Paper deep dive
Auditing the Synthetic Memoir: Measuring Scene-Level Confabulation in LLM-Generated Autobiography Against the Documented Record of the Life It Describes
Heather Renze
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/26/2026, 4:02:19 AM
Summary
This paper presents a scene-level audit of LLM-generated autobiography, measuring confabulation against a subject-specific ground-truth corpus. The study audited 366 daily entries drafted by an LLM (gpt-4o/o3/o4-mini-high) using only a template and quotes, finding a 96.7% verification-failure rate. The dominant failure mode was 'grounded drift' (real people/settings in invented scenes). Regenerating content with grounding in the subject's corpus improved verification but left an 83.3% failure rate. The study highlights the unreliability of LLM-generated labels for fact-checking and the need for robust grounding.
Entities (10)
Relation Signals (7)
Heather Renze Serenze → authored → After the Unicorn
confidence 95% · the memoir manuscript After the Unicorn
Heather Renze Serenze → authored → Birth of a Unicorn
confidence 95% · the memoir Birth of a Unicorn
LLM-generated autobiography → exhibits → grounded drift
confidence 95% · the dominant failure mode is grounded drift
LLM-generated autobiography → has → verification-failure rate
confidence 95% · 354 of 366 days fail, 96.7%
o3-pro → designed → Template
confidence 90% · OpenAI o3-pro designed the template
GPT-4o → generated → LLM-generated autobiography
confidence 90% · gpt-4o / o3 / o4-mini-high generated entries
HEATHER_CORPUS → usedfor → ground-truth corpus
confidence 90% · a 353k-word corpus of the subject’s published writing and talk transcripts (HEATHER_CORPUS)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:When a large language model (LLM) is asked to write a person's life, how much of what it writes actually happened? We present a scene-level case-study audit - the first quantified audit of LLM-generated autobiography against a subject-specific ground-truth corpus that we are aware of, based on an unsystematic literature search. The subject and the author of this paper are the same person: a 366-day "page-a-day" book of first-person anecdotal entries was drafted with a conversational LLM whose documented inputs were a template, two exemplar days, and each day's quote - not her corpus - and every day was subsequently audited at the anecdote-scene level against an independent verification corpus using a four-level rubric fixed before analysis. We define the verification-failure rate as the share of days not rated VERIFIED (scene positively corroborated): 354 of 366 days fail, 96.7% (Wilson 95% CI 94.4-98.1%). Only 12 days contain a corroborated scene; 19 days (5.2%) assert claims actively contradicted by the record; the dominant failure mode is grounded drift - real people, employers, and settings inside invented scenes - though its measured share varies across raters. Independent re-rating replicates the headline (no evidence the original rate was inflated) while showing that the four-way taxonomy has only fair-to-moderate reliability. Regenerating the same days with current named models reproduces 100% verification failure under the same inputs; grounding generation in the subject's corpus significantly improves the verification rate while leaving substantial residual failure (83.3%). We contribute the measurement, a reusable audit instrument whose WEAK/UNVERIFIED boundary we show to be unreliable, and a grounding remedy with quantified effect.
Tags
Links
- Source: https://arxiv.org/abs/2608.23640v1
- Canonical: https://arxiv.org/abs/2608.23640v1
Trouble viewing inline? Open PDF directly →
Full Text
60,051 characters extracted from source content.
Expand or collapse full text
Auditing the Synthetic Memoir: Measuring Scene-Level Confabulation in LLM-Generated Autobiography Against the Documented Record of the Life It Describes Heather Renze Serenze Global serene@serenze.org ORCID: https://orcid.org/0009-0001-8789-4611 August 22, 2026 Abstract When a large language model (LLM) is asked to write a person’s life, how much of what it writes actually happened? We present a scene-level case-study audit — the first quantified audit of LLM-generated autobiography against a subject-specific ground-truth corpus that we are aware of, based on an unsystematic literature search. The subject and the author of this paper are the same person: a 366-day “page-a-day” book of first-person anecdotal entries was drafted with a conversational LLM whose documented inputs were a template, two exemplar days, and each day’s quote — not her corpus — and every day was subsequently audited at the anecdote-scene level against an independent verification corpus using a four-level rubric fixed before analysis. We define the verification-failure rate as the share of days not rated verified (scene positively corroborated): 354 of 366 days fail, 96.7% (Wilson 95% CI 94.4–98.1%). Only 12 days contain a corroborated scene; 19 days (5.2%) assert claims actively contradicted by the record; the dominant failure mode is grounded drift — real people, employers, and set- tings inside invented scenes — though its measured share varies across raters. Independent re-rating replicates the headline (no evidence the original rate was inflated) while showing that the four-way taxonomy has only fair-to-moderate reliability. Regenerating the same days with current named models reproduces 100% verification failure under the same inputs; grounding generation in the subject’s corpus significantly improves the verification rate while leaving sub- stantial residual failure (83.3%). We contribute the measurement, a reusable audit instrument whose WEAK/UNVERIFIED boundary we show to be unreliable, and a grounding remedy with quantified effect. 1 Introduction Systems that simulate people from their data are moving from research prototypes [1–3] into personal use: digital twins, synthetic biographies, and ghost-written life narratives built by prompting an LLM with a person’s corpus and asking it to write “her memories” in the first person. Evaluation of such systems has concentrated on attitudes, survey answers, and surface facts [2, 3]. The episodic layer — the specific scenes a system asserts about a life — has not been audited against the documented record of that life, although laboratory work shows conversational AI can implant false memories in human subjects [4]. This paper reports a complete, naturalistic instance of exactly this failure mode, caught and measured before publication. Between drafting sessions, a professional author (the present paper’s 1 arXiv:2608.23640v1 [cs.AI] 23 Aug 2026 author) produced a year-long gift book: 366 daily entries (≈543 words each on average; 198,949 words in total), written in the first person about her own life. Repository documentation states plainly that the entries were “AI-padded”: drafted with a conversational LLM (generator docu- mented, Section 7; its documented inputs were the template and the day’s quote — see below). Before the derived page-a-day edition was assembled, every one of the 366 days was fact-checked at the scene level against an independent ground-truth corpus assembled for that purpose, using a rubric fixed on July 13, 2026 — before any analysis in this paper was performed. We ask three questions. RQ1: What fraction of generated anecdotal “memories” survive scene- level verification against an independent record? RQ2: What forms does the failure take? RQ3: Where do contradictions concentrate, and what remediation does that suggest? The study is falsifiable in the ordinary way. Had half or more of the 366 scenes been rated verified, the framing “generation without grounding” would fail; had verification failures been dominated by unverified rather than weak days, the “grounded drift” account in Section 6 would fail. Both outcomes were live possibilities before tallying. We contribute: (a) a quantified audit of confabulation in personal-narrative generation against subject-specific ground truth (n = 366 day-narratives; 96.7% fail positive scene-level verification under our definition); (b) codification and release of a reusable scene-level audit instrument — the four-verdict rubric quoted verbatim plus a nine-theme recurring-false-premise taxonomy opera- tionalized as fixed, published keyword screens; and (c) a remediation workflow (lesson-only rewrite; author-adjudication precedence) specified during the same project; and (d) an inter-rater reliabil- ity analysis with two independent re-raters, yielding the paper’s most transferable methodological finding: scene-level fabrication audits are reliable at the binary level (does this day’s scene check out?) but only fair at the taxonomic level, because the rubric’s WEAK/UNVERIFIED boundary turns on an undefined notion of “setting.” A note on positionality shapes everything that follows: the auditing author and the audited subject are the same person. Section 8 treats the consent, privacy, and self-audit implications; Section 4.6 treats the methodological ones. 2 Related Work Hallucination and confabulation in LLMs. Hallucination is surveyed extensively: factuality vs. faithfulness taxonomies [5, 6], mitigation surveys [7, 8], and a 2026 comprehensive review [9]. Sui et al. [10] argue for the term confabulation — fluent, confident fabrication that fills gaps — which we adopt: the audited entries are not random noise but coherent, confident first-person nar- rative. ReFACT [11] benchmarks scientific confabulation detection with fine-grained annotations; our instrument is analogous but subject-grounded rather than encyclopedic. Factuality measurement of long-form generation. FActScore decomposes long-form text into atomic claims and checks each against a knowledge source [12]; long-form factuality benchmarks evaluate models against world facts [13]. Our audit is a scene-level analogue: a day-entry is decom- posed into its anecdote scene, and the reference work is not an encyclopedia but an independent corpus documenting one life. No prior benchmark we are aware of uses a verified human-life corpus as ground truth. Person simulation and digital twins. Generative agents produce believable behavior from profiles and memories [1]. Interview-grounded agents simulate 1,000 real individuals [2], with fidelity evaluated on attitudes and survey responses; persona work explicitly separates surface-fact from deep persona fidelity [3]; digital-twin construction from conversation logs is demonstrated for counselors [14]. Our results supply the missing episodic dimension: even surface episodic content — employers, 2 dates, named events — fails verification in the large majority of generated days. False memories and identity risks. Chan et al. [4] show LLM conversational partners amplify false memories in witness interviews — the consumption side of the risk. Our study documents the complementary production side at scale: an LLM asked to write a person’s past produces unsupported scenes in 96.7% of days under our definition. Persona framing also perturbs factual QA [15], consistent with our observation that first-person narrative framing invites invention. Character consistency failures. RoleBreak and SHARP treat character hallucination in role- playing systems as an attack surface [16, 17]. We quantify the same drift class when the “character” is a real, documented person and the drift is measured against her actual records rather than a prompt specification. Gap. Hallucination is surveyed; long-form factuality is benchmarked against encyclopedic truth; person simulators are built and evaluated on attitudes; false-memory risks are shown in laboratories; character drift is characterized adversarially. No published study audits LLM-generated autobiography scene-by-scene against the documented record of the person it describes. This paper fills that gap. 3 Data and Study Design The study stands on three corpora, all fixed before this paper’s analysis session (August 21, 2026); the audit labels themselves predate it by five weeks. Suspect corpus (what was audited). A year-long “page-a-day” book of 366 daily first-person entries (≈543 words/day; 198,949 words in total), drafted for a gift edition. The source calendar’s own README documents the entries as “AI-padded”: long-form drafts written with a conversa- tional LLM. Session-log provenance recovery (Section 7) has since documented the generator of record: OpenAI o3-pro designed the template and gpt-4o / o3 / o4-mini-high generated entries from the quote list in ChatGPT conversations (July–August 2025), with a Cursor composer run on August 1, 2025 executing a one-shot prompt. A derived edition compresses each entry to a ≈175-word page (64,198 words total). The audit month-files are headed “page-a-day,” but their anecdote rows correspond scene-for-scene to the long-form entries, not the compressed pages: the May 1 audit row, for example, describes an Evernote scene (three stakeholders’ deliverables, cry- ing at 2 a.m.) that appears only in the long source, and the October 10 row’s “empty Pacific horizon” phrasing likewise matches only the long text. We therefore treat the long-form entries as the text that was audited. What the generator received is documented precisely (Section 7; results/original_generation_prompt.txt): a template, two exemplar days carrying personal facts, and the day’s quote from commonsense.csv — not the ground-truth corpus. Only replication Arm C receives corpus excerpts; the 2025 original and Arms B/B2 receive the template-plus-quote inputs alone; any additional biographical material the 2025 run carried in conversation context or drew from pretraining is undocumented. This matters for interpretation: the “grounded drift” pattern of Section 5 arose without the record in front of the model. The audited unit is the day- entry, each organized around one anecdotal “memory” scene plus a lesson. The book project’s build pipeline (a derived ≈150–200-word page per day, plus a 367-row quote index) derives mechanically from these entries; the day-entries are the atomic object of audit. Ground-truth corpus (what was checked against). Assembled by the audit to be indepen- dent of generation: (i) the memoir manuscript After the Unicorn with its editorial anecdote ledgers; (i) the memoir Birth of a Unicorn; (i) a 353k-word corpus of the subject’s published writing and talk transcripts (HEATHER_CORPUS); (iv) a speaking record and knowledge base maintained for her professional digital twin; (v) a music catalogue with dated life-era annotations; (vi) targeted 3 web checks. The calendar source itself, its page-a-day derivative, and the ChatGPT conversation backups were excluded as non-independent — they could not corroborate what they generated. Audit labels (the measurements). Every day received a scene-level verdict on July 13, 2026, recorded in twelve month tables (one row per day: date, anecdote gist, verdict, supporting source, evidence note), summarized in a fact-check summary document. Critical items were adjudicated by the subject directly; four explicit author confirmations are recorded, and the summary states that author confirmations take precedence over any ruling. This label set — not any new judgment — is the dataset analyzed here; Sections 4.3–5 are strictly mechanical parsing and aggregation of documents that existed before this paper was begun. Timeline. Entries generated July–August 2025 (documented; per-day dates unrecorded) → scene-level audit completed July 13, 2026 → remediation designed (same date) → dataset parsed and analyzed August 21, 2026. No audit label was created or modified by this paper. 