Paper deep dive
Faithful by Definition: Emotion Analysis via Natural Semantic Metalanguage Explications
Frank Xing, Erik Cambria
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 7/5/2026, 9:26:05 AM
Summary
The paper introduces a neurosymbolic interface for emotion analysis based on Natural Semantic Metalanguage (NSM). It proposes the Emotion Explication Schema (EES), a twelve-slot representation that maps input text to a closed vocabulary of semantic primes via a learned parser. A fixed, rule-based decision list (the mapper) then computes the emotion label from these explications, ensuring causal faithfulness. The system includes a per-line NLI-based verification interface to audit the parser's accuracy. Experimental results on the EmoExpl-1200 corpus show that while the parser's raw accuracy is comparable to black-box models, it provides a verifiable and inspectable decision basis for first-person, event-based emotion analysis.
Entities (7)
Relation Signals (4)
Frank Xing â authored â Faithful by Definition: Emotion Analysis via Natural Semantic Metalanguage Explications
confidence 100% · Frank Xing * and Erik Cambria #
Emotion Explication Schema â containsslots â semantic primes
confidence 100% · The Emotion Explication Schema (EES), a twelve-slot closed-vocabulary representation... organized into twelve typed slots
DeBERTa-v3-large â usedas â Verifier
confidence 100% · The verifier is a DeBERTa-v3-large NLI model.
Natural Semantic Metalanguage â providesbasisfor â Emotion Explication Schema
confidence 95% · We operationalize Natural Semantic Metalanguage (NSM)... as a two-segment definitional pathway
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Explanations for emotion classifiers are usually produced post hoc, with no guarantee that they reflect the computation behind the label. We present an explication interface for event-based emotion analysis. A parser maps the input text to an explication, a short script in the closed vocabulary of Natural Semantic Metalanguage organized into twelve typed slots, and a fixed decision list of rules transcribed from published semantic definitions computes the label from the explication alone. The faithfulness guarantee is therefore causal and definitional, while all empirical risk lives in the learned parser, which the per-line entailment interface makes auditable against the input. On crowd-sourced event descriptions, our fine-tuned parser reaches 0.33 accuracy and 0.48 selective accuracy on a small held-out set, suggesting that the interface trades insignificant accuracy difference to a black-box model for a verifiable, inspectable decision basis for first-person event-based emotion analysis. We also release EmoExpl-1200 with per-line verification metadata and the full rule set.
Tags
Links
- Source: https://arxiv.org/abs/2607.00661v1
- Canonical: https://arxiv.org/abs/2607.00661v1
Trouble viewing inline? Open PDF directly â
Full Text
48,599 characters extracted from source content.
Expand or collapse full text
Faithful by Definition: Emotion Analysis via Natural Semantic Metalanguage Explications Frank Xing * and Erik Cambria # * University of Reading # Nanyang Technological University z.xing@henley.ac.uk cambria@ntu.edu.sg Abstract Explanations for emotion classifiers are usually produced post hoc, with no guarantee that they reflect the computation be- hind the label. We present an explication interface for event- based emotion analysis. A parser maps the input text to an explication, a short script in the closed vocabulary of Natural Semantic Metalanguage organized into twelve typed slots, and a fixed decision list of rules transcribed from published seman- tic definitions computes the label from the explication alone. The faithfulness guarantee is therefore causal and definitional, while all empirical risk lives in the learned parser, which the per-line entailment interface makes auditable against the input. On crowd-sourced event descriptions, our fine-tuned parser reaches 0.33 accuracy and 0.48 selective accuracy on a small held-out set, suggesting that the interface trades insignificant accuracy difference to a black-box model for a verifiable, in- spectable decision basis for first-person event-based emotion analysis. We also release EmoExpl-1200 with per-line verifi- cation metadata and the full rule set. 1 Introduction Emotion analysis now informs content moderation, public- health screening, customer research, and the evaluation of conversational agents (Rajamanickam et al. 2020; Ma et al. 2020). In these settings a label alone is rarely sufficient; prac- titioners need to know why the system produced it, and regula- tors increasingly require the same. The dominant explanation formats do not meet this need. Post-hoc token attributions frequently disagree with the modelâs actual decision process (Jacovi and Goldberg 2020), and free-text rationales gener- ated alongside an answer can rationalize the computation instead of reporting it (Lanham et al. 2023; Madsen, Chandar, and Reddy 2024). The field has responded mainly by mea- suring unfaithfulness more carefully, by using counterfactual conflicts, or by optimizing explanations toward faithfulness proxies; all routes are post hoc (Cesarini et al. 2024) and leave the central guarantee missing. This paper pursues a constructive alternative for one task family: the prediction pathway itself serves as the explanation. We operationalize Natural Semantic Metalanguage (NSM), a linguistic-semantics program that defines word meanings through a closed set of semantic primes (Section 2), as a two-segment definitional pathwayËy = g(f Ξ (x)). A parser Preprint. Copyright with authors. frustration X feels something sometimes a person thinks something like this: I want to do something I canât do this because of this, this person feels something bad X feels like this Figure 1: A prototypical NSM explication that defines frus- tration entirely in semantic primes (Wierzbicka 1999). f Ξ maps text to an explication in a twelve-slot schema over the prime vocabulary, and a transparent mapperg, derived from the published definitions, computes the label from the explication and nothing else. The second segment is fixed by semantic theory; the first segment, which targets a cognitively and sensorily more primitive representation, is the only part that must be learned (Figure 2). Three developments make the design feasible now. LLMs can