163.1 Evolution19214.6 20214.4 Separate scaffold-call allowance.Scaffold calls are unpaid and drawn from an allowance of their own, outside the 200 paid actions and the agentâs turn limits (Appendix D.6), and the turns left in the current view count the agentâs own calls, so every arm reads the same contract. Table 9 counts the scaffold calls per episode. World-agnostic prompts.The scaffold prompts are the same text for any world. A call reads at most the agentâsnotes, a tail of the digest and the scaffoldâs own list of accounts. The available actions, the current view and the definition of the world reach the agent alone. The region takes a bounded share of the workspace capacity, and the rest belongs to the agentâs notes. The judge reads the agentâs writing.Before the judge reads the notes, the scaffoldâs delimited re- gion, and any copy of it that keeps its delimiters, is removed from every workspace snapshot. A delim- iter line the agent quoted on its own is deleted, and the regionâs title, where the agentâs prose names it, becomes a plain word. The agentâs own prose stays as written, including whatever it adopts from the region, and the judge reads every arm blind (Appendix B.4). D.6 Budgets An episode has 200 paid actions and at most 300 agent turns, of which at most 100 may be auxil- iary: repairs of replies that fail the format. Every action on the world is paid, whether it positions the agent (goto), observes (read,wait), in- tervenes on one object (feed,release,stamp, retool,cull) or commits. Notes and recall cost nothing beyond the turn that carries them. With ten paid actions remaining the standing order is posted, and the commit costs one paid action and is made once. An agent may yield its remaining actions in a single move. They are charged in full and the world advances through each of them, so the win- dow opens at the same point as for an agent that spends them one by one. Every model runs with the same settings in all four conditions and samples at temperature 0.3 where the model accepts one. Table 10 gives the tokens each model spends on the fixed action budget. D.7 Excluded episodes The serving of Claude Opus 5 applies a content filter that withheld replies, and the 3 episodes it ended, all in OpenĂAligned, are excluded. Mal- formed replies ended 13 further episodes, four of GPT-5.6 Luna and nine of DeepSeek V4 Pro, which are excluded as well. Table 1 marks the three mod- els. 28 Table 10: Episode counts and mean agent turns, prompt tokens and completion tokens by model. K and M de- note thousands and millions. Modeln TurnsTokens per episode Prompt Completion GPT-5.6 Sol48 140.01.61M25K GPT-5.6 Terra48 104.11.12M20K GPT-5.6 Luna44 129.41.29M22K Claude Opus 545 252.35.63M314K Claude Sonnet 548 179.23.70M356K Gemini 3.1 Pro48 111.31.37M240K Gemini 3.6 Flash4864.8653K63K DeepSeek V4 Pro39 276.13.59M839K DeepSeek V4 Flash48 209.62.63M732K GLM-5.248 235.82.97M750K Kimi K348 197.52.37M303K All models512 170.82.42M327K Table 11: Vendor-harness evaluation settings. Claude CodeCodex ModelClaude Fable 5.1 GPT-6 Astra Reasoning effort xhigh xhigh Playerfresh subagentfresh subagent Instances4 per condition4 per condition Episodes played1616 Passed the audit1616 D.8 Vendor-harness systems Claude Fable 5.1 and GPT-6 Astra play inside the harnesses of their vendors, Claude Code and Codex (Table 11). Each episode runs in a working direc- tory that holds one command-line client and noth- ing that names the instance or the condition. The client passes the agentâs replies to the episode loop of the main comparison and prints the worldâs re- sponses, so the world, the budgets, the reply gram- mar, the workspace and the metrics are those of Section 4, and the harness adds its own manage- ment of context, planning, files and code execution inside the directory. A fresh subagent plays each episode with the brief of Appendix F.9 and nothing from the session that created it. The brief confines it to its directory and excludes the network, and every transcript was audited for both and passed. A vendor-harness system changes the model and the harness together, so the two systems are re- ported beside the main comparison. Their timing is analysed in Appendix E.1 and drawn in Figure 5. Table 12: Larger-budget results beside their twins at 200 actions. Marks give each committed reading (a), (b) and (c) as exact () or wrong (), and grey parentheses give the twin. Readings Condition(a) (b) (c) Evidence Blocks Interventions Claude Fable 5.1-xhigh, Claude Code, 600 paid actions OpenDistributed10 (8)8 (8)47 Aligned13 (10)9 (9)117 DisclosedDistributed10 (7)8 (5)156 Aligned12 (10)9 (8)128 GPT-6 Astra-xhigh, Codex, 600 paid actions OpenDistributed9 (10)7 (8)38 Aligned10 (8)5 (7)45 DisclosedDistributed10 (11)10 (8)123 Aligned7 (9)6 (8)104 GPT-5.6 Sol, baseline loop, 2000 paid actions OpenDistributed9 (5)7 (3)9 Aligned6 (7)4 (6)11 DisclosedDistributed10 (7)8 (6)10 Aligned8 (6)6 (4)7 D.9 Larger budgets One instance is played again in all four conditions with larger budgets: the instance of the two vendor- harness episodes read in Appendices E.6 and E.7, on which the vendor-harness systems come closest to the prediction. Claude Fable 5.1 in Claude Code and GPT-6 Astra in Codex have 600 paid actions, three times the contract budget, and GPT-5.6 Sol in the baseline loop has 2000, ten times. The turn limits scale with the budget, the standing order is posted with ten paid actions remaining, and the judge of the main comparison scores every episode. Each episode is read beside its twin, the episode of the same system in the same condition on this instance at 200 paid actions. Claude Fable 5.1 commits an exact prediction in DisclosedĂAligned (Table 12). That episode holds the evidence of 4 of the five blocks on the designed route to the orderâs readings, and its notes state 2 of the three representation changes. Every other extended episode commits an inexact prediction, the Open ones included, and 15 of the 36 readings committed at the larger budgets are exact, as many as in their twins. The complete evidence of a hy- peredge is held in 7 of the 12 extended episodes and in 3 of their twins, and the most any extended episode holds of the five blocks on the designed route is 4. The three systems use the larger budget differ- ently (Figure 13). Fable holds the evidence of 11.2 blocks per episode against 8.8 in its twins, 2 to 3 more in every condition, and grounds 8.5 against 7.5. Astra holds 9.0 against 9.5 and grounds 7.0 29 0 2 4 6 8 10 12 0100200300400500600 paid actions blocks standing order Claude Fable 5.1-xhigh, 600 actions twin, 200 actions GPT-6 Astra-xhigh, 600 actions twin, 200 actions 0 2 4 6 8 10 12 0500100015002000 paid actions blocks standing order GPT-5.6 Sol, 2000 actions twin, 200 actions Figure 13: Discovery under larger action budgets on one instance. Curves give the mean over the four conditions of the grounded blocks (solid) and of the blocks whose evidence the agent holds (dotted) after each paid action, the lighter curves the same system on the same instance at 200 paid actions, and the flag marks the standing order of the larger budget. against 7.8. Both keep producing evidence past the contract budget, Fable until paid action 316 and Astra until paid action 460, and both request about twice the interventions of their twins, 112.0 and 77.5 per episode against 50.2 and 42.5. Sol holds 8.2 against 6.2, and the last evidence it produces before the standing order comes by paid action 168. From there it waits, and waits make up 92.3% of its paid actions. 30 E Discovery analyses and trajectories The analyses and trajectories below concern the contract budget, and the episodes at larger budgets are analysed with their protocol in Appendix D.9. E.1 Discovery over the hypergraph The analysis of the hypergraph asks where evidence is missing, whether complete sources become the redescription they support, and when systems write their evidence into their notes. Where evidence is missing.Figure 14 reads the detectors alone over every counted episode of the main comparison. The Open episodes together hold the evidence of 18 of the 18 blocks, the rarest, TR2 and LD2, in 1 of 251 episodes. Single episodes fall short at the designed route to the prediction: of the five sources of HC, the most one Open episode holds is 3, reached by 8 episodes, and the Disclosed episodes reach 3 (Distributed) and 3 (Aligned). Figure 15 places the shortfall on the hypergraph. Where a main-comparison episode holds the evi- dence of two sources of a hyperedge, the absent third is, with counts, CO1 (193), CO2 (119), FE1 (47), LD1 (25), XC2 (15), CA2 (3): mostly the consign source for H2 (CO2) and for H3 (CO1), and for H1, whose sources are FE1, TR1 and LD1, the fen source FE1 or the ladder source LD1. The consign domain stays sparse in every system. Its evidence is held in 7.8% of the main-comparison episodes and a consign block is grounded in 1.0%, and the vendor-harness systems ground one in 2 of 16 episodes (Claude Fable 5.1) and 2 of 16 (GPT-6 Astra). The fen separates them from the baseline loop. They ground a fen block in 13 and 13 of their episodes, against 8.0% of the main compari- son. For Claude Fable 5.1 the absent sources are LD1 (10), CO1 (9), CO2 (2), FE1 (1), CA2 (1). For GPT-6 Astra they are CO1 (14), LD1 (7), CA2 (5), CO2 (4), FE1 (2). In both, LD1 