4 Method 4.1 Verdict taxonomy The rubric below is quoted verbatim from the audit summary dated July 13, 2026 (FACT_CHECK_ SUMMARY.md); it is a pre-stated instrument, not a post hoc scheme. Each day’s central anecdote scene received exactly one verdict: • verified — “specific scene corroborated”; • weak — “real setting/person, invented specific scene”; • unverified — “no anchor; generic invention”; • contradicted / flagged — “asserts something FALSE about her life.” The classes are ordered by evidential relation to the record, not by severity alone: verified requires positive corroboration; unverified records absence of evidence; contradicted requires positive disproof. weak is the pivotal class for person simulation: the system correctly retrieved real grounding material (employer, city, person) and invented the specifics around it. 4.2 Audit procedure For each of the 366 days, the auditor extracted the day’s anecdote scene from the entry, searched the ground-truth corpus for corroboration or contradiction, recorded the verdict with a supporting source citation and an evidence note in the month table, and flagged candidate false premises. Where a claim turned on facts only the subject could settle (e.g., whether a childhood accident occurred at age 13 or in adulthood; whether a treatment history included chemotherapy), the subject issued an explicit written ruling, and the summary’s precedence rule — “Author confirmations (2026-07-13, take precedence)” — binds all other rows. Four such rulings are recorded. Health-related content in this paper is restricted to what those rulings themselves state. 4.3 Parsing and analysis procedures Two scripts, both published with the paper, convert the audit tables into every number in Section 5. 4 01_parse_factcheck.py reads the twelve month tables and emits one CSV row per day. Verdict cells are normalized by fixed rules in fixed order: any cell containing “contradicted” → contra- dicted; else prefix “verified” → verified; else prefix “unverified” → unverified; else prefix “weak” → weak; anything else is reported as unparsed, never silently dropped. Rows are keyed on parsed month/day; the parser reports duplicates and unparsed candidate lines explicitly. Output: 366 rows, zero duplicate keys, zero unparsed verdicts (parse_report.txt). 02_analyze.py asserts n = 366, tallies verdicts overall and per month, computes Wilson score intervals (z = 1.96), applies the nine fixed keyword screens of Section 4.5 over each row’s gist+note text, codes each populated Source cell against seven named source classes plus two fallbacks (multi- label; a cell may match several), and writes stats.json, four figures, and two tables. Every number, table, and figure in this paper is reproduced by running these two scripts in sequence; stats.json is published so values can be diffed without re-running. 4.4 Definition of the verification-failure rate Because this rate is the paper’s headline quantity, we state its definition explicitly and report a denominator check. Throughout this paper, verification-failed day≡ any day whose verdict is not verified, i.e., weak∪ unverified∪ contradicted = 227 + 108 + 19 = 354 of 366 days = 96.7% (Wilson 95% CI [94.4%, 98.1%]). We name this quantity a verification failure rather than “fabrication”: an unverified scene may be real-but-unrecorded, so the defensible claim is that the day failed positive scene-level verification, not that its content is false. The definition remains deliberately conservative in the direction of undercounting failure: machine-generated scene content must survive positive corroboration to be counted as verified. The term fabricated is reserved in this paper for the contradicted class only (5.2%), whose content the record positively disproves. The source audit summary states its own approximate tally, “≈350 of 366 days carry a fabricated anecdote,” computed from pre-parse counts (≈9 verified days) before exact table parsing. Exact parsing yields 12 verified days. The two figures agree within rounding of the approximate tally (354/366 vs.≈ 350/366; both round to≥95%). All inferential statistics in this paper use the parsed denominators; we flag any reader-facing number that depends on which convention is used. 4.5 Recurring false-premise screens The audit identified recurring false premises — specific false claims about the life that resurface across many days. To make their footprint reproducible rather than anecdotal, each premise was operationalized as a fixed regular-expression screen over the audit rows’ gist and note text (patterns fixed verbatim in 02_analyze.py). These are co-occurrence screens: a day counts if its audit row references the premise theme, not because a human re-confirmed causality per day. Counts are therefore ceilings for per-premise incidence and may overlap across premises (three dates appear under two premises). 4.6 Limitations of single-rater verdicts We describe the rating instrument exactly as it is. The 366 verdicts were assigned by an LLM auditor working against the named independent sources: every month-table row carries an evidence note citing specific source locations, and the rubric and precedence rule were fixed before tallying. The author (subject) adjudicated a small number of critical items only — four explicit written 5 confirmations are recorded in the audit summary — and those rulings take precedence over any other row. No second rater participated in the original audit, so no inter-rater reliability statistic accompanied the labels, and rater drift across the twelve month tables cannot be measured post hoc. The consequence must be stated plainly: the dependent variable of this study is an LLM judg- ment, which means this paper measures LLM confabulation using LLM-produced labels — a circu- larity we do not pretend away. Section 5.6 reports a blind independent re-rate of a 60-day sample by a second rater from a different model family, designed to bound exactly this concern. 5 Results 5.1 The headline distribution Table 1 and Figure 1 give the verdict distribution over all 366 days. Under the definition of Sec- tion 4.4, 354 of 366 days (96.7%, Wilson 95% CI 94.4–98.1%) failed positive scene-level verification. Only 12 days (3.3%) contain a scene that survived positive corroboration. At the other extreme, 19 days (5.2%, CI 3.3–8.0%) assert claims the record actively contradicts — the machine did not merely imagine, it misstated documented fact about a documented person. In the original audit, the bulk of verification failures is weak — real settings, employers, and people arranged inside invented scenes — and weak is the single largest category in all three independent ratings of the sampled data, though its measured share is rater-sensitive (Section 5.6: 43–82% across raters). Table 1: Scene-level verdicts over 366 generated day-narratives. “Failed verification” is any non- verified day (Section 4.4); the source summary’s approximate tally (≈350/366) is consistent with this exact count. VerdictDays % of 366 VERIFIED123.3% WEAK22762.0% UNVERIFIED10829.5% CONTRADICTED / flagged195.2% Failed verification (any non-VERIFIED class) 35496.7% 5.2 Monthly variation Figure 2 shows the cross-tabulation by month. Three patterns stand out. First, verification is rare everywhere and clustered: six months (January–March, June, July, August) contain zero verified days; May contributes 4 of the 12. Second, contradicted days are unevenly clustered rather than uniform: August (6), April (4), and December (4) account for 14 of 19. The repository records no generation order or dates for the month files, so we draw no chronological conclusion from this clustering. Third, the unverified share swells mid-year (July: 18 of 31 days), marking stretches where entries detached from the corpus entirely and invented generically. 