generate explications that respect the prime vocabulary (Baartmans et al. 2025); constrained decoding enforces the closed vocabulary so legality is a guaranteed decoder prop- erty; mature natural language inference (NLI) models make per-proposition verification practical at corpus scale; and appraisal-annotated corpora supply event descriptions with the cognitive granularity the schema requires (Troiano, Ober- lĂ€nder, and Klinger 2023). The paper makes three contributions. First, we introduce the Emotion Explication Schema (EES), a twelve-slot closed- vocabulary representation for emotional construals, and a definitional rule mapper over thirteen emotions. We are, to our knowledge, the first to combine an NSM-grounded closed vocabulary with instance-level emotion classification and a per-line verification interface, with three scoped properties: (i) the explication-to-label segment is causally faithful by con- struction (structural transparency: the label is computed from the explication alone); (i) the vocabulary is a closed, indepen- dently motivated prime set; and (i) the text-to-explication segment is verifiable line by line through entailment. We scope all claims to first-person, event-based emotions (Sec- tion 2). Second, we release EmoExpl-1200, an instance-level explication-annotated corpus over crowd-sourced event de- text: âI got the management positionâ parser f Ξ (learned) Emotion Explication Schema EXPER.IEVALFEEL-GOOD TRIGGER SOMEONE-DIDE-TARGET SELF AGENCY IKNOWNONE WANTWANTO-KNOW NONE REALIZ. HAPPENEDBODYNO TIMEBEFORE-NOWINTENS.PLAIN mapper g (fixed) emotion: pride NLI verifier: each line entailed by the text? FEEL-GOOD⧠SELF ⧠AGENCY=I â pride Figure 2: The explication interface, shown on a corpus instance. The parser turns âI got the management positionâ into the twelve-slot explication (centre); the fixed mapper fires the highest-priority rule whose conditions the explication satisfies, here FEEL-GOODâ§EVAL-TARGET=SELFâ§AGENCY=I (the bold slots), and returns pride, a positive self-evaluation distinct from the broad joy route. Only the parser is learned, so causal faithfulness of the mapper step holds by construction; each explication line is a proposition the verifier checks against the input text (dashed edge), with claims flowing from the explication to the verifier and entailment judgments flowing back along the two parallel arcs. scriptions derived from crowd-enVent (Troiano, OberlĂ€nder, and Klinger 2023), with per-line verification metadata. 1 Third, we document a replicable methodology for revising defini- tional rules under regression safeguards, including a case in which textual evidence forced a revision to be rolled back. 2 Background and Notation Natural Semantic Metalanguage. NSM analyzes mean- ing through roughly sixty-five semantic primes: simple, cross- linguistically attested concepts such as I, SOMEONE, DO, HAP- PEN, WANT, KNOW, FEEL, GOOD, and BAD (Wierzbicka 1996; Goddard and Wierzbicka 2014). Primes are posited as in- definable; every other word sense is defined by an explica- tion, a short prototypical script in primes. For emotions this yields a paraphrase of the prototypical eliciting situation. A pride-like state reads, in part, âI did something; I think this is good; I think something good about myself; I feel some- thing good because of thisâ (Wierzbicka 1999). Crucially, the explication is the definition of the concept, fixed by the theory independently of any classification task. For instance, the explication of frustration reads, in part, âI wanted to do something; I now know I cannot do this; because of this I feel something badâ (Wierzbicka 1999), which our schema records as WANT=WANT, REALIZATION=NOT-CAN, EVAL- UATION=FEEL-BAD (Figure 1). NSM has been applied at scale to emotion vocabularies organized by elicitor category and by cognitive template (Wierzbicka 1999); Appendix A reproduces these classifications for reference. Notation. We writex â Xfor an input text,c â Cfor an explication, andy â Yfor an emotion label, withYthe thirteen categories plus a reservedABSTAINsymbol. The system is the composition Ëy = g f Ξ (x) , f Ξ :XâC, g :CâY,(1) 1 Code, the EmoExpl-1200 corpus, and the rule set are available at https://github.com/fxing79/ebm. where the parserf Ξ is the only learned component and the mappergis a fixed function specified in advance. We call this a definitional pathway: all empirical risk lives inf Ξ , whileg is determined by the NSM definitions and never sees a label during training. The Emotion Explication Schema. We discretize the prime configurations relevant to event-based emotion into a finite product space. Definition (EES). The Emotion Explication Schema is the typed productC = Q 12 s=1 V s over twelve slots with finite value setsV s (Table 1), partitioned into event structure, ap- praisal, and expression. A rendererr :C â 2 P expands an assignment into a set of prime-vocabulary propositionsP, one per active slot. An assignment is legal iff every value lies in itsV s and every rendered line parses under the prime grammar. The schema is the engineering object the paper studies; it indexes prime configurations and does not extend the prime inventory. An instance pairs the twelve-slot assignment with a residue field, which records construals the primes cannot express and stays empty under full coverage, and a free-text notes field. 3 Related Work Faithfulness, disambiguated. âFaithfulnessâ labels sev- eral distinct questions, and our guarantee concerns exactly one. Explanation faithfulness asks whether an explanation reflects the computation behind a prediction (Jacovi and Gold- berg 2020); work here either measures the property (Matton et al. 2025; Madsen, Chandar, and Reddy 2024; Yeo, Sata- pathy, and Cambria 2025; Siegel et al. 2025) or optimizes toward it (Wang and Atanasova 2025; Chuang et al. 2026). We instead establish it by construction for one segment. The other senses are orthogonal: contextual and situated faithful- ness concern grounding in supplied documents (Ming et al. Group SlotValue set V s event EXPERIENCERI, SOMEONE, PEOPLE TRIGGER SOMETHING-HAPPENED, SOMEONE- DID-SOMETHING, NOTHING, NONE AGENCYI, SOMEONE-ELSE, NO-ONE, NONE REALIZATIONHAPPENED, NOT-HAPPENED, CAN, NOT-CAN, MAYBE, NONE TIME-DIR.BEFORE-NOW, NOW, AFTER-NOW appraisal WANTWANT, NOT-WANT, NONE EVALUATIONFEEL-GOOD, FEEL-BAD, NEITHER EVAL-TARGETSELF, OTHER, EVENT, OBJECT, NONE