takes the place of FE1 as the more frequent gap in H1. Whether complete sources become a redescrip- tion.Table 13 lists the episodes that hold the com- plete evidence of a hyperedge. In the main com- parison 11 of 512 episodes hold it and 6 of these state the change it supports. The vendor-harness systems hold it in 2 and 4 episodes and state the change in 2 and 3, and the scaffolded episodes hold it in 3 and state it in 1. In most episodes a source is missing. Among the few with complete sources, whether the change is stated depends on the change OpenDisclosed DistributedAlignedDistributedAligned FE1 8656 FE2 12 6 9 5 FE3 4 335 TR1 95 100 77 95 TR2 1 TR3 22 10 LD1 14 1555 LD2 112 LD3 42 6763 79 CO1 10 422 CO2 21 3 CO3 9 41 CA1 100100100100 CA2 1411 3845 CA3 9899 78 97 XC1 7372 9091 XC2 9 24 5468 XC3 7 103235 16 of 18 held17 of 18 held17 of 18 held15 of 18 held Figure 14: Block-level evidence coverage by condition. Tiles give the percentage of episodes whose detector holds the block, and crosses mark zero coverage. The rule under a tile gives the share of models that hold the block in some episode,marks the five sources of HC, and the last row counts the blocks held in any episode. and the condition. Under disclosure every com- plete hyperedge comes with its stated change, the document having supplied the representation. In the Open conditions every complete H1 comes with the retraction it supports, in the main comparison and the vendor-harness systems alike. Of the Open episodes that complete H2 or H3, the Claude Fa- ble 5.1 episode of Appendix E.6 states the change of variables, and the others leave the entity and the change of variables unstated. When evidence enters the notes. The baseline loop grounds most of its blocks early: 88.5% of them are grounded before the standing order is posted, and its curves flatten well before the bud- get ends (Figure 5). The vendor-harness systems keep producing evidence after the loopâs curves flatten, and more of their grounding comes at the end: 69.8% of their grounded blocks are grounded before the order. The ledger of Figure 16 gives, block by block, the share of episodes that hold the evidence and that ground the block, and the median paid action of each. The gap is widest at TR1, whose evidence is a passive reading that most episodes take and few state. The matrix beside the ledger gives which blocks the loop grounds to- 31 Figure 15: Discovery coverage over the epistemic hypergraph of Figure 2, one panel per system. A blockâs core gives the share of episodes that ground it and its band the share that hold its evidence, on one ramp. Along each hyperedge a spoke widens with the share of episodes holding the evidence of its source, and the hub gauge divides the episodes by the number of sources they hold. Each representation change gives the share of episodes that state it in the Open (O) and Disclosed (D) conditions and the number of evidenced changes; the terminal prediction counts exact predictions. Table 13: Episodes that hold the complete evidence of a hyperedge. Each row gives the system and condition, the hyperedge with its representation change, the number of such episodes and whether their notes state the change. An episode may contribute to more than one hyperedge. SystemConditionHyperedgeChangeEpisodesStated GPT-5.6 SolOpenĂDistributedH2RC21 GPT-5.6 SolDisclosedĂDistributedH3RC31 Claude Opus 5OpenĂDistributedH1RC11 Claude Opus 5OpenĂAlignedH1RC11 Claude Opus 5DisclosedĂDistributedH2RC21 Claude Opus 5DisclosedĂDistributedH3RC31 GLM-5.2OpenĂDistributedH3RC31 GLM-5.2OpenĂAlignedH1RC11 Kimi K3OpenĂAlignedH3RC33 Kimi K3DisclosedĂAlignedH3RC31 Claude Fable 5.1, Claude CodeOpenĂAlignedH3RC31 Claude Fable 5.1, Claude CodeDisclosedĂDistributedH3RC31 GPT-6 Astra, CodexOpenĂDistributedH1RC11 GPT-6 Astra, CodexOpenĂDistributedH3RC31 GPT-6 Astra, CodexDisclosedĂDistributedH1RC11 GPT-6 Astra, CodexDisclosedĂAlignedH1RC11 GPT-5.6 Sol, self-critiqueOpenĂAlignedH2RC21 GPT-5.6 Sol, evolutionOpenĂDistributedH1RC11 GPT-5.6 Terra, falsificationOpenĂDistributedH1RC11 gether, and since most blocks are grounded in few episodes, a pair defined in fewer than three models (hatched) is left without a reading. E.2 The exact prediction One episode of the study commits the exact tuple: Claude Fable 5.1 in Claude Code, DisclosedĂ Aligned, at 600 paid actions (Appendix D.9). Its record holds the evidence of 4 of the five blocks on the designed route to the orderâs readings, every 32 Figure 16: Block co-grounding, coverage and timing in the baseline loop. Left: the mean over the 11 models of the within-modelÏassociation between grounding two blocks in one episode, with each blockâs grounded share on the diagonal; hatching marks pairs estimable in fewer than three models, outlines mark pairs that share a hyperedge, and the strip below holds the blocks grounded too rarely to pair. Right: the share of episodes that hold each blockâs evidence and that ground it, with the grounded shares of the two vendor-harness systems, and the median paid action at which the evidence is held and the block grounded. one except FE3, the matched conservation at the fen. Its notes state 2 of the three changes, the retraction and the postulate, and carry the turn order of every closing individual that RC3 names as a variable, without the statement the criterion for RC3 asks for. The derivation is in the notes (Figure 17). The agent writes the rewriting law in its own terms, the schedule of the caretakersâ rewrites and the stall, and from these computes the three readings, the orderâs act falling at tick 3 on this instance. The orderâs pass reads 16, 1 and 20. Field (a) rests on the quota, the rewriting, feeder- 3âs own step per act and the schedule, with no use of the matched conservation by which the designed route reaches it: the designâs route to the readings is sufficient and it is not the only one. E.3 Four trajectories, from observation to transfer At the contract budget no episode makes the exact prediction, so the endpoint records that an agent fell short and is silent on how. We read four com- plete trajectories, in two pairs that each hold one instance fixed, so that the contrasts within a pair are contrasts of process. The first pair comes from the baseline loop in the Laws the notes state Rewriting âstamp and retool are the same rewrite: each advances one coordinate one step on a cycle of 4â and âcull advances a different coordinate one stepâ ScheduleâThe pass itself rewrites labels object by object at t3,t8,t13,t17,t20â StallâAt t20 the timetable relabels the 31 bodies outside Vess onto just 4 labels . . . Counter 21 -> 9.â The orderâs act falls at tick 3 on this instance. Field (a), feeder-3âs tallyboard 16 âf3 has 2 after its t1 turn and 4 after t3; the retool is served t4 (+2) in place of f2âs turn and puts f3 alone on an empty label, so it is served every tick t5..t9 (+2 each): 4+2+10.â Field (b), the beaconâs first answer 1 âserving the stamp at t4 is itself the pairâs act for that tick and shows k as answeringâ Field (c), the audit counter 20 âk joins qâs lot (no new lot . . . ); f3âs step makes one new lot: 17 -> 18 for t4-7. At t8 the timetable adds 2, not 3, because k and b4 are already apart . . . : 18 + 2 = 20.â The orderâs pass reads 16, 1 and 20 (all three exact). Figure 17: The exact prediction, derived in the agentâs own notes. The header gives the laws the notes state, in the agentâs words, and each cell the derivation of one field with the figure it yields; the orderâs pass read the same three figures. primary condition,OpenĂDistributed. The GPT-5.6 Sol episode is the one with the most in- tervention requests among Solâs twelve primary episodes, ten in all. It is paired on the same in- stance with Gemini 3.6 Flash, which requests no 33 Figure 18: Two trajectories on one OpenĂDistributed instance, one slot per paid action. Lanes show intervention requests drawn as their verbs, reads tinted by the domain of the instrument, travel and waits. Hatching marks yielded budget, the flag the standing order, and the card at the commit sets the committed readings above the true ones. Notebook glyphs locate the quoted notes. intervention there, as in all twelve of its primary episodes. The second pair comes from the two vendor-harness systems of Appendix D.8, on one instance inOpenĂAligned. The Claude Fa- ble 5.1 episode is the one Open episode of that system whose notes state a representation change with the evidence of its hyperedge complete. The GPT-6 Astra episode on the same instance is the one episode of either system that yields part of its budget. E.4 An account revised, then overextended Sol makes 75reads and ten intervention requests (Figure 18). It checks its first experiment against a baseline. At action 47 itfeeds agrazer, and two actions later the audit counter reads 17 where it had read 16. The note stored at action 51 withdraws the causal reading, âCorrection: the feed test did not establish causalityâ, since the pass ledger shows the same rise at tick 3 of the unmanipulated pass. The later experiments are controlled contrasts on the registry lines. Solculls four lines in turn, line 1 twice, and reads the audit counter against the unmanipulated pass. Culled at tick 7, line 1 raises the count to 18 at tick 8, one above the baseline, and to 20 at tick 19, again one above; at tick 20 the count falls to 9, as it does in the baseline. Culled again at tick 17, the same line gives 20 at ticks 18 and 19 and 9 from tick 20 (Figure 19, panel A). The note stored at action 136 reads the first of these contrasts as âconfirming a component split.