6 VERIFIEDWEAKUNVERIFIEDCONTRADICTED 0 50 100 150 200 250 Days (n = 366) 12 (3.3%) 227 (62.0%) 108 (29.5%) 19 (5.2%) Figure 1: Verdict distribution, n = 366. 5.3 Recurring false premises (RQ2) Table 2 and Figure 3 show the nine recurring false premises and their screen footprints (31 dis- tinct days; three dates screen under two premises). The taxonomy spans four fabrication families: invented relations (nonexistent children, named strangers); invented capabilities and events (blue- water sailing, an adult car crash, a conference keynote on Antarctica); misattributed facts (treatment history, first keynote, education, Evernote role); and inverted dispositions (fearing a diagnosis that in fact went unnoticed). One screen (P8, education) was overturned by same-day author adjudica- tion — the education content it flags was author-confirmed as real — and is retained in the table to document the instrument’s correction trail, not as a confirmed error count. PremiseFamilyScreen hits P1 Invented children / “the kids”invented relations5 P2 Blue-water sailor personainvented capability4 P3 Treatment-history misattribution misattributed fact4 P4 Invented adult car accidentinvented event5 P5 Evernote role inflationmisattributed fact5 P6 AntarctiConf “on the ice” fiction invented event3 P7 First-keynote misattributionmisattributed fact3 P8 Education misattributionmisattributed fact (reversed on adjudication)3 P9 Invented named third partiesinvented relations2 Table 2: Recurring false premises; fixed keyword screens over audit rows (Section 4.5). Per-day lists: Appendix C. 5.4 What the audit checked against Of 366 audit rows, 114 (31.1%) cite no independent source — the day’s scene had no anchor in the record at all. The remaining 252 rows carry 333 source-class mentions (multi-label): memoir After the Unicorn 134 (53% of sourced rows), anecdote ledger 94 (37%), memoir Birth of a Unicorn 7 Jan Feb MarAprMayJun Jul AugSepOctNovDec 0 5 10 15 20 25 30 Days VERIFIEDWEAKUNVERIFIEDCONTRADICTED Figure 2: Verdicts by month, n = 366. 68 (27%), asserted-real settings without a named source 16, and music catalogue, web, knowledge- base, and talk-transcript citations in single digits each (Figure 4). The memoirs and the ledger — curated, author-controlled documents — carry over 80% of the verification load, which is itself a finding: personal archives are the only ground truth that exists for a life. 5.5 What survives, and what fails Table 3 lists all 12 verified days; the survivors are almost entirely public-career facts (Evernote tenure and scale metrics) or scenes the subject has told publicly for years. Appendix A lists all 19 contradicted days with their gists. The asymmetry is instructive: what survives is what the corpus says often; what fails is what sounds like the life — poignant, quotable, structurally perfect for a page-a-day format. 5.6 Inter-rater reliability: two independent re-raters Because the original labels are themselves LLM-produced (Section 4.6), two independent raters — both from a different model family than the original auditor — re-rated random samples of 60 of the 366 days (uniform day-of-year draw, fixed seed 20260821, published before any rating), applying the same verbatim rubric to the same long-form entries against the same ground-truth corpus. Their exposure conditions differed, and we disclose both. Rater B performed a source-provenance task in the same session before rating and saw the original audit rows for two days (neither in its sample), the audit summary’s four author confirmations, and aggregate counts; all 60 of its ratings were recorded and frozen before any comparison. By contrast, the replication raters D and E of Section 5.7 were fully blind. Rater C was fully blind: a fresh session, forbidden from reading the audit files, the derived dataset, and this paper. Binary result (the headline’s foundation). On whether a day’s scene survives verification — VERIFIED vs. not — raw agreement is 95.0% (rater B, 57/60) and 98.3% (rater C, 59/60). We deliberately report no binary Cohen’s κ: it is degenerate here. At a VERIFIED base rate of ≈1–5%, and with rater C assigning VERIFIED to zero of its 60 days, one response column has no variance and κ collapses toward zero regardless of agreement; quoting it would understate near- 8 Table 3: All 12 days whose central scene was positively corroborated. DateVerdictAnecdote gist (abbreviated) April 2VERIFIED Joined Evernote as employee number eight; a mentor said "build a culture" April 6VERIFIED Hit one hundred million users at Evernote; no celebration May 9VERIFIED Lived on a sailboat in Mexico & an RV at Disney World; travel light May 13VERIFIED Asked every founder "what aren’t you saying?"; never turned it on her own dying marriage. . . May 14VERIFIED Underestimated as a woman at THQ; winning College Jeopardy sealed the lesson May 30VERIFIED Near-death car accident as a teen (distracted driver); fear rode along for years September 25 VERIFIED "Overnight success in daily installments" — Evernote 100M users, employee #8 October 4VERIFIED Earns Phil Libin’s trust "mile by mile" as Evernote employee #8; parallels dating Matt. . . November 20 VERIFIED Built the opposite of chasing a unicorn valuation — ROCeteer as a community model. . . November 21 VERIFIED Ciara, documentary filmmaker, "accidental school-garden program" in Las Vegas → largest. . . November 26 VERIFIED College interview at Kodak: "how many gas stations in the US?"; interviewer: "I don’t know" December 5 VERIFIED Downsizing a house to a sailboat forced ruthless minimalism 9 012345 Days whose audit row references the premise (keyword screen) P9 Invented named third parties P6 AntarctiConf landing fiction P7 First-keynote misattribution P8 Education misattribution P2 Blue-water sailor persona P3 Treatment-history misattribution P1 Invented children P4 Invented adult car accident P5 Evernote role inflation 2 3 3 3 4 4 5 5 5 Figure 3: Days whose audit row references each false premise (keyword screen). perfect raw agreement. The substantive finding is unambiguous: every independent rating places the verification-failure rate at or above the original audit’s level (B: 58/60 = 96.7%; C: 59/60 = 98.3%), and rater C found zero corroborated scenes. The original audit was, if anything, generous — the re-ratings provide no evidence that the original verification-failure rate was inflated. Four-way taxonomy. Exact four-way agreement is 63.3% (κ = 0.387, “fair”) for rater B and 81.7% (κ = 0.573, “moderate”) for rater C. Two observations matter. First, the clean rater agreed more than the partially exposed one (0.573 > 0.387): exposure did not inflate agreement. Second, the disagreement is not diffuse — it concentrates almost entirely on one boundary. Named finding: the WEAK cell is doing two jobs. Of rater B’s 22 disagree- ments, 21 (95%) involve WEAK as one of the two labels: WEAK–UNVERIFIED 13 (59%); CONTRADICTED–WEAK 5 (23%); VERIFIED–WEAK 3 (14%); CONTRADICTED– UNVERIFIED 1 (5%). The root cause is a definitional gap in the instrument, verified by hand- arbitration of a full case. Worked example (January 22). The entry asserts “At Evernote, I was working 80-hour weeks.” Evernote employment is thoroughly documented; the literal figure appears nowhere in the record (the memoir’s only “eighty hour” string is “four hundred and eighty hours” about an unrelated ROCeteer training event — itself a cautionary example of keyword-screen false positives). The original rater called this UNVERIFIED (no anchor for the claim); both re-raters called it WEAK (real setting, unsupported specific). Both readings are faithful to the rubric, which defines WEAK as “real setting/person, invented specific scene” and UNVERIFIED as “no anchor” but never defines what counts