KNOWLEDGEKNOW, NOT-KNOW, MAYBE-KNOW, NONE expr. OTHERS-KNOWCAN-KNOW, NONE BODYYES, NO INTENSITYVERY, PLAIN, SMALL Table 1: The twelve-slot EES and its closed value sets. The full Cartesian space has|C| â 1.9Ă10 6 legal assignments, of which the mapper names thirteen emotion regions and can elaborate further. The schema targets first-person, event- based emotions; nested time, relational, and aspectual con- struals fall outside it. 2025; Si et al. 2026; Tang et al. 2025; Huang et al. 2025); rea- soning faithfulness concerns whether stated steps support an answer (Sui et al. 2025; Somov et al. 2026); confidence faith- fulness concerns verbalized confidence (Liu et al. 2025; Miao and Ungar 2026); reconstruction fidelity concerns sparse de- compositions (Oldfield et al. 2025; Gao et al. 2026); and one further sense concerns fidelity to oneâs own judgment under pressure (Zhao et al. 2025). Inside bottleneck models, struc- tural faithfulness (the label is computable only from the inter- mediate layer) is cheap, while input faithfulness (that layer truly describes the input) is the hard part. Concept-bottleneck LLMs inherit the first and neither guarantee nor verify the second (Sun et al. 2025); our design assigns the structural part to the definitional mapper and routes all input-faithfulness risk to the parser, where the per-line verifier measures it. Closed-vocabulary neurosymbolic emotion analysis, ap- praisal theory, self-conscious emotions, and NSM in NLP. A neurosymbolic tradition already pairs a closed primitive vocabulary with neural models (Cambria et al. 2026); The differences here are: (1) our intermediate layer is a propo- sitional script with a grammar rather than a flat polarity lexicon, and (2) we parse at the instance level rather than concepts lookup. Generic concept-bottleneck LLMs (Koh et al. 2020; Sun et al. 2025), however, satisfy neither. Ap- praisal theories (Ortony, Clore, and Collins 1988; Scherer 2009; Smith and Ellsworth 1985; Troiano, OberlĂ€nder, and Klinger 2023) are already propositional and rule-based; the difference is the linguistic motivation of our vocabulary and the per-line verifiability of discrete propositions. Shame and guilt are distinguished in psychology by global self-blame versus specific behavior-blame (Tangney and Dearing 2002; Tracy and Robins 2004); our mapper separates them with a single prime (OTHERS-KNOW versus AGENCY), a deliber- ate simplification we adopt for transcribability. Automated explication at the word-sense level is demonstrated by Baart- mans et al. (2025) and motivates our parser; mechanistic work finds appraisal concepts steerable in LLM emotion in- ference, with agency steering converting sadness into guilt (Tak et al. 2025), which motivates but does not validate our symbolic slot-flips. A probing experiment that tests whether the parserâs hidden states encode the slots is planned as future work. 4 Method The interface has three components over EES: a learned parser f Ξ , a fixed mapper g, and a per-line verifier NLI. 4.1 The parser f Ξ The parser maps text to a twelve-slot assignment. Under schema-constrained decodingf Ξ (x) â Cfor everyxand Ξ, so legality is one by configuration; we therefore report measured legality under free decoding (Section 6), where it is informative. We study a zero-shot prompted parser and fine- tuned variants (Section 5). All empirical risk concentrates inf Ξ ; the verifier of Section 4.3 audits it line by line. A set of written construal conventions governs how an annotator resolves slots that the text underdetermines (for example, a completed event narrated in the present tense takes TIME- DIRECTION=BEFORE-NOW; a slot leaves its default only on explicit textual support). These conventions are part of the annotation protocol, not of the guarantee. 4.2 The definitional mapper g Definition (Mapper).gis a priority-ordered decision list of thirteen rules(e k ,Ï k ) 13 k=1 , wheree k â Yis an emotion andÏ k is a conjunction of slot-value conditions transcribed from the published NSM explication ofe k . For an explication c,g(c) = e k for the smallestkwithc |= Ï k , andg(c) = ABSTAIN if no rule fires. The order is specific-before-general, so a particular pos- itive emotion shadows the broad joy rule, and abstention rate is itself reported. Table 2 lists representative rules; the full thirteen-rule list is in Appendix A. The self-conscious pair is instructive: shame and guilt share FEEL-BAD and a self-directed evaluation and are separated by a single prime, whether others can know (OTHERS-KNOW) versus whether the experiencer is the agent (AGENCY). Becausegreads onlyc, two properties hold by construc- tion. The label is invariant to anything not encoded inc; and for any single-slot editcâ c âČ the label changes exactly as the firing conditions dictate, so the Causal Consistency Rate equals1.0as an identity. This is true of any deterministic function ofc, so it is not by itself an XAI contribution. The scientific risk lives entirely in whethercfaithfully represents x, which the parser must get right and the verifier measures (Section 6). Figure 3 illustrates the identity: editing OTHERS- KNOW alone moves a guilt explication into the higher-priority shame rule. k e k condition Ï k ·reliefWANT=NOT-WANT, REALIZATION=NOT- HAPPENED, EVALUATION=FEEL-GOOD ·shameEVALUATION=FEEL-BAD,EVAL- TARGET=SELF,OTHERS-KNOW=CAN- KNOW ·guilt EVALUATION=FEEL-BAD,EVAL- TARGET=SELF, AGENCY=I ·prideEVALUATION=FEEL-GOOD,EVAL- TARGET=SELF, AGENCY=I ·trust EVALUATION=FEEL-GOOD,EVAL- TARGET=OTHER, KNOWLEDGE=KNOW ·angerEVALUATION=FEEL-BAD, AGENCY=SOMEONE-ELSE,EVAL- TARGET=OTHER last joy EVALUATION=FEEL-GOOD, TRIGGERâ SOMETHING-HAPPENED, SOMEONE-DID Table 2: Representative rules fromg, in priority order. Each Ï k is transcribed from a published explication; no condition is learned. The full list is in Appendix A. slotexplication cedit c âČ EVALUATIONFEEL-BADFEEL-BAD EVAL-TARGETSELFSELF AGENCYII OTHERS-KNOWNONECAN-KNOW g(·)guiltshame Figure 3: A single-slot counterfactual. Flipping OTHERS- KNOW satisfies the higher-priority shame rule, sogrelabels guiltâshame exactly as the definitions prescribe. 