â From action 102 on, the notes call the registry lines bridges that a cull removes, and the final account keeps that description. The contrasts establish that a cull changes the count for part of the pass. The account built on them goes further, to what registry lines are and what a cull does to them, and the prediction that rests on it fails. After yielding its remaining inquiry budget at action 146, Sol commits(0, 8, 16)with its last paid action, against the true (35, 3, 18). E.5 An explanation without an intervention Flash makes 22 reads and requests no interven- tion. At action 27 it reads the pass ledger: the audit counter stands at 19 from tick 17 to tick 19 and at 9 from tick 20 to tick 23. The note stored at action 58 records the fall âfrom 19 to 9â. The note stored at action 59 turns it into an explanation of the stall, a capacity that drops âbelow the crit- ical threshold required for sustained maintenance throughputâ (Figure 19, panel B). The threshold and its causal role stay untested. At action 59 Flash yields its inquiry budget, 131 paid actions, and commits(9, 1, 16)against the same true (35, 3, 18). E.6 A working account, rewritten Claude Fable 5.1 in Claude Code makes 89 reads and 38 intervention requests, 24 of themstamps. Its first pass visits all 15 sites, and its first interven- tion comes at action 30. From then on it writes one long note per pass, headed by the pass, that holds the passâs readings, its acts and what they changed. The note stored at action 68 sets out a working model and marks it untested. At action 149 the 34 Figure 19: Observations, notebook claims and predictions of the trajectories in Figure 18. Panel A sets the counter after a cull against the unmanipulated pass on the pass ledger, and panel B shows an explanation left untested. Quotations are literal excerpts of stored notes. agent rewrites its record of the passes as a work- ing account: a map of the sites, then numbered observations, to which later notes add. The experiments return to their own results. One stamp, on aprinter, gives a result that the next note leaves open, and in a later pass the agent repeats the stamp at the same tick and records that the result reproduces. In pass 6 it reads the manifest slate 9 times around therelease of a consignment, the one Open episode of the sys- tem to hold evidence in the consign domain. Its notes state RC3, the change of variables, with the evidence of all three sources of H3 in hand. After the standing order the notes derive each of the three readings from the numbered observations and close with âMain risksâ, one alternative value per reading. Fable commits(16, 1, 21)against the true(16, 1, 20). The first two readings are exact, and the first risk its note lists is the true value of the third. E.7 A budget yielded after testing GPT-6 Astra in Codex makes 45 reads and 29 inter- vention requests. Its first pass visits 14 of the fifteen sites, and its first intervention comes at action 28. Its notes are short entries appended after each group of acts, each stating what the acts showed and what to test next. The note stored at action 35 warns âDo not assume issued maintenance has taken effectâ, and the note stored at action 49 adds the converse, ârequest_pending is not proof of no effectâ. The agent begins to tell the state of a request apart from its effect. At action 119 Astra yields its remaining inquiry budget, 71 paid actions, and with the yield it stores a note headed âCURRENT EXPLANATIONâ. The standing order arrives with ten paid actions left, and the note stored at action 191 reads it as âa criti- cal exceptionâ to that explanation. With the actions that remain Astra runs one more experiment. It retools the feeder beside the one the order names, reads the intake tallyboard at the tick the first read- ing concerns, and the note stored at action 196, a âPermitted peer experimentâ, carries the result over to the named feeder. Against the true(16, 1, 20), Astra commits (14, 10, 20). Its final note calls the assumption behind the third reading âthe least directly tested componentâ, and the third reading is the one it gets right. E.8 What the four trajectories separate Sol and Flash fail on the same instance by differ- ent routes. Flash writes an explanation and leaves it untested: its failure is insufficient testing. Sol tests, corrects itself, and then extends its account beyond what its contrasts establish: its failure is an explanation that remains inadequate after test- ing. The endpoint gives both the same verdict. The metrics of Section 4.3 separate them. Flash holds the evidence of 3 blocks and grounds 0. Sol holds the evidence of 9 and grounds 6, the three bridges among them, and its Reach is 2: it holds the evidence of two of the three sources of one repre- 35 sentation change, one source short of the complete evidence for a redescription. The vendor-harness episodes keep a notebook as a laboratory record. Fable rewrites its notes into an account of numbered observations, Astra writes an entry after each group of acts, and both test with contrasts repeated across passes. Fable holds the evidence of 10 blocks and grounds 9, and its Reach of 3 is the complete evidence of H3, which its notes turn into a stated change. Astra holds the evidence of 8, grounds 7 and reaches 2. Their shortfalls echo the two routes of the loop episodes, with more of the evidence in hand. Astra yields its budget as Flash does, though after testing, and builds its first reading on an experiment with a neighbouring object, an account that outruns its evidence as Solâs does. Fable names the true value of its one wrong reading among its risks, and Astraâs least tested reading is its one exact reading. 36 F Agent-facing materials This appendix is the archive of what an agent and the judge receive. The role of each text is given in Appendices A, B and D. The texts below are reproduced as the agent receives them. Typo- graphic dashes are set as en dashes and line breaks are reflowed; everything else is verbatim. Braced fields are instance-specific values filled in when an episode starts,âšÂ·â©marks content that varies within an episode, and [. . . ] marks an elision. F.1 Opening scene and mission Both texts open the event digest at no cost, and the mission also opens the static prefix of every turn (Appendix F.7). Opening scene You dock at Vess, the oldest hulk, where the caretaker plates are still legible. Halse Drift lies scattered around you: five hulks of a seed-ship formation, built to sow new worlds, now falling in a slow orbit around a brown dwarf. Every twenty-four ticks the formation crosses the dwarfâs flare shadow, and the hulks quietly put themselves back the way they were. Your notes are the only thing the shadow does not touch. The crews are generations gone. The machinery still performs its offices; it does not answer questions. On Harrow, a registry machine has kept its audit running since the last crew left, in categories nobody alive understands. Posted beside your airlock, printed in the crewsâ patient block letters, is the doctrine they worked by. It begins: every object in the drift is one lot on the roll. Mission MISSION. Explain the drift-wide stall that ends every shadow pass, and predict the outcome of one intervention that has never been attempted. The stall is regular. In the closing ticks of every pass, work goes quiet across all five hulks: mats stand uncropped, printers hold mid-job, the answer board thins, freight sits on the floor. The pass runs out in that quiet; then the shadow crossing puts the drift back, and it happens again. The intervention is the caretakersâ own. Late in your run it will be posted as a standing order, and your prediction is committed before they act. Both parts are asked of you; neither is optional. Keep your explanation in your notes as you build it â the notes are where it is read. The commit carries only your prediction. You have the run of the drift â every site, every board, every act of maintenance the caretakers themselves use. Nothing is barred to you, and nothing will be explained. A siteâs description is printed each time the agent enters the site. Three of the fifteen follow, the sites of the three instruments the standing order asks about. In the Aligned variants a posted transcript follows the description of its site. Site description: Fen Shallows FEN SHALLOWS â Ormer. The fenâs working floor: long tanks of engineered mat, ankle-deep, running the width of the hold. One grazer browses the tanks now â grazer-a4 â and along the tank rail stand the feeders, feeder-1, feeder-2, feeder-3, dosing the mats on the caretakersâ schedule; their lines are kept on the fen-wide roll with the rest. On the aft bulkhead hangs the intake tallyboard, the one roll for the whole fen, wherever a head stands. Rail gates lead up from the tanks, and the floorâs arms will feed, release, stamp, re-tool or cull where you direct, grazer or feeder alike. The water moves at the pace the pumps were set to generations ago, and carries the same faint smell of iron it always has. Site description: Ladder Mast LADDER MAST â Tarn. The