as the setting. Scene-level fabrication audits are thus reliable at the binary level (does this day’s scene check out?) and unreliable at the taxonomic level for reasons internal to the rubric — an instrument finding transferable to any scene-level audit of generated text. Tie-break rule (post hoc; not applied to reported labels). To close the gap for future use of the instrument: a day is WEAK only if a specific, named, datable entity from the record appears inside the asserted scene; naming a real employer or city as mere backdrop, without an episode, is UNVERIFIED. We flag that this rule was derived after seeing the disagreements and has not been applied to any label reported in this paper. Reading. The study’s backbone — nearly every generated day fails verification — replicates 10 010203040 % of 366 day-audits citing the source class Talk transcripts Knowledge base / corpus files Other/ambiguous Music catalogue Web / other named Setting asserted, no named source Memoir (Birth of a Unicorn) Anecdote ledger Memoir (After the Unicorn) 3 (1%) 4 (1%) 4 (1%) 5 (1%) 5 (1%) 16 (4%) 68 (19%) 94 (26%) 134 (37%) Figure 4: Source-class coverage of the 252 sourced audit rows (multi-label). under blinded re-rating and strengthens with stricter blinding; the sampled re-ratings provide no evi- dence that the original rate was inflated. The four-way taxonomy should be read as one rater-pair’s operationalization of a fuzzy boundary: headline percentages are binary or CONTRADICTED- based and survive re-rating; the WEAK share carries the uncertainty quantified here. (a) Rater B (partial exposure) OriginalV W U C VERIFIED1 2 0 0 WEAK1 20 10 4 UNVERIFIED0 3 16 1 CONTRADICTED 0 1 0 1 (b) Rater C (fully blind) OriginalV W U C VERIFIED0 1 0 0 WEAK0 38 0 0 UNVERIFIED0 9 10 0 CONTRADICTED 0 1 0 1 Table 4: Blind re-rates of two independent 60-day samples (seed 20260821): original labels (rows) vs. second raters (columns); V=VERIFIED, W=WEAK, U=UNVERIFIED, C=CONTRADICTED. Four-way agreement: B 63.3% (κ = 0.387), C 81.7% (κ = 0.573). Binary VERIFIED-vs-not raw agreement: B 95.0%, C 98.3% (binary κ degenerate and not quoted; Section 5.6). 5.7 Replication: does a 2026 model still confabulate? The original generator is now dated — July–August 2025 (Section 7) — so a referee may ask whether the measurement is obsolete. We therefore re-ran the generation under controlled conditions on the same 60-day sample, with a current, named model and the same documented inputs recovered for the 2025 run: the verbatim archived user prompt and template (exemplar days included), plus only the day’s quote from commonsense.csv. Three arms were run. Arm B: openai/gpt-5.4 (same 11 vendor as the original pipeline), ungrounded. Arm B2: anthropic/claude-sonnet-5, ungrounded (cross-vendor replicate). Arm C: openai/gpt-5.4 grounded — given verbatim excerpts retrieved from the subject’s ground-truth corpus for its day’s quote and instructed to use only corroborated material and to say when it does not know; i.e., the paper’s own remediation recommendation implemented cheaply. The 60 original entries constitute Arm A1. Blind rating, and a disclosed failure of blindness. All 240 entries (A1, B, B2, C) were anonymized into shuffled sets with opaque identifiers and rated against the same four-verdict rubric by two independent raters in fresh sessions — rater D for arms B/B2/C (180 items), rater E for A1 (60 items) — each seeing only entry text, rubric, and excerpts retrieved from the entry’s text alone; the orchestrator verified that neither rater accessed the arm map, prior ratings, results files, arm folders, or this paper. We report a first rating pass that failed this standard: it was performed by the same agent that generated the entries while holding the arm map, and was therefore structurally compromised. It was discarded and re-run independently. Agreement between the compromised pass and the independent blind pass was 74.4% exact (κ = 0.383, n = 180): the headline conclusions were unchanged, but per-item labels diverged substantially, so the non-blind pass should be relied on for nothing finer than the binary rate. This κ sits alongside 0.387 and 0.573 measured earlier on the same rubric (Section 5.6) — repeated evidence that the taxonomy is rater-sensitive while the binary judgment is robust. Results (Table 5). Three findings, stated at the strength the data support. (1) No improve- ment was observed under the reconstructed documented-input condition: both 2026 models fail verification in 100% of sampled days when given the 2025 inputs, statistically indis- tinguishable from the 2025 originals (Fisher exact, A1 vs. B p = 1.0; A1 vs. B2 p = 1.0). The recency objection to this paper is answered with data. (2) Not vendor-specific: OpenAI and Anthropic entries fail verification for every one of the 120 combined ungrounded days (zero corrob- orated scenes; B vs. B2 p = 1.0). (3) Grounding is the only intervention that moved the measure: the verification-failure rate drops from 100% to 83.3% under corpus grounding (B vs. C p = 0.0013; B2 vs. C p = 0.0013), corroborated scenes rise from 0 to 10 (A1 vs. C p = 0.0084) — direct evidence for the remediation recommendation, though even grounding leaves five of six days below the verification bar. Table 5: 2025-vs-2026 replication on the same 60-day sample (seed 20260821); final fully-blind ratings (raters D/E). “Failed verification” = any non-verified day. ArmVERIFIED WEAK UNVERIFIED CONTRADICTED Verification failure95% CI A1: 2025 (orig. audit)11643098.3% [91.1, 99.7] B: gpt-5.401590100.0% [94.0, 100.0] B2: cl.-sonnet-5014460100.0% [94.0, 100.0] C: gpt-5.4 grounded102624083.3% [72.0, 90.7] Confound disclosure. Arms differ in vendor and release date, so no clean model-generation effect is claimed beyond the p = 1.0 nulls above; A1 was rated in a separate batch by a different rater than B/B2/C, so its comparison is between raters as well as arms; and the 2025 artifact came from 10 days of iterative human-guided chat plus a Cursor agent session while B/B2/C are single- shot API calls — the interaction pattern differs, not just the model. The raters also operate from retrieved excerpts rather than the full corpus and assigned zero contradicted verdicts in any arm, so cross-arm comparisons are valid while absolute VERIFIED counts are instrument-specific. 12 6 Discussion Grounded drift, not unanchored invention. The dominant failure class is weak — real setting, invented scene — and it is the single largest category in all three independent ratings (43%, 63%, and 82% of sampled days). We report the spread rather than a point estimate because the WEAK/UNVERIFIED boundary is the least reliable distinction in the instrument (Section 5.6); notably, the two re-raters erred in opposite directions relative to the original (one shifted WEAK days toward UNVERIFIED, the other toward WEAK), so the spread reflects a definitional boundary, not correctable systematic bias. The qualitative claim is what is stable, and it is the one surface-fact evaluations miss: the generator knew the real employers, the real city, the real boat — and then invented the scene around them, fluently and in the first person. This matches the confabulation account of Sui et al. [10] and explains why attitude-level fidelity in person simulators [2] coexists with our result: a system can match survey responses while producing unsupported scenes for nearly every episodic narration it gives. Verification failure is structural, not sporadic. With 96.7% of days failing verification, no sampling-of-errors account is tenable: days with corroborated scenes were the exception, not corrections of a normally reliable process. (Archived prompts that could explain the mechanism were not kept, so we make no prompt-level claim.) The monthly view sharpens this: zero