4.3 Per-line verification and rule revision The renderer turnscinto discrete propositions, and a natu- ral language inference model scores the entailment of each proposition by the input textx. The per-item verification score is the fraction of lines entailed above a threshold; it is an audit channel, calibrated against human entailment on a held-out subset before any gating use. Conventions and rules stay frozen between revisions; a revision passes only through an adjudication round, and every change runs a regression suite of thirteen canonical self-checks, a fifty-two-item pilot, and the slot-flip consistency test. One revision was rolled back when held-out evidence showed it stole visceral grief into the disgust route (Section 6); this episode is the basis of our replicability claim. 5 Experimental Setup Corpus and splits. The corpus draws on crowd-enVent (Troiano, OberlĂ€nder, and Klinger 2023), where each writer described an episode for a prompted emotion; the prompted emotion is the reference label, whose imperfection Section 6 quantifies. We annotate1,200items into EmoExpl-1200 with per-line verification metadata, and draw a156-item stratified subset (twelve per emotion) for two parallel passes: an LLM- annotator pass in which one of the most powerful models TierSourcen Pauthor pilot52 Aautomatic metrics, full batch1,200 Ssilver (Claude Fable 5, guidelines)156 Hhuman dual annotation99 goldadjudicated human gold98 evalheld-out routing set36 Table 3: Evidence tiers and sizes. Tier A reports label-free automatic metrics, S model agreement, H inter-annotator reliability. (Claude Fable 5) applies the written guidelines, and a gold pass by two trained human annotators applying the same guidelines. Because this annotator and the parser are both LLMs and may share construal biases, we treat the LLM- annotator pass as weak supervision, never as inter-annotator agreement; a cross-family pass with a non-Claude annotator (Llama-3.3-70B-Versatile via Groq) on the same156items returns the same slot reliability pattern at a lower scalar (mean α = 0.429vs. within-Claude0.491), confirming the model- agreement reading. The held-out evaluation set is36items drawn from the gold/silver core with a fixed seed; the leakage- free training pool is the remaining gold+silver explications (206clean pairs, or1,164with weak parser-preannotation augmentation). Table 3 lists every evidence tier. Parser configurations.The zero-shot parser is a prompted instruct model with constrained decoding. Fine-tuned parsers apply LoRA toLlama-3.2-1B,Llama-3.2-3B, and Llama-3.1-8B: rank16(α=32) for 1B and32(α=64) for 3B and 8B, dropout0.05, on attention and MLP projections; learning rates2/1.5/1Ă10 â4 ;3â5epochs; effective batch 16; 8B uses 4-bit NF4 QLoRA. The 3B and 8B runs add the weak-augmentation pool. Each scale trains and is evaluated on a single A100 in well under an hour. At inference we re- port results under free decoding to make the learned-legality claim falsifiable. Baselines. We compare four unconstrained families on the same36-item held-out set. (1) Black-box classi- fiers: sequence-classification heads on the same Llama-3.2- 1B/3.2-3B/3.1-8B backbones (LoRA, identical split), and a RoBERTa-large (Liu et al. 2019) classifier trained on the full corpus as an upper bound. (2)Appraisal pipeline: a logistic mapping from the twenty-one gold crowd-enVent appraisal dimensions (Troiano, OberlĂ€nder, and Klinger 2023) to emo- tion, trained on the corpus minus eval; it is given oracle appraisal at test time. (3) Concept-bottleneck LLM (Sun et al. 2025); (4) a SenticNet 9 lexicon baseline (Cambria et al. 2026), mean/max/min of four affective primitives plus polar- ity and Plutchik category counts, fed into a logistic head. The verifier is a DeBERTa-v3-large NLI model. Metrics. Accuracy against the prompted emotion, absten- tion rate, selective accuracy on routed items (El-Yaniv and Wiener 2010); explication legality; Krippendorffâsαfor reli- ability; the Causal Consistency Rate forg; and the per-line verification score. Each figure carries one tier label of Table 3. 1B3B8B parser scale 0.1 0.2 0.3 0.4 0.5 accuracy on held-out 36 ours (acc) ours (selective) black-box (acc) gold-expl (0.31) Figure 4: Accuracy against parser scale on the held-out set. The fine-tuned parser (red) improves with scale; a same- supervision black-box classifier (blue) is higher in raw accu- racy but the differences to ours diminish with scale; both are not statistically distinguishable from gold-explications. 6 Results and Analysis Each figure carries its evidence source. Automatic metrics (A) cover the full1,200-item batch. Silver figures (S) come from the156-item subset and report model agreement, never IAA. Gold figures (H) come from the human dual-annotation pass and its adjudication. Accuracy means agreement with the prompted emotion. 6.1 Main classification The zero-shot parser with constrained decoding reaches 38.3%accuracy at13.8%abstention on the full batch (selec- tive44.4%, legality100%, A), and36.5%on the156-item subset. Fine-tuning sharpens routing on the held-out36items (Table 4, Figure 4): the 8B parser routes twenty-five items at0.480selective accuracy and0.333overall, abstaining on eleven, and all three scales hold legality at100%under free decoding, so the parser learns to stay legal without a decoding constraint. Routing the human gold explications through the same mapper yields0.306(0.458selective); this is the gold explication routed throughg, not an upper bound. On raw accuracy the oracle-appraisal pipeline is numerically high- est (0.361), the same-scale black box next (0.389), and our parser0.333; the concept-bottleneck LLM reaches0.222and SenticNet 9 reaches0.250. We do not claim an accuracy win. Atn=36the exact binomial intervals on all of these overlap (Appendix C, Table 15), so no pairwise difference is statis- tically resolvable; the interfaceâs distinguishing property is the verifiable decision basis (and a higher selective accuracy, 0.480, on the items it routes). 6.2 Guideline executability and reliability The LLM-annotator explications reach53.8%accuracy at6.4%abstention (selective57.5%) against the parserâs 36.5%/11.5%/41.3%(Table 5, S). The17-point gap reflects partly prompting and training differences, not just guideline System (held-out, n=36) Routed Sel. accAccLegality Fine-tuned 1B (ours)14/360.3570.1391.00 Fine-tuned 3B (ours)28/360.3210.2501.00 Fine-tuned 8B (ours)25/360.4800.3331.00 Gold explication (routed)24/360.4580.3061.00 Black-box 1B36/360.2780.278n/a Black-box 3B36/360.3060.306n/a Black-box 8B36/360.3890.389n/a Black-box RoBERTa â 36/360.1110.111n/a Appraisal â 36/360.3610.361n/a Concept-bottleneck LLM36/360.2220.222n/a SenticNet 9 lexicon36/360.2500.250n/a Table 4: Main results on the held-out36-item set. Fine-tuned parsers (ours) route through the mapper and may abstain. Direct classifiers never abstain.