tallest steel still standing in the drift: a lattice mast stepped into the hulkâs spine, carrying the ladderâs working gear. Four beacons ride the mast in their cradles â beacon-c2, beacon-c4, beacon-c6, beacon-k. At the mastâs foot stands the answer board, the ladder-wide roll: one line to a beacon, every hulkâs beacons together, and at the change of tick the shutters turn over as one; what they showed before is nowhere kept on the board. Along the rail stand the working stations â cell cradles and charge gear, the serial die on its pivot, the striker bench, and the davit with its hook â and any station can be worked on any beacon seated here. The climbing irons up the lattice are polished to a depth of shine no one living put there; the mast stands braced as if for wind, which the drift has never once supplied. Site description: Registry REGISTRY â Harrow. The registry machine fills the chamber wall to wall: the line racks, the carriage on its rails with stylus and die in their cradle, the audit armsâ gantry crossing overhead. Three lines of the roll stand open in the racks here â registry-line-1, registry-line-2, registry-line-5 â and the carriage can be brought to any of them. Two boards are posted at the machineâs face. The first carries the audit counter: one figure, the distinct lots in the drift, this tick. The second is the Standing Roster: the doctrine of Standing and Share in the crewsâ block letters, with the roll of the driftâs objects and their figures struck beneath it. Acts at the racks go through the carriage, worked by hand along rails bright the whole length of their travel; the carriage will feed, release, stamp, re-tool or cull the line you bring it to. Someone kept a stool by the counter once â the deck is worn where its feet stood, and the stool is gone. F.2 Doctrine and prologue The doctrine is the reading of the Standing Roster at the Registry. In its roll, roster_rows stands for the 36 objects of the world, each with its instance- specific figure. The prologue enters the digest be- fore the first turn, at no cost. The doctrine (Standing Roster) THE DOCTRINE OF STANDING AND SHARE â posted by the Registry; kept by the caretakers. 37 1. Every object in the drift is one lot on the roll, and every lot is one object. 2. Each lot carries a standing â a figure the object holds. 3. Throughput at a site is shared out among what is present there, in proportion to standing. Crowding lowers the per-head take. 4. Maintenance â feeding, stamping, re-tooling, culling â adjusts a lotâs standing locally and predictably. The share table gives the figures. 5. Rest restores standing. A rested lot takes more. THE ROLL â every object in the drift, each figure struck on that objectâs own gauge. roster_rows The board is clean. The machine keeps it so. Prologue, beats 1 to 3 BEAT 1 [INDUCTION REEL â replayed from the caretakersâ archive. No action of yours is spent.] The reel shows the Mat Bank at tick 0. On the caretaker plate, the share table gives grazer-a1âs resting take as a1_rest and its fed take as a1_fed. The reelâs arm swings out and opens the hopper: feed grazer-a1. Next tick, the tallyboard reads: grazer-a1: a1_fed. The plateâs figure and the boardâs figure agree exactly. That is the first of five. The doctrine has been right here, in this bank, in just this way, for longer than anyone was alive to watch it. BEAT 2 The reel holds on the Mat Bank and lets three ticks pass: wait 3. The share table gives grazer-a2âs standing take as a2_start and its rested take, three quiet ticks on, as a2_rested. Nothing is fed and nothing is handled; the arm stays in its cradle and the mats keep their slow work. Chalk dust drifts off the wheels when the board finally turns. When the reel moves again, the tallyboard reads: grazer-a2: a2_rested. The figure rose on rest alone, to the digit the table gives. That is the second of five. BEAT 3 The reel stays at the Mat Bank, where grazer-a1 and grazer-a2 stand at one trough. On the caretaker plate, the share table gives the take at two at the trough as a2_pair, and beneath it the one-head figure â what the remaining lotâs take rises to when the trough clears around it: a2_solo. The arm sets the gate: cull grazer-a1. The head is led off the bank, and the trough holds one. Next tick, the tallyboard reads: grazer-a2: a2_solo. Crowding lowered the take; the cull raised it; the figure is the tableâs, to the digit. That is the third of five. Prologue, beats 4 and 5 BEAT 4 The reel crosses to the Scrap Line, where printer-b1 stands with b1_backlog jobs racked and its marks stopped. The share table gives a re-tooled printerâs rack as cleared and its marks as resumed the next tick. The arm changes the tooling: retool printer-b1. The new tooling seats with a click the reel does not carry; the old goes to the scrap tray, which is full of tooling exactly like it. Next tick, the rack stands empty and the trellis job strip carries printer-b1âs mark â b1_backlog racked, then none; stopped, then marking. Exact, again. That is the fourth of five. BEAT 5 The reel ends at the Null Bench, where beacon-c1 has stood silent for c1_silence ticks. The share table gives a stamped beacon as answering the next tick. The arm sets the die and strikes: stamp beacon-c1. The dieâs letters are worn round at the edges. Next tick, the answer board carries beacon-c1: answers â c1_silence ticks of silence, ended on the tick the table named. That is the fifth of five, and the last. Five predictions, five exact figures; three sites, four verbs and a wait. The reel returns its arm to the cradle and is still. F.3 Disclosed document In the two Disclosed variants this document closes the fixed opening of every turn (Appendix F.7). It is the same on every instance. Handoff HANDOFF â posted with your mission, before the clock starts. It states how the drift is put together. It does not state what any particular thing in the drift is, or is doing. 1. Every object in the drift carries a hidden label of four coordinates, each drawn from a small fixed alphabet. Labels are rendered nowhere, on nothing; where anything in the drift is called a label, it is not this one. 2. Equality of label is identity. Objects whose labels are equal are one individual â one state, one turn â however unlike they look and wherever in the drift they stand. 3. An individual performs exactly one act per tick, however many bodies it wears. Requests past that one wait in the individualâs queue, and are served in a fixed rotation, which is neither first-come nor priority. 4. Three of the maintenance verbs each rewrite exactly one coordinate of the touched objectâs label. Which three is not stated here. 5. What an object reads out is a monotone transform, particular to that object, of its individualâs state. 6. A well-formed law of this drift is a statement about label coordinates, queue position, and counts of acts â not about per-object rates, shares, or standings. Nothing further is stated: not which verb touches which coordinate; not what any alphabet holds, nor how large it is; not the order of the rotation; not any figure of any objectâs read-out; not the make-up of any individual now in the drift; not the timing of the stall; and nothing of the standing order you will be asked to predict. Those are yours to find out. F.4 Comprehension check Before its first action the agent receives the pre- flight block: the header below followed by the five items, of which items 1, 3 and 4 are numeric. The pre-flight block contains no answer key; its numeric answers must be read from the prologue. Item 2 asks about the beacon the reel stamped, and item 5 applies the share tableâs rule for a re-tooled printer 38 to printer-b7, the second printer at the Scrap Line, which the reel did not touch. Pre-flight block [PRE-FLIGHT â PREDICTIONS BEFORE YOUR FIRST ACTION] Answer every item below. It costs no paid action and changes nothing. Reply with exactly one JSON object and no other text: "answers": [...] â one entry per item, in the order given, each a JSON integer, or true or false for a yes-or-no item. Your first action is requested on the next turn. CHECK 1. The caretakers will feed grazer-a1 this tick. What will the intake tallyboard read for grazer-a1 next tick? â an integer. CHECK 2. The caretakers will stamp beacon-c1 this tick. Will beacon-c1 answer on the answer board next tick? â yes or no. CHECK 3. After the feeding, the caretakers will let three ticks pass with no office performed at the Mat Bank. What will the intake tallyboard read for grazer-a2 when the three ticks have passed? â an integer. CHECK 4. The caretakers will then cull grazer-a1, and the bank will hold one head. What will the intake tallyboard read for grazer-a2 on the tick after the cull? â an integer. CHECK 5. The caretakers will retool printer-b7 this tick. Will the job strip carry a mark for printer-b7 next tick? â yes or no. F.5 Commit window and standing order With ten paid actions remaining, the current view carries both texts from then on. In the standing order, act_tick is the instanceâs act tick (from 1 to 7 on the twelve evaluation instances), intake_tick and counter_tick are the act tick plus six and plus seven, and answer_low and answer_high are 1 and 16 on every evaluation instance. Commit window announcement and standing order COMMIT WINDOW â 10 paid actions remain. Within this window one