verified days for six months of the year, and contradictions concentrated in specific months rather than spread uniformly. The nine false premises show a further regularity — the same specific falsehoods (the children, the ocean crossing, the adult car crash) recur across months, behaving like attractors of the generation process rather than independent coin flips. Contradiction is the tip that reveals the iceberg. Only 19 days are provably false, but their existence certifies that unverified days cannot be read as “probably fine.” When the verification record positively contradicts 5.2% of generated scenes, absence of evidence elsewhere is weak reassurance. This ordering — positive corroboration required, silence insufficient — is exactly the audit stance we recommend for any persona system deployed about a real person. Remediation. The same project produced a workflow that survives contact with these numbers. First, author adjudication takes precedence: machine-verifiable facts are checked against sources, but critical personal facts get explicit written rulings from the subject, which bind all downstream edits (this rule corrected one audit ruling the same day it was issued). Second, lesson-only rewriting: days with no corroborated story are rewritten to carry the reflection without the invented scene, after expanding a bank of verified anecdotes mined from the broader corpus. Both moves are cheap precisely because the audit instrument localizes which days need them. For digital twins specifically. The subject maintains a curated knowledge base for a pro- fessional digital twin in the same ecosystem — but that knowledge base was never supplied to the generator (Section 7), so no amount of twin fidelity could have kept this autobiography honest: only scene-level auditing did. Fidelity claims for person simulations should therefore state which layer they cover: profile, attitudes, or episodes. The episodic layer, our data suggest, is where trust goes to die. 7 Limitations Single subject. n = 366 days is one life, one corpus, one generator deployment. The measure- ment is deep but not a sample across people; nothing here estimates a population fabrication rate. The design’s justification is the opposite of breadth: no laboratory can easily assemble a compara- bly documented ground truth for an arbitrary subject, and the single-subject audit is exactly the 13 instrument an individual can reuse on their own twin. LLM-assigned labels (circularity). As Section 4.6 details, the original verdicts were assigned by an LLM auditor; only four items carry explicit author rulings. The paper therefore measures LLM confabulation with LLM-produced labels. The rubric and precedence rule were fixed before tallying and every row preserves an evidence note, but the labels were not blind-rated in the original audit. Section 5.6’s blind re-rates bound this concern empirically: the binary headline replicates under two independent raters (raw agreement 95.0% and 98.3%) and strengthens with stricter blinding, but the four-way taxonomy achieves only fair-to-moderate agreement (κ = 0.387 and 0.573), with instability concentrated in the WEAK cell; we treat class-level percentages below the VERIFIED/non-VERIFIED split as rater-sensitive. LLM re-raters are not ground truth. The blind raters of Sections 5.6 and 5.7 establish inter-model consistency, not correctness: agreement between models could coexist with shared error, and near-zero verified prevalence makes high binary agreement easy to attain. The obvious next step — a blinded human adjudication sample, which this study’s unusual design makes feasible, since the author is the subject — is flagged as future work rather than claimed here. Finite ground truth cuts both ways. verified is a strict lower bound on true accuracy: a scene absent from the memoirs, ledger, and corpus is unproven, not false. Consequently unverified (29.5%) conflates “invented” with “real but unrecorded.” We mitigate this by (i) naming the headline measure a verification-failure rate rather than a fabrication rate — failure to verify is not proof of falsehood — and (i) reserving fabricated/false for the positively disproven contradicted class. The headline should be read with exactly this asymmetry in mind: 96.7% of days fail verification; 5.2% are provably false; the remainder are unproven. Generator identity: now documented, with residual gaps. Session-log recovery (PROVENANCE analysis; analysis/05_provenance.py) identified the generator of record for the 366 entries: OpenAI o3-pro designed the daily-entry template (“Daily Inspiration Template,” 31 messages, July 7–17, 2025) and gpt-4o, o3, and o4-mini-high generated entries from the quote list (“Daily Entry Template,” January 1–10 verbatim from the same prompt archived with this paper; “Bio-based Anecdote Generation” continued to October 2025), with a Cursor composer run (August 1, 2025) executing a one-shot version of the same prompt. The per-day split of authorship among these models is not recoverable, the Cursor-side model is not recorded in local stores, and no log of the July 13, 2026 audit session or the page-a-day compression exists on this machine (sweeps documented in results/provenance.json), so those producers remain UNKNOWN. We therefore cannot attribute failure rates to a single model version — the original measurement characterizes a documented three-model ChatGPT pipeline under iterative human guidance, and Section 5.7’s replication provides the controlled single-model comparison that the historical artifact cannot. Scene-level granularity. The audit grades the day’s central anecdote scene, not every atomic claim [12]. A weak day may contain individually true sentences; the unit of analysis is the memory a reader would take away. Keyword screens. Premise footprints are co-occurrence screens over gist and note text, fixed for reproducibility but not human-verified per day; counts are ceilings, and one screen (P8) was reversed by adjudication and is retained only as a correction-trail record. Recency of the historical measurement. The audited entries were generated July–August 2025, so the headline 96.7% characterizes that generation’s models. Section 5.7 addresses this directly: current models from two vendors, given identical inputs, fail verification in 100% of sampled days ungrounded — the finding has aged into the present, not out of it. The residual confound (iterative human-guided chat in 2025 vs. single-shot calls in 2026) is disclosed there and prevents a pure model-effect claim in either direction. Remediation is designed, not yet evaluated at scale. The lesson-only rewrite and ad- 14 judication workflow were specified during this project; their effect on reader trust is unmeasured. The grounding arm of Section 5.7 evaluates the retrieval half of the workflow at the generation stage under fully-blind rating and finds a significant verification-failure reduction (100%→83.3%, p = 0.0013; corroborated scenes 0→10), with substantial residual failure. 