â marks systems given extra information (RoBERTa trained on the full corpus; appraisal given oracle ratings). Annotation (156-item subset)AccAbstainSel. acc LLM-annotator (model)0.5380.0640.575 Parser (zero-shot)0.3650.1150.413 Table 5: Guideline executability (S). Routed through the same mapper, silver outscores the zero-shot parser by17points. The four-cell split is44both correct,40silver-only,13parser- only, 59 both wrong. executability: the cross-family pass on the same156items, replacing Claude Fable 5 with Llama-3.3-70B-Versatile, re- turns mean per-slotα = 0.429(vs.0.491within-Claude) and preserves the slot ordering, so silver is model agreement rather than IAA (Xing et al. 2020). Per-slot, raw agreement averages0.803across the twelve slots while Krippendorffâs αaverages0.491, or0.536excluding the variance-degenerate EXPERIENCER (Figure 5, Appendix B). The three slots that carry most routing decisions, EVALUATION, WANT, and AGENCY, holdα℠0.68, while five fall below0.45and de- fine the calibration priorities. Divergences are directional: the parser reads human-caused events as agentless happenings (34items on TRIGGER), marks present-tense retellings as NOW (24on TIME-DIRECTION), and over-specifies diffuse evaluations to persons (21 on EVAL-TARGET). Human pass and adjudication.Two trained annotators in- dependently labeled99items from the highest-priority review stratum. Raw agreement averages0.809whileαaverages 0.384, replicating the inflation pattern. Determinacy splits by slot type: four event-structure slots holdαbetween0.67 and0.78, while five collapse below0.30(KNOWLEDGE 0.04, TIME-DIRECTION0.19, INTENSITY0.23, EVAL- UATION0.25, plus degenerate EXPERIENCER). Conflicts are one-directional. Routed through the mapper, the human explications abstain on roughly80%of items (the two anno- tators agree on abstaining for64of99), exposing a routing frontier rather than estimating accuracy. The interface pre- filled slots from parser predictions, and each annotator kept EVALUATION AGENCY WANT EVAL-TARGET REALIZATION TRIGGER TIME-DIR OTHERS-KNOW INTENSITY BODY KNOWLEDGE EXPERIENCER 0.0 0.2 0.4 0.6 0.8 Krippendorff's α silver (n=156) human (n=99) Figure 5: Per-slot Krippendorffâsα: silver model agreement (blue,n=156) and the first human pass on the adversarial stratum (yellow,n=99). Event-structure and core appraisal slots reachα℠0.67; KNOWLEDGE and the variance-degenerate EXPERIENCER collapse. Exact numbers appear in Appendix B. those defaults on0.85and0.89of slots against0.81between themselves, so the first-batch figure measures anchored veri- fication. A convention vote resolved the directional conflicts into a gold standard of98items under one principle: a slot takes a marked value only on explicit textual evidence. EVAL- UATION took a calibrated exception: a bare negative life event (e.g. divorce) maps to FEEL-BAD; we acknowledge this encodes a mild world-knowledge prior. The adjudicated gold abstains on85of98items, confirming the routing fron- tier. 6.3 Distributional evaluation Per-emotion accuracy ranges from joy (0.70) down to trust (0.09); the prominent off-diagonal flows in Table 6 localize the failures. The largest single source is a slot-rule interaction: fourteen silver explications carry KNOWLEDGE=NOT-KNOW from discovery framings, thirteen route to surprise, and the slot has the lowestαin the corpus. The adjudicationâs evi- dence gate on KNOWLEDGE targets exactly this. A second localization concerns the mapper:28.5%of parser explica- tions satisfy more than one rule, resolved by priority; most collisions are specific-over-general by design, and one re- flects a true overlap, shame ahead of guilt on66items. On the adjudicated gold the tie rate is zero. Typed abstention, priority-order sensitivity, and a distributional evaluation that scores the mapper as a 13-way distribution over fired rules (priority-decay strictly dominates single-winner, lifting top-3 by six points at n=1,200) are reported in Appendix C. 6.4 Verification and counterfactual analysis The entailment verifier assigns a mean per-item score of 0.201, with487of1,200items at zero (Appendix D, Fig- ure 7), while selective accuracy stays at44.4%across this range (A). Verification and routing respond to different sig- nals: the verifier is conservative on prime-style minimal sen- tences, a known cost of the closed vocabulary, while routing depends only on slot values. We therefore treat the score as 0.00.20.40.60.8 selective accuracy joy no-emotion relief shame surprise anger boredom disgust fear guilt pride sadness trust mean 0.37 Figure 6: Selective accuracy by prompted emotion (A, n=1200). Recovery is strong for joy, anger, and no-emotion and weak for trust and fear; the dashed line is the overall mean (0.38). an audit channel and defer any gating decision to calibra- tion. The Causal Consistency Rate equals1.0as an iden- tity (Section 4); single-slot flips reproduce the documented contrasts, including guiltâshame through OTHERS-KNOW alone (Figure 3). To probe whether the contrast survives at the text level, we built a42-pair counterfactual benchmark over six slot-edit families (Table 7) where each pair(x,x âČ ) rewrites the input to flip a designated slot. The parser flips the intended slot on76%of pairs and the definitional path- way carries the contrast to the expected emotion on36%. Clean-flip is near zero by design: a real text rewrite shifts surface signals across multiple slots, so the parserâs inci- dental updates are expected, and the structural-transparency property still localizes the mapper-side causal effect under gold irouted jn % of i trustjoy4650.5 pridejoy4447.3 guiltshame3436.6 surprisejoy3436.6 reliefjoy3032.6 shameguilt2527.5 disgustanger2324.7 disgustsurprise2223.7 boredom no-emotion 1718.5 fearsurprise1617.2 Table 6: Prominent off-diagonal routing flows on the full batch (A,n=1200); the complete row-normalized confusion matrix is Figure 8 in Appendix D. Positive emotions collapse toward the broad joy route and the self-conscious pair swaps in both directions; both patterns trace to the lowest-reliability slots. Familyslot flipnslot label causal F1OTHERS-KNOW7 0.430.290.00 F2WANT + REALIZATION7 1.000.570.14 F3KNOWLEDGE7 0.860.140.00 F4AGENCY + EVAL-TARGET7 0.710.430.00 F5TIME + REALIZATION7 1.000.570.00 F6EVAL-TARGET only7 0.570.140.00 allâ42 0.760.360.02 Table 7: Counterfactual benchmark. Slot and label are the rates at which the parser flips the targeted slot and routes both endpoints to the expected emotions; causal requires both, with clean-flip gating it to near zero by design. g. Per-family rates track the slot reliabilities of Figure 5: the high-αmulti-slot relief recipe (F2) leads at label-flip0.57; KNOWLEDGE (F3) and the EVAL-TARGET-only swap into a residue region (F6) bottom at 0.14. 