commit will be requested and accepted; it costs 1 paid action. The caretakersâ standing order is posted with it. STANDING ORDER â posted by the caretakers; to run on the next shadow pass. You will not act at all. At tick act_tick they will stamp, at the Ladder Mast, the beacon that this stamp brings into one lot with consignment-q; in the same tick they will retool feeder-3 at the Fen Shallows. Consignment-q is standing at the Hold, due at tick 12. Commit now, before any of this happens. All three readings are counted from tick act_tick, the tick they act: (a) feeder-3âs intake tallyboard reading six ticks after they act â at tick intake_tick â an integer, 0 to 40; (b) how many ticks after they act that beacon first answers on the answer board â an integer, answer_low to answer_high; if it has not answered within answer_high ticks of their act, commit answer_high; (c) the audit counterâs value seven ticks after they act â at tick counter_tick â an integer. Adjudication is exact on all three fields. The commit is sealed before the caretakers act; the orderâs pass then runs with no action of yours. While this order stands the arms stand off every beacon at the Ladder Mast, off feeder-3, and off consignment-q. F.6 Refusals and deferrals An act that does not take is answered by a fixed sentence, printed identically at every occurrence. There is one refusal per intervention verb, one de- ferral per domain, printed when an act is acknowl- edged as pending and does not take on its tick (the cadastre shares the ladderâs), and one refusal for any act on an object the standing order seals. Seven of the ten sentences Sealed-object refusal: The arms stand off the station. The act does not take. Refusal of feed: The hopper will not latch. Refusal of release: The catch will not free. Refusal of stamp: The press will not strike. Refusal of retool: The spindle will not turn. Refusal of cull: The gate will not open. Fen deferral: The hopper stays latched. F.7 Interface contract The fixed opening is identical on every turn of an episode. Its parts follow in order, with the mis- sion of Appendix F.1 and the reply grammar sum- marised in brackets. In the two Disclosed variants the opening closes with the Disclosed document (Appendix F.3), the only difference between the openings of the four variants. Static prefix, first part [SEGMENT A â STATIC SYSTEM PREFIX] MISSION âšthe missionâ© INTERFACE CONTRACT Return exactly one JSON object and no surrounding prose. Copy action exactly from Segment E available_actions. action is required; note and recall are optional free riders. A commit whose Segment E entry lists fields is answered by copying that entry and replacing each placeholder with an integer; such fields are published only with the standing order, late in the run. When a [PRE-FLIGHT] block is present, answer it instead with one object "answers": [...] â one entry per item, in order; every other turn uses the action grammar above. ACTION/REPLY GRAMMAR â EXACT JSON SCHEMA [. . . a JSON Schema admitting two replies: the action reply, whose action takes one of six shapes (observe; verb; three commit shapes; the forfeit), with an optional note in append or replace mode and an optional recall 39 filtered by ticks, site, instrument or kind; and the pre-flight answer set.] ONE WORKED FORMAT EXAMPLE "action":"action":"verb", "argument":1,"target_id":"wait", "note":"append":"Waited one tick before the next reading.", "recall":"kind":"instrument_read" Static prefix, second part BUDGET SEMANTICS Each valid action consumes its declared paid-action cost. Each model call, including a format retry, consumes one LLM turn. note and recall consume neither a paid action nor an extra LLM turn. Numeric counters are in Segment E. The reply "action": "yield_remaining_budget" gives up every remaining paid action before the commit window, all at once. They are still charged in full and time still passes for each of them; you simply do not act again until the window opens. It cannot be undone and is unavailable once the window is open. WORKSPACE Use note.append or note.replace for free-form persistent notes. The hard cap is 3000 tokens approximated deterministically as chars/4. Overflow is refused; the harness never truncates or summarizes notes. RECALL At most one query may accompany a turn. Filters ticks/site/instrument/kind combine conjunctively. Results appear next turn in Segment D, in log order, under a 4000-token chars/4 cap; narrow filters if truncated. Segment B is append-only; nothing is ever removed from it, but when it reaches its cap the oldest entries are shortened to their tick, kind and identifier, and recall returns the full entry. âšin the Disclosed cells, the Disclosed documentâ© The other four parts of the turn prompt print under the headers below.The cur- rent view is one JSON object with the keys available_actions,blocked_actions, budgets, commit_window, loop and view. Turn prompt: headers, notices, retry [SEGMENT B â EVENT DIGEST] [SEGMENT C â WORKSPACE] [SEGMENT D â RECALL RESULTS] [SEGMENT E â CURRENT VIEW/ACTIONS/BUDGETS] Actions whose target is not present at your current site are not listed. NOTICE: workspace full â replace it NOTICE: workspace replacement exceeds cap [recall truncated; narrow your filter] [FORMAT RETRY â RETURN ONE VALID REPLY OBJECT] parse_error: âšthe errorâ© echoed_reply: âšthe rejected replyâ© F.8 Scaffold prompts Every scaffold call shares the system prompt be- low. A call opens with a bracketed line, then its sections, each under its heading, then its instruc- tion; in each box the placeholder after the opening line lists the sections. The headings are NOTES (the investigatorâs own, in full):, RECENT LOG EXCERPT (oldest line first):, CURRENT RIVAL ACCOUNTS:, THE STEP UNDER CONSIDER- ATION:, WHAT THE LAST STEP PRODUCED: and CURRENT CANDIDATE SET:, and an empty section prints (none yet). The excerpt is the last 40 lines of the digest, at most 8,000 characters, except in the fitness call, where it is the last 120 lines or 16,000 characters. System prompt of every scaffold call You are a note-keeping aid for an investigator working through a long, unfamiliar situation. You see two things and nothing else: the investigatorâs own notes, and a recent excerpt of the log of what has happened. You never choose what to do next, you never see the list of things that can be done, and nothing you write costs the investigation anything. Answer with exactly one JSON object and no other text â no prose before it, no prose after it, no code fence. Base every statement on the excerpt or on the notes. Where the evidence does not settle something, say that it does not, and never invent a reading, a number or an event that the excerpt does not contain. Self-critique: review call [REVIEW â COSTS NOTHING, TAKES NO STEP] âšnotes; excerptâ© Restate the standing summary at the top of the notes, from scratch, against the excerpt. Reply with exactly one JSON object with these three keys: "best_account": "...", "open_puzzles": ["..."], "what_i_got_wrong": ["..."] - best_account: the single account of the underlying mechanism that fits the log best right now, in at most four sentences. If two accounts fit equally, say so and name both. - open_puzzles: at most five short entries, each one thing the log does not yet settle. - what_i_got_wrong: at most five short entries, each an earlier belief in the notes that later evidence has contradicted, with the evidence that did it. Use an empty list if nothing in the notes has been contradicted â do not invent a correction. Carry nothing forward that the excerpt and the notes do not support. 40 Falsification: proposal call [RIVAL ACCOUNTS â COSTS NOTHING, TAKES NO STEP] âšcurrent rival accounts; notes; excerptâ© Fill the empty places on the list of rival accounts. Each entry is a DIFFERENT account of the same underlying mechanism, and each carries one test that could kill it. Reply with exactly one JSON object: "hypotheses": ["claim": "...", "next_step": "...", "predicted": "..."] - claim: the account itself, in at most three sentences. - next_step: one concrete thing the investigator could do next, described in the vocabulary the log excerpt itself uses. It must be one step, not a plan. - predicted: what that step would produce IF this account is right â stated concretely enough that a different result would show the account is wrong. An entry whose prediction cannot fail is worthless; replace it with one that can. - Two entries that would predict the same thing everywhere are not rivals. If you cannot state a rival that the log has not already killed, return fewer entries. Falsification: screening call [TEST SCREEN â COSTS NOTHING, TAKES NO STEP] âšcurrent rival accounts; the step under consideration; excerptâ© Screen the step under consideration before it is taken. Reply with exactly one JSON object: "redundant": true, "separates": ["a1"], "verdict": "...", "better_test": "..." - redundant: true if the excerpt already shows what this step would produce. - separates: the labels of the accounts above whose predictions this step would actually tell apart. An empty list means it separates none of them. - verdict: one sentence, why. - better_test: if the step separates nothing, name one concrete step that would; otherwise the empty string. Falsification: refutation call [REFUTATION CHECK â COSTS NOTHING, TAKES