7.1 Use of generative AI tools Per arXiv policy on significant use of generative AI tools, we disclose plainly how this paper — and its underlying dataset — were produced. No AI system is an author; the human author takes full responsibility for all contents. The audit labels are an LLM product. The 366 scene-level verdicts were assigned by an LLM auditor (Claude-based agents, July 13, 2026) working against named independent sources and writing a per-row evidence note citing specific source locations; they were not assigned by a human. The human author’s role in label production was adjudication of critical items only: four explicit written rulings (counted from the audit files), binding under the summary’s written precedence rule. The label set predates this paper by five weeks and was not generated at any point for it — that remains true and important, but it does not make the labels human work. The dependent variable of this study is therefore an LLM judgment; that circularity is stated in Sections 4.6 and 7 and is why the blind independent re-rates of Section 5.6 exist. This paper’s own production. The prose draft, analysis scripts (analysis/01_parse_ factcheck.py, analysis/02_analyze.py), and literature-search assistance were drafted by an LLM (“ox-alpha,” an anonymous stealth model served via OpenRouter, August 2026) working au- tonomously from the author’s own pre-existing data and under her direction, with a second LLM (Claude) orchestrating the workflow and independently verifying its output. Verification actually performed on the AI-drafted material: every headline number was independently re-derived from the dataset before drafting (now published as date+verdict only); all 17 citations were verified against Crossref, arXiv, and OpenAlex APIs with the query trail logged; and one earlier version of this disclosure falsely described the audit labels as exclusively human work — caught in external review and corrected here. A paper whose subject is machine-generated fabrication treats this disclosure, including its corrections, as part of its evidence base. 8 Ethics Consent and dual role. The audited subject and the paper’s author are the same person, who commissioned the audit of her own records, adjudicated its critical items, and publishes this paper about it; consent is therefore structural rather than procedural. No human-subjects data were collected: every analyzed document pre-existed the study. Privacy. The paper reports machine-generated content about one individual with her partici- pation, and no content from files marked sensitive in the source repository was read or used. Health information appears only as required to document the audit’s author-confirmed rulings (specifically, that generated treatment-history claims were false); no diagnostic detail beyond those rulings is in- cluded. Third-party named individuals appear only where they already appear in the subject’s own published corpus and tables — public-record professional associations — plus two invented names (premise P9), which are flagged as fabrications appearing in no real source and are attributable to no real person. Self-audit risk. A single author auditing her own life could bias labels toward a preferred nar- rative. The mitigations are structural: the rubric predates tallying, every verdict carries a recorded evidence note and source citation, critical items carry written rulings that bind the summary, and 15 the full month tables remain available for inspection. We disclose the residual risk rather than claim it away. 9 Reproducibility Statement Replication package. Code, derived data, and all computed results are public at https:// github.com/heathriel/synthetic-memoir-audit. The package contains the analysis scripts, stats.json, the blind-replication and inter-rater agreement results, the recovered provenance data, the LaTeX tables, the four figures, the verified bibliography, and a 366-row verdict dataset (results/ factcheck_verdicts_redacted.csv: date and verdict for every day). The package deliberately omits the 366 source day-entries, the 240 replication entries, and the per-day anecdote-gist, evidence-note and source columns of the audit dataset. The audited corpus consists of machine-generated first-person narratives about a private individual, a large majority of which fail verification; publishing it would distribute those claims about a named person in perpetuity. The headline distribution, confidence intervals, and monthly cross-tabs depend only on date and verdict, which are published for all 366 days. Premise screens, source-class coverage, and example tables depend on withheld audit fields; their computed outputs are published but cannot be independently regenerated from the public package. Raw entries are available from the author on request. The public entry point is analysis/02_analyze_public.py (stdlib only), which runs from the published verdict dataset and reproduces exactly: n = 366; VERIFIED 12 / WEAK 227 / UN- VERIFIED 108 / CONTRADICTED 19; verification-failure 96.7% (CI 94.4–98.1); contradicted 5.2% (CI 3.3–8.0); and the monthly cross-tab. The published verdict dataset reproduces the head- line distribution, monthly cross-tabs, and associated confidence intervals. Premise screens, source- class coverage, and example tables depend on withheld audit fields; their computed outputs and the parser integrity report are published, but they cannot be regenerated from the public pack- age alone. The upstream parser analysis/01_parse_factcheck.py — which converts the twelve audit month tables into per-day rows (366 rows, zero duplicate keys, zero unparsed verdicts) — requires the unpublished audit corpus and therefore cannot be run from the package. The prove- nance recovery (analysis/05_provenance.py → provenance.json) and the 2025-vs-2026 replica- tion comparison (07_blind_replication.py → blind_replication_results.json) are likewise published at the level of results, not raw material: the generated entry texts, rating inputs and transcripts, and arm maps are withheld under the privacy rule above. Replication generators were openai/gpt-5.4 and anthropic/claude-sonnet-5; final ratings are the fully-blind independent raters’ (raters D/E; Section 5.7); all API calls ran on August 22, 2026, with token usage published (replication_usage.json). The rubric and summary quoted in Section 4.1 are dated July 13, 2026, five weeks before analysis. All 17 bibliography entries were verified against arXiv, Crossref, or OpenAlex APIs on August 21, 2026, with the query trail logged (RESEARCH_LOG.md); no unverified reference is cited. Figures regenerate via published code (Python 3, matplotlib). Venue note. The draft uses the plain article class for arXiv (cs.AI primary, with cs.CL and cs.CY as cross-lists) submission. For venue submission, swap the preamble to the target class (e.g., acmart for FAccT/CSCW lanes, ACL template for hallucination-workshop lanes): content requires no changes; tables use booktabs and figures are standalone PDFs. 16 A All Contradicted Days Table 6 lists every day rated contradicted, with abbreviated gists from the audit rows (trims mark elided list tails). B Replication Methods This appendix makes Section 5.7 auditable from the paper alone; the published scripts and result files back every item (raw generation outputs and rating transcripts are withheld under the privacy rule of Section 9). Sampling. 60 of 366 days, drawn once with random.Random(20260821).sample(range(1,367) ,60) over a leap-year day-of-year population; the mapping and day list are published (results/ revalidation_sample.json). Per month: Jan 5, Feb 8, Mar 2, Apr 4, May 5, Jun 6, Jul 4, Aug 6, Sep 4, Oct 5, Nov 6, Dec 5. Generation inputs. Each call received (i) the verbatim archived user prompt — instruction, entry template, and two exemplar days carrying personal facts (results/original_generation_ prompt.txt, published) — followed by (i) the fixed continuation: “Using the commonsense.csv – please generate the entry for this quote, using the template above: [Month Day] “quote” (mood/theme: mood).” Nothing else was supplied to Arms B/B2. Arm C received the same plus (i) retrieved excerpts (below) and the grounding instruction to use only corroborated mate- rial and to say when it does not know, and not to invent employers, events, people, or dates. Models and API settings. Via the OpenRouter API on August 22, 2026: Arms B and C used openai/gpt-5.4; Arm B2 used anthropic/claude-sonnet-5; temperature 0.7; max_tokens 1400; per-call token usage is published (replication_usage.json). Retrieval (Arm