6.5 Inherent disagreement and ground truth The silver annotator flagged32of156items (20.5%) as admitting two defensible construals,14(9.0%) as unrecover- able under masking, and three as vicarious; these rates bound end-to-end accuracy on this corpus well below100%. A low αmay raise a foundational objection: if trained annotators disagree this much, perhaps the items have no determinate label (Plank 2022; Pavlick and Kwiatkowski 2019; Uma et al. 2021). Two replies are decisive. First, the corpus places ground truth at the source (the writerâs self-report about an ex- perienced episode), so disagreement among readers measures recoverability from a degraded text rather than indetermi- nacy. Second, determinacy is slot-heterogeneous: the median disagreeing pair differs on two of twelve propositions, so a label conflict such as shame versus guilt decomposes into ten agreed and one or two contested propositions, a distinction a black-box label cannot register. Where two construals survive adjudication, the mapper can emit both with the pivotal slot that separates them (Röttger et al. 2022; Uma et al. 2021). 7 Discussion Scope of the guarantee.The faithfulness guarantee covers the explication-to-label segment only; the parser remains learned and carries all empirical risk. The phrase âfaithful by definitionâ must always be read with the segment qualifier attached. Verification, not the guarantee, does the empirical work, and our verifier is conservative on prime-vocabulary sentences, which depresses raw scores; we therefore report it as an audit signal to be calibrated, never as a gate. How to read the reliability numbers. Two biases shape the first human pass and we flag both. The stratum is adver- sarial by selection, which lower-bounds the agreement of a random batch; the interface also pre-filled slots from parser predictions, anchoring annotators and inflating agreement above blind annotation. Because one bias deflates and the other inflates, the per-slot determinacy pattern, stable across both, carries the interpretive load rather than the scalar α. Theory dependence and expressivity.The contested uni- versality claims of NSM are not load-bearing here: we use only the closed vocabulary, the propositional format, and the availability of independently motivated definitions. The frozen schema cannot express every construal: remembered fear needs nested time, vicarious pride a relation slot, termi- nated relief an aspectual distinction; we flag such instances, report the residue rate, and restrict the headline claims to event-based emotions. Future work. The current evidence base is constrained by annotator hours, compute, and API budget; what fol- lows is the planned, not the completed, programme. (i) A blind, random-stratum human pass would replace the present anchored, adversarial estimate. (i) Parser probing would test whether slot decisions are driven by the documented features or by surface artefacts. (i) Scaled-up annotation would tighten the held-out interval beyondn=36. (iv) Cali- brating the verifier against human entailment judgments turns the audit channel into a usable gate. (v) A controlled simu- latability study (Lyu, Apidianaki, and Callison-Burch 2024) would test verifiability head-on. (vi) A cross-lingual evalua- tion would test whether slot-level construals transfer across the languages in which the primes are attested. 8 Conclusion We present an explication interface for event-based emotion analysis, in which faithfulness is relocated from a post-hoc measurement target to a structural property of the pipeline architecture. On crowd-sourced event descriptions, we fine- tuned an 8B parser that reaches 0.33 accuracy and 0.48 selec- tive accuracy on the held-out set (n=36), statistically indis- tinguisable from a same-scale black-box emotion predictor (0.39 acc. and ~0.50 sel. acc.) or human reconstruction (0.39 acc. and 0.48 sel. acc.). Without compromising accuracy, however, our interface makes the analytical steps of emotion auditable line by line. We release the schema, the mapper, and EmoExpl-1200 with per-line verification metadata, together with the rule-revision protocol that produced them. In future work, probing the parserâs hidden states for slot encoding would test whether the interface reflects the modelâs internal working mechanism. References Baartmans, R.; Raffel, M.; Vikram, R.; Deringer, A.; and Chen, L. 2025. Towards Universal Semantics With Large Language Models. arXiv:2505.11764. Cambria, E.; Mao, R.; Zhang, X.; Xiao, L.; Shen, T.; and Anand, A. 2026. SenticNet 9: Generative Commonsense for Emotion AI via Conceptual Primitive Discovery and Time Shift Mechanism. IEEE Transactions on Computational Social Systems, 13(3): 4086 â 4095. Cesarini, M.; Malandri, L.; Pallucchini, F.; Seveso, A.; and Xing, F. 2024.Explainable AI for Text Classification: Lessons from a Comprehensive Evaluation of Post Hoc Meth- ods. Cogn. Comput., 16(6): 3077â3095. Chuang, Y.-N.; Wang, G.; Chang, C.-Y.; Tang, R.; Zhong, S.; Yang, F.; Wen, A.; Du, M.; Cai, X.; Braverman, V.; and Hu, X. 2026. FaithLM: Towards Faithful Explanations for Large Language Models. In Proceedings of EACL. El-Yaniv, R.; and Wiener, Y. 2010. On the Foundations of Noise-free Selective Classification. J. Mach. Learn. Res., 11: 1605â1641. Gao, Y.; Meng, Q.; Zhou, Y.; and Pan, L. 2026. Towards In- trinsic Interpretability of Large Language Models: A Survey of Design Principles and Architectures. arXiv:2604.16042. Goddard, C.; and Wierzbicka, A. 2014. Words and Meanings: Lexical Semantics Across Domains, Languages, and Cultures. Oxford University Press. Huang, Y.; Chen, S.; Cai, H.; and Dhingra, B. 2025. To Trust or Not to Trust? Enhancing Large Language Modelsâ Situated Faithfulness to External Contexts. In Proceedings of ICLR. Jacovi, A.; and Goldberg, Y. 2020. Towards Faithfully Inter- pretable NLP Systems: How Should We Define and Evaluate Faithfulness? In Proceedings of ACL. Koh, P. W.; Nguyen, T.; Tang, Y. S.; Mussmann, S.; Pierson, E.; Kim, B.; and Liang, P. 2020. Concept Bottleneck Models. In Proceedings of ICML. Lanham, T.; et al. 2023. Measuring Faithfulness in Chain-of- Thought Reasoning. arXiv:2307.13702. Liu, G. K.