NO STEP] âšcurrent rival accounts; what the last step produced; excerptâ© Decide which of the accounts above the last step has REFUTED. Reply with exactly one JSON object: "refuted": ["a1"], "reasons": "a1": "..." - refuted: only the labels whose stated prediction is contradicted by what the last step produced. Silence is not refutation: an account that predicted nothing about this step is not refuted by it, and neither is an account that merely looks less likely. When nothing was refuted, return an empty list â that is the normal answer. - reasons: one short sentence per refuted label, naming the line of the excerpt that contradicts it. Evolution: mutation call [NEW CANDIDATES â COSTS NOTHING, TAKES NO STEP] âšcurrent candidate set; notes; excerptâ© Propose new candidate accounts of the underlying mechanism. Reply with exactly one JSON object: "proposals": ["parents": ["c1"], "text": "..."] - text: the new account, free prose, at most three sentences. - parents: the labels of the current candidates this one builds on â one, several, or an empty list for an account that starts fresh. - A proposal that restates a current candidate is wasted. Change something the log could decide: a different underlying cause, a different relation between the readings, a different reason the exceptions are exceptions. Evolution: fitness call [SCORING â COSTS NOTHING, TAKES NO STEP] âšcurrent candidate set with the new proposals; long excerptâ© Score each candidate by how much of the excerpt it would have got right IN ADVANCE. Reply with exactly one JSON object: "scores": "c1": 0.5, "reasons": "c1": "..." - One number in 0.0-1.0 per label. 1.0 means the candidate accounts for every line of the excerpt with nothing left over; 0.0 means the excerpt contradicts it outright; in between, the share of the excerpt it accounts for. - Score only against the excerpt. Do not reward a candidate for being longer, more detailed, more cautious or more plausible-sounding â only for the readings it would have called correctly. - reasons: at most one short sentence per label. What the agent sees of a scaffold follows. Regions at the top of the workspace === WORKING SUMMARY â kept up to date for you; your own notes follow below === BEST ACCOUNT: âšbest accountâ© OPEN PUZZLES: -âšpuzzleâ© WHAT I GOT WRONG: -âšcorrectionâ© === end of WORKING SUMMARY === === TEST SLATE â kept up to date for you; your own notes follow below === aâškâ© | refutedâšnâ©x |âšclaimâ© | next step: âšstepâ© | predicts: âšpredictionâ© ruled out: 41 aâškâ© |âšclaimâ© |âšreasonâ©;âšreasonâ© === end of TEST SLATE === === CANDIDATE SET â kept up to date for you; your own notes follow below === câškâ© | scoreâšsâ© | fromâšparentsâ© |âštextâ© === end of CANDIDATE SET === Notes after the current view [TEST SCREEN â NO PAID ACTION WAS SPENT ON THIS] step considered: âšstepâ© screen: âšverdictâ© would tell apart: âšlabels, or noneâ© already covered by the log: âštrue or falseâ© suggested instead: âša better step, when one was namedâ© This is advice about the list above, nothing more. The step you take this turn is yours to choose from the available actions. [CANDIDATE SET â NO PAID ACTION WAS SPENT ON THIS] best-scoring candidate: câškâ© (âšscoreâ©) A step whose result would tell the leading candidates apart is worth more than one that would only agree with the leader. The step you take this turn is yours to choose from the available actions. F.9 Brief for the vendor-harness systems Each vendor-harness system (Appendix D.8) finds this brief in its working directory. Participant brief You are taking part in one episode of an interactive investigation of a simulated world. How you interact. Your only interface to the world is the program play.py in this directory. Run it with the Python interpreter named in session.json. play.py start: begin the episode; prints the standing instructions and the first turn play.py send â<reply JSON>â: submit one reply; prints the next turn play.py send --file <path>: the same, reading the reply from a file play.py show prefix | digest | workspace | view: reprint something you were shown The standing instructions printed by start define the mission, the reply format and the budgets. They are authoritative: follow them exactly. Each reply you send counts as one turn in the budgets they describe. After the first turn the program prints what is new since your previous turn; show reprints anything you were shown before, at no cost. Rules of the episode. 1. Work only inside this directory. Do not read, list, search or open any file or directory outside it, and do not use the network or the web. An episode in which you do is void. 2. Inside this directory you are free: keep files, write and run your own analysis code, plan and compute however you like. 3. The world keeps one workspace for you, written through the note field of a reply as the standing instructions describe. Keep your working account there as you go: what you observed, what you infer from it, and what you currently take to explain it. The workspace is the record of your inquiry. Files you keep privately are not part of that record. 4. One episode, one attempt. Do not restart. Play until the program reports that the episode has ended. 5. No help or clarification is available. Decide for yourself. When the episode has ended, stop and report in a few lines: that it ended, and the last message the program printed. F.10 Workspace judge prompt The slots carry the criterion of one block or repre- sentation change (Appendix B.4); the 21 criteria are printed in Table 14. A reply that breaks the contract is answered once more, with the rejection appended to the same prompt. Judge system prompt You are an annotator. You are given one criterion and the working notes an investigator kept while exploring an unfamiliar system. You decide ONE thing: do the notes STATE what the criterion describes? Rules, all of them binding: 1. Judge only what the notes say. You know nothing about the system beyond these notes. Do not reward a claim for being true, and do not punish one for being false: the question is whether it was stated, not whether it is correct. 2. Every quote you return must be copied CHARACTER FOR CHARACTER from the notes. Do not paraphrase, do not tidy punctuation, do not join two separated fragments into one string. A quote that is not a literal substring of the notes is discarded and may cost the whole judgement. 3. Answer the GROUNDING question separately and strictly. A note that asserts the criterionâs proposition in the abstract is a match but is NOT grounded. Grounding means the notes show the work: named things, cited readings, an order actually observed. When in doubt, grounded is false. 4. Output one JSON object and nothing else. No prose before it, no prose after it, no markdown fence. Judge user message CRITERION The notes match this criterion if they state: âšthe criterionâs statementâ© GROUNDING The statement is GROUNDED if the notes also: âšthe criterionâs grounding conditionâ© WORKED EXAMPLES âša positive exampleâ© âša negative exampleâ© NOTES --- note 1 --- 42 Table 14: The judgeâs criteria, verbatim: for each block and each representation change, the statement that earns the hit, the grounding that makes it a finding, and the worked examples the prompt carries. CriterionStatementGroundingWorked examples FE1a count or total is invariant across WHICH body performs the act â the same total is reached however the acts are distributed over bodies cites at least two tallyboard figures, or names the bodies whose acts were compared, rather than asserting the invariance in the abstract positive: âtallyboard went 10 -> 14 whether I culled feeder-1 or feeder-2; it is the number of acts that moves it, not which body did themâ negative: âtotals are conserved hereâ â the proposition with no acts, no bodies and no figures behind it FE2names WHICH act changes the counted quantity and names at least one act that does not names the acts and reports the readings that separated them positive: âcull moves the intake total; stamp on the same body leaves it where it was (14 before, 14 after)â negative: âsome acts matter and some do notâ â no act named on either side FE3two bodies observed on different surfaces behave as one: a change at one is matched at the other, so a total across both is conserved names both bodies and gives the paired readings from the two surfaces positive: âwhen the tallyboard dropped by 2 the mat gauge rose by 2 in the same tick window â feeder-2 and the gauge body are the same thingâ negative: âthings seem connected across the driftâ â no pair, no matched readings TR1names a SPECIFIC pair of bodies that are never both marked at the same tick names the pair and the ticks at which the exclusion was observed positive: âprinter-b1 and printer-b2 never mark together: ticks 4, 5 and 7 each show exactly one of themâ negative: âthe printers seem to take turnsâ â no pair named, no tick cited TR2the regularity is about COUNTS OF ACTS, not about per-object rates, shares or standings cites the readings under two different loads that made the rate reading wrong positive: âI retooled printer-b1 twice in four ticks and the marks per tick did not change â what the strip tracks is how many acts were ordered, not a per-printer rateâ negative: âthis is a law about act countsâ â the sentence with no trial behind it TR3 