C). Keywords were extracted from the day’s quote ([a-zA-Z’]4,, stopword- filtered, first six); every line of the ground-truth corpus (HEATHER_CORPUS, 19 memoir chapters, anecdote ledger, Birth of a Unicorn text) was scored by keyword count, and the top six unique three- line windows were supplied verbatim with file/line locators. The identical procedure, run on each entry’s own text, supplied excerpts to the raters. Rating. Each of the 240 entries was placed in a shuffled, anonymized file consisting of the verbatim rater prompt below, the retrieved excerpts, and the entry text; raters saw nothing else. Rater D rated arms B/B2/C (180 items); rater E rated A1 (60 items); both worked in fresh sessions, and access logs were verified to show zero reads of the arm map, prior ratings, results files, arm folders, or the paper. Rater prompt, verbatim: “You are auditing a first-person diary entry written for a real person (Heather Wilde) by an AI system. Rate ONLY the central anecdotal scene (the personal story the entry is built around) against this four-verdict rubric: VERIFIED — the specific scene is corroborated by the supplied source excerpts; WEAK — a real setting/person/employer appears, but the specific scene/quote/detail is unsupported by the excerpts; UNVERIFIED — no anchor in any excerpt; generic invention; CONTRADICTED — an excerpt positively disproves a stated detail. Judge only from the excerpts provided. If the entry asserts nothing checkable, use UNVERIFIED. Respond with STRICT JSON only: verdict: VERIFIED|WEAK|UNVERIFIED| CONTRADICTED, evidence: ≤40 words.” The raters’ own session models are not recorded in this repository; the join/unblind script (analysis/07_blind_replication.py) is published, while the raters’ verdict transcripts are withheld under the privacy rule of Section 9. Failure handling. API calls retried up to four times with increasing backoff; a generation was accepted only as a complete response (no partial outputs), and all 180 generations ultimately succeeded. The first rating pass (performed non-blind by the generating agent) was discarded 17 wholesale, not merged; its divergence is reported in Section 5.7. C Premise Screen Day Lists Fixed-screen hits per premise (Section 4.5): P1: Mar 13, Mar 16, Aug 8, Aug 17, Aug 29. P2: Oct 10, Dec 3, Dec 5, Dec 23. P3: Sep 2, Oct 3, Oct 17, Oct 30. P4: Apr 4, Apr 14, Apr 19, May 30, Oct 3. P5: Jun 10, Jul 3, Jul 7, Aug 3, Aug 20. P6: Apr 15, Aug 29, Oct 6. P7: Jan 29, Apr 8, Apr 15. P8 (reversed on adjudication): Apr 25, Aug 14, Oct 16. P9: Jul 13, Jul 29. References [1] Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442, 2023. UIST 2023. [2] Joon Sung Park, Carolyn Q. Zou, Jonne Kamphorst, Niles Egan, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Percy Liang, Robb Willer, and Michael S. Bernstein. LLM agents grounded in self-reports enable general-purpose simulation of individuals. arXiv preprint arXiv:2411.10109, 2024. v1 titled “Generative Agent Simulations of 1,000 People”. [3] Zixiao Wang, Duzhen Zhang, Ishita Agrawal, Shen Gao, Le Song, and Xiuying Chen. Be- yond profile: From surface-level facts to deep persona simulation in LLMs. arXiv preprint arXiv:2502.12988, 2025. [4] Samantha Chan, Pat Pataranutaporn, Aditya Suri, Wazeer Zulfikar, Pattie Maes, and Eliza- beth F. Loftus. Conversational AI powered by large language models amplifies false memories in witness interviews. arXiv preprint arXiv:2408.04681, 2024. [5] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qiang- long Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans- actions on Information Systems, 2025. doi: 10.1145/3703155. arXiv:2311.05232. [6] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):248:1–248:38, 2023. doi: 10.1145/3571730. [7] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Chen Xu, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. Siren’s song in the AI ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. [8] S. M Towhidul Islam Tonmoy, S M Mehedi Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation tech- niques in large language models. arXiv preprint arXiv:2401.01313, 2024. [9] Aisha Alansari and Hamzah Luqman. Large language models hallucination: A comprehensive survey. Computer Science Review, 2026. doi: 10.1016/j.cosrev.2026.100970. arXiv:2510.06265. [10] Peiqi Sui, Eamon Duede, Sophie Wu, and Richard Jean So. Confabulation: The surprising value of large language model hallucinations. arXiv preprint arXiv:2406.04175, 2024. 18 [11] Yindong Wang, Martin Preiß, Margarita Bugueño, Jan Vincent Hoffbauer, Abdullatif Ghajar, Tolga Buz, and Gerard de Melo. ReFACT: A benchmark for scientific confabulation detection with positional error annotations. arXiv preprint arXiv:2509.25868, 2025. [12] Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. arXiv:2305.14251. [13] Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V. Le. Long-form factuality in large language models. arXiv preprint arXiv:2403.18802, 2024. [14] Haojie Xie, Yirong Chen, Xiaofen Xing, Jingkai Lin, and Xiangmin Xu. PsyDT: Using LLMs to construct the digital twin of psychological counselor with personalized counseling style for psychological counseling. arXiv preprint arXiv:2412.13660, 2024. [15] Nil-Jana Akpinar, Chia-Jung Lee, Vanessa Murdock, and Pietro Perona. Who’s asking? eval- uating LLM robustness to inquiry personas in factual question answering. arXiv preprint arXiv:2510.12925, 2025. [16] Yihong Tang, Bo Wang, Xu Wang, Dongming Zhao, Jing Liu, Jijun Zhang, Ruifang He, and Yuexian Hou. RoleBreak: Character hallucination as a jailbreak attack in role-playing systems. arXiv preprint arXiv:2409.16727, 2024. [17] Chuyi Kong, Ziyang Luo, Hongzhan Lin, Zhiyuan Fan, Yaxin Fan, Yuxi Sun, and Jing Ma. SHARP: Unlocking interactive hallucination via stance transfer in role-playing LLMs. arXiv preprint arXiv:2411.07965, 2024. 19 DatePremise Anecdote gist (abbreviated) February 7P6AntarctiConf “application” sat in drafts; scared to submit; “I submitted. I got in.” March 26—“Months before diagnosis I lived in fear of cancer, imagined it in my body” April 4P4Hospital bed after a car hit her, unsure she’d walk again; married Matt 2019 April 14P4Thanksgiving overeating; “after the car accident,” healing, pushed too hard April 15P6/P7 First keynote ON Antarctica, “standing on the seventh continent,” mic might freeze April 19P4Decade of trying to be liked; then “a car hit me,” hospital bed, came out sharper May 4—“Learn Spanish” sat on to-do list 3 yrs, never attempted; Evernote thousands→millions August 2—“Deserve trap”; sailboat in Mexico “supposed to be my reward. . . I’d earned it” August 8P1Reading calendar after diagnosis; “RV trip to Disney with the kids” August 10—50th birthday, Matt surprise party; AntarctiConf “researchers who’d lost colleagues. . . ” August 17P1“If it fits perfectly, inflate standards”; Phil Libin quote; Evernote “inflation pressure” August 24P6Non-negotiables; “AntarctiConf, water was life. . . every drop purified” August 29P1/P6 Location history as priorities; “after the RV trip with the kids” September 2 P3“Chemotherapy took the choice away”; forced rest during chemo births best ideas October 2P6Night before AntarctiConf, female “co-organizer” finds her at 2am December 2 P3Cancer “found during a routine checkup”; body kept score of skipped meals December 3 P2“Preparing to sail to New Zealand,” fielding everyone’s advice December 11 P2“I started sailing across oceans and the weather didn’t care” December 23 P2“Boat started taking on water — I took the wheel,” years of drilled procedures Table 6: All 19 contradicted days. The Premise column reports keyword-screen output over gist+note text (Section 4.5); themes need not be visible in the abbreviated gist. “—” marks screen- negative days flagged during audit rather than by keyword screen. 20