-M.; Yona, G.; Caciularu, A.; Szpektor, I.; Rudner, T. G. J.; and Cohan, A. 2025. MetaFaith: Faithful Natural Language Uncertainty Expression in LLMs. In Proceedings of EMNLP. ArXiv:2505.24858. Liu, Y.; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy, O.; Lewis, M.; Zettlemoyer, L.; and Stoyanov, V. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Ap- proach. arXiv:1907.11692. Lyu, Q.; Apidianaki, M.; and Callison-Burch, C. 2024. To- wards Faithful Model Explanation in NLP: A Survey. Com- putational Linguistics, 50(2). Ma, Y.; Nguyen, K. L.; Xing, F. Z.; and Cambria, E. 2020. A Survey on Empathetic Dialogue Systems. Information Fusion, 64: 50â70. Madsen, A.; Chandar, S.; and Reddy, S. 2024. Are Self- Explanations from Large Language Models Faithful? In Findings of ACL. Matton, K.; Ness, R. O.; Guttag, J.; and Kıcıman, E. 2025. Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations. In Proceedings of ICLR. ArXiv:2504.14150. Miao, M. M.; and Ungar, L. 2026.Closing the Confidence-Faithfulness Gap in Large Language Models. arXiv:2603.25052. Ming, Y.; Purushwalkam, S.; Pandit, S.; Ke, Z.; Nguyen, X.-P.; Xiong, C.; and Joty, S. 2025. FaithEval: Can Your Language Model Stay Faithful to Context, Even If "The Moon is Made of Marshmallows". In Proceedings of ICLR. Oldfield, J.; Im, S.; Li, Y.; Nicolaou, M. A.; Patras, I.; and Chrysos, G. G. 2025. Towards Interpretability Without Sacri- fice: Faithful Dense Layer Decomposition with Mixture of Decoders. In Advances in Neural Information Processing Systems (NeurIPS). Ortony, A.; Clore, G. L.; and Collins, A. 1988. The Cognitive Structure of Emotions. Cambridge University Press. Pavlick, E.; and Kwiatkowski, T. 2019. Inherent Disagree- ments in Human Textual Inferences. Transactions of the Association for Computational Linguistics (TACL). Plank, B. 2022. The "Problem" of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation. In Proceedings of EMNLP. Rajamanickam, S.; Mishra, P.; Yannakoudakis, H.; and Shutova, E. 2020. Joint Modelling of Emotion and Abusive Language Detection. In Proceedings of ACL, 4270â4279. Röttger, P.; Vidgen, B.; Hovy, D.; and Pierrehumbert, J. 2022. Two Contrasting Data Annotation Paradigms for Subjective NLP Tasks. In Proceedings of NAACL-HLT. Scherer, K. R. 2009. The Dynamic Architecture of Emotion: Evidence for the Component Process Model. Cognition and Emotion, 23(7). Si, S.; Zhao, H.; Gao, C.; Bai, Y.; Wang, Z.; Gao, B.; Luo, K.; Li, W.; Huang, Y.; Chen, G.; Qi, F.; Zhang, M.; Chang, B.; and Sun, M. 2026. Teaching Large Language Models to Maintain Contextual Faithfulness via Synthetic Tasks and Reinforcement Learning. In Proceedings of AAAI. Siegel, N. Y.; Heess, N.; Perez-Ortiz, M.; and Camburu, O.- M. 2025. Verbosity Tradeoffs and the Impact of Scale on the Faithfulness of LLM Self-Explanations. arXiv:2503.13445. Smith, C. A.; and Ellsworth, P. C. 1985. Patterns of Cognitive Appraisal in Emotion. Journal of Personality and Social Psychology, 48(4). Somov, O.; Chaichuk, M.; Ershov, G.; Vafin, K.; Seleznyov, M.; Panchenko, A.; and Tutubalina, E. 2026. Breaking the Chain: A Causal Analysis of LLM Faithfulness to Intermedi- ate Structures. arXiv:2603.16475. Sui, Y.; He, Y.; Liu, N.; He, X.; Wang, K.; and Hooi, B. 2025. FiDeLiS: Faithful Reasoning in Large Language Models for Knowledge Graph Question Answering. In Findings of ACL. Sun, C.-E.; Oikarinen, T.; Ustun, B.; and Weng, T.-W. 2025. Concept Bottleneck Large Language Models. In Proceedings of ICLR. Tak, A. N.; Banayeeanzade, A.; Bolourani, A.; Kian, M.; Jia, R.; and Gratch, J. 2025. Mechanistic Interpretability of Emotion Inference in Large Language Models. In Findings of ACL. ArXiv:2502.05489. Tang, X.; Li, J.; Hu, K.; Du, N.; Li, X.; Zhang, X.; Sun, W.; and Xie, S. 2025.CogniBench: A Legal-Inspired Framework and Dataset for Assessing Cognitive Faithful- ness of Large Language Models. In Proceedings of ACL. ArXiv:2505.20767. Tangney, J. P.; and Dearing, R. L. 2002. Shame and Guilt. Guilford Press. Tracy, J. L.; and Robins, R. W. 2004. Putting the Self Into Self-Conscious Emotions: A Theoretical Model. Psychologi- cal Inquiry, 15(2). Troiano, E.; OberlĂ€nder, L.; and Klinger, R. 2023. Dimen- sional Modeling of Emotions in Text with Appraisal Theo- ries: Corpus Creation, Annotation Reliability, and Prediction. Computational Linguistics, 49(1). Uma, A. N.; Fornaciari, T.; Hovy, D.; Paun, S.; Plank, B.; and Poesio, M. 2021. Learning from Disagreement: A Survey. Journal of Artificial Intelligence Research (JAIR). Wang, Y.; and Atanasova, P. 2025. Self-Critique and Re- finement for Faithful Natural Language Explanations. In Proceedings of EMNLP. Wierzbicka, A. 1996. Semantics: Primes and Universals. Oxford University Press. Wierzbicka, A. 1999. Emotions Across Languages and Cul- tures: Diversity and Universals. Cambridge University Press. Xing, F.; Malandri, L.; Zhang, Y.; and Cambria, E. 2020. Financial Sentiment Analysis: An Investigation into Common Mistakes and Silver Bullets. In Proceedings of COLING. Yeo, W. J.; Satapathy, R.; and Cambria, E. 2025. Towards Faithful Natural Language Explanations: A Study Using Ac- tivation Patching in Large Language Models. In Proceedings of EMNLP. Zhao, Y.; Deng, Y.; Ng, S.-K.; and Chua, T.