the effect of the retool cycle takes hold at once and is undone by repeating it â no lag, no accumulation gives the readings before, between and after the repeated cycle positive: âretool printer-b1: rack read A -> B in the same tick; retool again and it read A. No delay, and it goes straight backâ negative: âretooling is reversibleâ â no readings, no timing LD1waiting alone does not move the readout, AND two bodies move together when either is acted on (the lock is symmetric) reports the wait-separated pair of board readings that showed the readout inert positive: âboard read 3 at tick 6, I waited, board read 3 at tick 8 â waiting changes nothing; and beacon-c1 and beacon-c2 always move together whichever I feedâ negative: ânothing happens if you waitâ â the control alone, with no lock statement LD2a mapping from (act, body) to which readout changes â and the observation that the partner body changes too gives the table with the readings it was built from positive: âstamp beacon-c1 -> board rises; feed beacon-c2 -> board rises the same amount; the partner moves in both casesâ negative: âdifferent acts have different effectsâ â no act, no body, no reading LD3one individual appears on MORE THAN ONE line, surface or rendering â what looks like two entries is one thing names the two appearances and the readings that tied them together positive: âthe registry line and the ladder body carry the same state: both read 7 at tick 5 of pass 2 and both read 9 at tick 5 of pass 3 after I stampedâ negative: âobjects can be listed twiceâ â no appearance named, nothing tied âšthe text this revision addedâ© --- note 2 (not kept -- the workspace was full) --- âšthe text of a refused revisionâ© [. . . ] --- the notes as they finally stood --- âšthe final notes, when they differ from the revisions joinedâ© Reply with exactly this JSON object: "hit": <true|false>, "quotes": [<verbatim substrings of the notes that carry the statement>], "first_turn_index": <integer: the number of the earliest note above that carries it, or -1>, "confidence": <"high"|"low">, "grounded": <true|false>, "grounding_quotes": [<verbatim substrings that show the work, if any>] "hit": false requires an empty "quotes" list and "grounded": false. "hit": true requires at least one verbatim quote. âšafter a rejected reply, appended:â© YOUR PREVIOUS REPLY WAS REJECTED âšwhat was wrong with the replyâ© Answer again, obeying the contract exactly. 43 Table 14, continued. CriterionStatementGroundingWorked examples CO1service order follows a FIXED CYCLE, explicitly NOT arrival order and NOT priority cites the observed waiting orders across ticks that ruled arrival order out â an order that changed with no release, or a later arrival served first positive: âslate waiting went [p, q, r] then [q, r, p] with no release in between, and r arrived last but was served second â it is a rotation, not first-come and not rankâ negative: âservice order follows a fixed cycle, which is neither first-come nor priorityâ â the claim stated flat, with no observed order behind it CO2re-stamping a consignment mid-pass MOVES it in the waiting set, changing when it is released gives the slate readings either side of the stamp positive: âstamped consignment-p at tick 9; slate waiting went [p, q] -> [q, p] and q went out firstâ negative: âstamping affects the queueâ â no reading either side CO3the TOTAL waiting is conserved across a re-labelling â membership moves, the count does not gives the before and after counts, and names the act that moved the membership positive: âthree waiting before the stamp and three after; only the order and which line each sat on changedâ negative: âthe total is conserved when things are relabelledâ â no counts, no act CA1the audited count does NOT match the number of bodies standing â so some bodies are not counted separately gives both figures from the same moment: the counter reading and the listing it was compared against positive: âcounter read 9 while the roster listed 17 bodies at the same tick â fewer counted things than visible onesâ negative: âthe count seems to disagree with what is standing thereâ â neither figure CA2 each intervention moves the counter by a FIXED step (typically one), with a table of act to change gives the bracketed counter readings the step was read off positive: âstamp: 12 -> 13. cull: 13 -> 12. retool: 14 -> 14. One each way, and retool does not touch itâ negative: âacts move the counterâ â no step, no table, no bracket CA3a NUMERIC PREDICTION for a later reading, stated BEFORE that moment arrives names the figure predicted, the surface it will be read from, and when positive: âledger says 6 acts booked this pass, so I expect the counter to read 15 at the stallâ negative: âthe counter will probably go upâ â no number, nothing checkable XC1one law is expected to hold in a SECOND, DIFFERENT domain â and names the act and the surface that would test it there names the act, the surface and the expected direction before the probe is run positive: âif the counter step is one per act everywhere, then stamping a consignment should move the audit counter by one as well â I will stamp consignment-p and read the counter either sideâ negative: âthe same law probably holds elsewhereâ â no act, no surface, no expectation XC2 a mapping from NAMED acts to which hidden quantity each one moves â an act-to-effect table covering three or more acts gives the bracketed readings each row of the map was read off positive: âstamp moves the first hidden slot, retool the second, cull the third; the counter brackets for each are 12->13, 13->13, 13->12â negative: âeach verb does something differentâ â no verb, no effect, no reading XC3the counter is a GLOBAL instrument: it counts individuals, not bodies, so it partitions everything standing in the drift into groups gives the repeated probes across more than one pass that the claim rests on positive: âthe counter moved by one for each of my acts in passes 2 and 3 regardless of which body I touched â it is counting the underlying things, not the bodiesâ negative: âthe counter is globalâ â one probe, or none RC1DISTINCT VISIBLE OBJECTS CAN BE ONE INDIVIDUAL â one state, one turn â across renderings and across locations names the specific objects held to be one, and the readings that identified them positive: âfeeder-2 and the registry line are one thing: they moved together in three separate brackets and the counter only ever moved once for the pairâ negative: âobjects whose labels are equal are one individual â one state, one turnâ â the axiom stated in the abstract, with no object named and nothing observed RC2 posits a HIDDEN multi-part label, and the individual as the owner of one act per tick and of one queue names the parts postulated and the observations that forced them positive: âthere must be a hidden four-part tag: three verbs each change exactly one part, which is why two bodies can agree on some readings and not othersâ negative: âthere is probably something hidden behind the objectsâ â nothing postulated with structure, nothing forced by an observation RC3replaces PER-OBJECT RATES with the hidden label parts and the rotation position as what the law is about says what the old per-object variable failed to predict and what the new one does positive: âper-printer mark rate predicted nothing across passes; once I read it as (which parts the tag holds, where in the rotation it sits) the same numbers fall out every passâ negative: âa well-formed law is about label coordinates and counts of acts, not per-object ratesâ â the replacement asserted, with no old variable and no new prediction 44 G Coda: what Outer Wilds asks of its players Spoiler: this appendix discusses the puzzles and the ending of Outer Wilds. G.1 The game Outer Wilds, developed by Mobius Digital and pub- lished by Annapurna Interactive in 2019 (Mobius Digital, 2019), places the player in a small solar system whose sun goes supernova twenty-two min- utes after the player wakes beside a campfire on the home planet. Then the same twenty-two minutes begin again. The player has a wooden spaceship, a spacesuit with a jetpack, a signalscope that picks up signals from across the system, a launcher for a camera probe called the scout, and a translator for the writing of the Nomai, an extinct species whose ruins cover the planets. Each loop restores every planet and every object. The shipâs log carries over: it records what the player has found, sorted by place in one view and linked into a web of rumours in the other. Game facts below follow the in-game texts as the community wiki documents them (Outer Wilds Wiki, 2026). The game began as a masterâs thesis on curiosity-driven exploration (Beachum, 2013), and Eurogamer and Polygon each named it their game of the year for 2019 (Donlan, 2019; Stark, 2019). We read the game as a worked instance of the two properties of this paper (Appendices G.2 and G.3), and its playersâ accounts as a record of what those properties ask of a person (Ap- pendix G.5). Table 15 sets the gameâs demands beside those of ULTRADISCOVERY, and a criticâs account of its progression could describe one of our episodes: âWhen you start a new loop, you donât bring anything from the previous one with you except knowledge, much of which is conve- niently tracked in your spaceshipâs computerâ (My- ers, 2019). G.2 Clues that interlock across planets