-S. 2025. Align- ing Large Language Models for Faithful Integrity Against Opposing Argument. In Proceedings of AAAI. A Schema Reference and Full Rule Set This appendix collects the rest of the interface specification: the NSM source classifications the schema discretizes and the complete thirteen-rule mapper. The NSM literature clas- sifies emotion concepts by elicitor category (Table 8) and by cognitive template (Table 9), and gives short prime-only ex- plications for each concept (Figure 1 in the main text is one). Our twelve-slot EES discretizes these prime configurations rather than replacing the prime inventory; Table 10 then lists the complete mapper that reads the resulting slots. Table 10 lists all thirteen rules of the mappergas a priority- ordered decision list (index0highest). The first rule whose conditions hold fires; no match yields ABSTAIN. Each condi- tion is transcribed from a published NSM explication; none is learned. B Per-Slot Annotation Reliability Table 11 gives the full per-slot raw agreement and Krippen- dorffâs α for the silver pass and the first human pass. C Mapper Behaviour: Typed Abstention, Priority Sensitivity, Distributional Evaluation, and Confidence Intervals This appendix reports four diagnostics of the mappergthat the main text summarizes: the composition of abstention by type (Table 12), the sensitivity of accuracy to rule ordering (Table 13), a distributional scoring ofgas a 13-way label dis- tribution (Table 14), and exact binomial confidence intervals on the held-out set (Table 15). D Additional Result Figures This appendix collects the two full-resolution result figures summarized in Section 6: the per-item verification histogram (Figure 7) and the complete routing confusion matrix (Fig- ure 8). Table 8: Emotion concepts analyzed with the NSM approach by Wierzbicka (1999). CategoryEmotion concepts Bad things happeningsad, unhappy, distressed, upset, sorrow, sorry, grief, despair, depressed Good things happeningjoy, contented, pleased, delighted, excited People doing bad thinganger, indignation, shocked, appalled, hurt Thinking about ourselvesremorse, guilt, shame, humiliation, embarrassment, pride, triumph Unclassified frustration, relief, disappointment, surprise, amazement, happy (glĂŒcklich, heureux, schastlivyi), frightened Table 9: Emotion concepts associated to cognitive templates as analyzed by Wierzbicka (1999). TemplateEmotion concepts First-person thought-plus-feeling(+) great, wonderful, terrific, awesome, fabulous (â) awful, dreadful, terrible Experiential evaluation(+) entertaining, delightful, fascinating, compelling, interesting, touching (â) boring, predictable Experiential with bodily reaction(+) gripping, exciting, stunning, suspenseful, tense (â) disgusting, sickening Lasting effect(+) powerful, memorable, haunting, inspiring (â) depressing, disturbing Cognitive evaluation (+) complex, excellent, outstanding, impressive, brilliant, clever, original (â) disappoint- ing, dismal, woeful # EmotionSlot conditions (conjunction) 0 reliefWANT=NOT-WANT, REALIZATION=NOT-HAPPENED, EVALUATION=FEEL-GOOD 1 surpriseKNOWLEDGE=NOT-KNOW, REALIZATION=HAPPENED 2 fearEVALUATION=FEEL-BAD, TIME-DIR=AFTER-NOW, REALIZATION=MAYBE 3 boredomTRIGGER=NOTHING, WANTâWANT,NOT-WANT, EVALUATION=FEEL-BAD 4 disgustEVALUATION=FEEL-BAD, BODY=YES, WANT=NOT-WANT, EVAL-TARGETâOBJECT,OTHER 5 shameEVALUATION=FEEL-BAD, EVAL-TARGET=SELF, OTHERS-KNOW=CAN-KNOW 6 guiltEVALUATION=FEEL-BAD, EVAL-TARGET=SELF, AGENCY=I 7 prideEVALUATION=FEEL-GOOD, EVAL-TARGET=SELF, AGENCY=I 8 trustEVALUATION=FEEL-GOOD, EVAL-TARGET=OTHER, KNOWLEDGE=KNOW 9 angerEVALUATION=FEEL-BAD, AGENCY=SOMEONE-ELSE, EVAL-TARGET=OTHER 10 sadnessEVALUATION=FEEL-BAD, REALIZATION=NOT-CAN 11 joyEVALUATION=FEEL-GOOD, TRIGGERâSOMETHING-HAPPENED,SOMEONE-DID 12 no-emotion EVALUATION=NEITHER Table 10: The complete mapperg. Shame precedes guilt (both share FEEL-BADâ§SELF, split by OTHERS-KNOW vs. AGENCY); the general joy rule is last so specific positive emotions shadow it. SlotAgree. (S) α (S) α (H) EXPERIENCER0.9940.00 â0.01 EVALUATION0.8650.7680.227 AGENCY0.8270.7370.607 WANT0.8210.6960.395 EVAL-TARGET0.6730.5480.702 REALIZATION0.7440.5420.788 TRIGGER0.7050.4950.656 TIME-DIRECTION0.6860.4490.192 OTHERS-KNOW0.8850.4370.316 INTENSITY0.7880.4250.230 BODY0.8850.4080.402 KNOWLEDGE0.7690.3910.081 Mean0.8030.4910.382 Table 11: Per-slot raw agreement and Krippendorffâsα. S = silver model pass (n=156); H = first human pass (n=99). High raw agreement co-occurs with lowαon skewed slots (EXPERIENCER, OTHERS-KNOW, BODY), so the chance- corrected coefficient is the one to read. Abstention typeFull batch36-eval No-rule (legal, no rule fires)141 (11.8%)12 Schema-residue (residue present)25 (2.1%)0 Illegal (not schema-legal)00 Total abstain166 (13.8%)12 (33.3%) Multi-rule (resolved by priority)342 (28.5%)â Table 12: Typed abstention (A). Abstention is dominated by no-rule-fires, not malformed explications, and multi-rule items never abstain because priority resolves every tie. Mapper configurationFull-batch acc NSM most-specific-first (ours)0.382 Swap: guiltâshame0.388 Random orders (N=200): min/mean/max0.295 / 0.343 / 0.392 Single-winner selective acc0.443 Set-valued (goldâ fired set)0.521 Table 13: Priority-order sensitivity (A). Our hand order sits at the96th percentile of random orders; the guilt/shame swap moves accuracy under a point. Set-valued accuracy is0.521 vs.0.443single-winner: about eight points live in priority resolution. Scoring ruleNLLBriertop-1top-3cov. single-winner2.206 0.833 0.398 0.498 0.862 uniform-fired2.140 0.826 0.328 0.558 0.862 priority-decay 2.126 0.817 0.398 0.558 0.862 Table 14: Distributional evaluation ofgon tier A (n=1,200) with Laplace smoothing (α=0.05). Priority-decay strictly dominates single-winner: top-3 recovers six points, NLL and Brier both fall. Set-valued accuracy (goldâfired set) is 0.449. System (n=36)Acc95% CI Fine-tuned 8B (ours)0.333 [0.19, 0.51] Black-box 8B0.389 [0.23, 0.57] Appraisal (oracle)0.361 [0.21, 0.54] Gold explication (routed)0.306 [0.16, 0.48] Table 15: Exact (ClopperâPearson)95%binomial intervals on the held-out set. All four overlap, so no pairwise accuracy difference is resolvable at n=36; we read Table 4 as a trade- off, not a ranking. 0.00.20.40.60.81.0 verifier score v (per item) 0 100 200 300 400 500 count ( n =1200 ) mean 0.20 Figure 7: Per-item entailment verification score on the full batch (A,n=1,200). The verifier is conservative on prime- vocabulary propositions:40.6%of items score zero and the mean is0.20. On the silver subset the uncalibrated score separates high-divergence items at AUC0.477and silver- correct items at AUC 0.576, significant on neither (S). anger boredom disgust fear guilt joy no-emotion priderelief sadness shame surprise trust ABSTAIN routed label anger boredom disgust fear guilt joy no-emotion pride relief sadness shame surprise trust prompted (gold) emotion 46135156519 53911117161119 2320111152219 141911414351634 2311294543428 6441221531 452654213214 1444331145 113065401341 7124916233829 412541143210 1134351408 121461564484 0.0 0.2 0.4 0.6 0.8 1.0 row fraction Figure 8: Row-normalized routing confusion matrix on the full batch (A,n=1,200); raw counts shown, the diagonal is boxed and the right column is abstention.