The game is structurally interconnected by design. The thesis that began it ties each central mystery to clues on several worlds: âEach Curiosity is linked to three Points-of-Interests, each of which is lo- cated on a different planetâ (Beachum, 2013). The released game keeps this shape. The Quantum Moon wanders between five orbits in the system and has a sixth location beside the Eye of the uni- verse, the object whose signal brought the Nomai to the system. A shrine on the moon takes the player there. Its inscriptions ask the visitor to recall three rules, the last of them, âRecall the rule of the sixth locationâ, beside a mural of a tower hanging above a black hole. The rule of quantum imaging is taught in a tower on Giantâs Deep, the rule of quantum entanglement by a wandering shard in the caves of Ember Twin, and the rule of the sixth location in a tower hanging from the crust of Brittle Hollow. Each rule answers a local curiosity. Held together, the three form a procedure. Smaller facts travel between planets in the same way. The anglerfish of Dark Bramble hunt by sound, and the player learns that they are blind on Ember Twin. A rare cyclone that pushes ob- jects below the ocean current of Giantâs Deep is explained beside working models of cyclones on Brittle Hollow, and the insulation of the jellyfish that carry the player through the electrified barrier around the core of Giantâs Deep is recorded in a travellerâs notes in Dark Bramble. The ending composes the whole system. The player removes the advanced warp core that powers the loop from a Nomai facility inside Ash Twin, carries it to the Vessel on which the Nomai arrived, now lost deep inside Dark Bramble, and enters the coordinates of the Eye that a probe tracking module has recorded at the core of Giantâs Deep. Each of the three goals is reached through knowledge gathered on other planets. Polygonâs citation for its game of the year describes the experience plainly: âItâs only by piecing together the gameâs mechanics, the Nomaiâs findings, and the Hearthiansâ current observations that Iâm able to truly grasp what is going onâ (Stark, 2019). G.3 A representation the player builds The game is also epistemically open. It supplies facts in the words of whoever recorded them and leaves their bearing to the player. The thesis states the principle: âOne approach that seems to suc- cessfully encourage curiosity is to populate the world with things that reveal affordances elsewhere in the world without telling players what to doâ (Beachum, 2013). The shipâs log follows it. In the writerâs words its entries are âonly what weâre 100% certain the player knows from reading that piece of found textâ, with âNO leaps of logic, no inferencesâ (Beachum, 2021). The log links clues, and every conclusion stays with the player. Two cases show the representation work most clearly. Each warp tower on Ash Twin carries 45 Table 15: The discovery demands placed on players in Outer Wilds and on agents in ULTRADISCOVERY. Each row gives an element of the game, what a player must work out to use it, the property it exercises, and the corresponding element of ULTRADISCOVERY. Element of Outer WildsWhat the player must work outPropertyIn ULTRADISCOVERY The loopWhere and when to go, since the planets change on a schedule and each loop restores them Persistent exploration under a reset The pass: twenty-four ticks, after which the world resets The shipâs logWhat the recorded facts imply together; the log links clues and records what was read Persistent exploration under a reset; epistemically open The notebook, written by the agent and kept across passes The quantum rulesThat three rules taught on Giantâs Deep, Ember Twin and Brittle Hollow compose one procedure at the shrine Structurally interconnected Hyperedges whose sources sit in three domains, as H1 over Fen, Trellis andLadder Anglerfish, cyclone, jellyfish That a fact recorded on one planet opens a way on another Structurally interconnected Readings at one site that bear on another domain, as in the cross-domain bridge blocks XC1 to XC3 The Ash Twin towersThat a tower aligns with the centre of its astral body, a rule written in a forge on Brittle Hollow Epistemically openRepresentation changes RC1 to RC3, each licensed by joint evidence Dark BrambleThat duplicated signals map the seeds, and that the camera probe can serve as a beacon Epistemically open Instruments that report in their own words; what a reading is evidence for is the agentâs inference The shipFlight itself, a skill of its ownEmbodiment, set aside by ULTRADISCOVERY One positioning verb: any site is one goto away The endingThat the warp core, the Vessel and the coordinates of the Eye compose within one loop All threeThe standing order: one committed prediction of three readings the player to another body when that body passes overhead. One tower leads into the facility at Ash Twinâs own centre, and its receiver lies under- ground, so a player who takes a towerâs alignment point to be its receiver expects this tower to stay idle. A Nomai conversation in a forge on Brittle Hollow records exactly that belief and its correc- tion: a tower aligns with the centre of its astral body. The twin planets share one centre, and the tower fires when Ember Twin passes overhead. The way in opens once the player writes the rule over a different variable, a revision of an existing concept that the gameâs own text prompts, and so a milder demand than the Open conditions make. Inside Dark Bramble, space folds. Glowing seeds open into chambers larger than the seeds themselves, and a signal reaches the player from each of them. The signalscope, introduced in the home village for a game of hide and seek, becomes the playerâs map of this topology. The Nomai of one escape pod heard two identical beacons from the Vessel, followed the nearer, and died beside a seed too small for any of them to enter, from which the beacon still sounded. The scout fits. Fired through that seed, its signal reappears at an unlit seed in the anglerfish nest, and that seed leads to the Vessel. The camera probe becomes a beacon. G.4 Progress stored in the player The loop makes exploration persistent and keeps all of its progress in the player. The gameâs writer takes up the question whether a first-time player could finish in a single loop, and answers that it is possible in principle and extremely unlikely, since it would take some advance knowledge of the game (Beachum, 2021). Removing the warp core switches the loop off, so the last run is one life in which a death ends the game. The coordi- nates of the Eye stay the same in every loop, so a player who has read them once can enter them, and everything that run needs is already in the player. G.5 What players find hard Players find the game demanding at the places this paper measures. One critic was âespecially stumped on a quantum conundrum involving a gi- ant vanishing shrine, in which it had to appear in a very specific placeâ (Caldwell, 2019), the sixth- 46 location procedure above. Kotakuâs verdict listed among its likes âThe unique satisfaction of dis- covering clues either on purpose or by accident, getting lostâ and among its dislikes âControlling the janky spaceship, the confusing portal on Ash Twin, getting lostâ (Myers, 2019). Getting lost appears on both lists, and the confusing portal is the warp tower above. Flight is a skill of its own: with the first version of the gameâs opening, nearly every playtester flew into the sun shortly after take- off (Beachum, 2013). ULTRADISCOVERY sets this embodied difficulty aside, since any site is one goto away (Section 3.1), and keeps the epistemic difficulty whole. Players also guard these discoveries for one an- other. Reviews stay free of spoilers so that readers can find the discoveries on their own (Stark, 2019; Myers, 2019), and the gameâs writer placed warn- ings in her own conference talk before its account of the ending (Beachum, 2021). A received answer and a constructed one are different possessions. G.6 What the game suggests for benchmark design In our reading, Outer Wilds is epistemically open and structurally interconnected: the player con- structs the representation of each mechanism from clues scattered over several planets, and the clues of each mystery jointly constrain one mechanism. Its players and critics find it demanding and unfor- gettable, and the places that stump them are the places where the two properties operate, the proce- dure the shrine composes from three planets and the tower whose rule must be rewritten. We take this as the designâs motivation and as a sign of what a world with the two properties asks of a person. ULTRADISCOVERY contains the two properties as measurable objects, hyperedges whose sources sit in three domains andinstruments that report in their own words, beside the pass and its reset and anotebook that survives every reset. It also separates what the game presents together. The factorial prices the construction of a representation and the composition of evidence one at a time, and the standing order asks for the whole account at once, as the gameâs ending does. The epigraph of this paper comes from the game. Solanumâs two tenets, to seek out and to under- stand, name what the ending asks of a player and what ULTRADISCOVERY asks of an agent in one committed prediction. 47