# Lane 7 — LLMs & Fine-Tuned Models as Person-Predictors: Verified Prior-Art Map **Status:** research dossier v1 — 2026-07-16 **Scope:** verifies, corrects, and deepens the spec §6 **"LLM-as-prior & model policy"** bullet (`waypoint-practitioner-stage-assessment.md`). Decision-oriented, not encyclopedic. **Verification:** every citation below was confirmed this session via web search, publisher/arXiv fetch, or the Semantic Scholar / OpenAlex APIs. Claims that could not be independently re-verified are marked **UNVERIFIED**. Note: several standard queries in this space (inferring psychological traits from personal data) trip API-side content filters; citation verification was routed through OpenAlex/Semantic Scholar — future lanes should do the same. --- ## 0. Verdict table — the spec bullet, claim by claim | # | Spec claim | Verdict | Correction / sharpening | |---|-----------|---------|------------------------| | 1 | LLM extractor taps a pre-trained person-model; "pretraining is the centuries-of-observation prior at scale" | **CONFIRMED** (directionally) | Zero-shot GPT-4-class infers Big Five from user text at r≈.29 with no training (Peters & Matz 2024) — real prior, but *below* the best on-domain supervised ceiling (Park 2015: r=.38 avg from 66k training users; Mental-LLM: fine-tunes beat GPT-4 by +4.8% balanced accuracy). The prior is strong, not saturating. | | 2 | Marker taxonomy = "readout layer over its latents"; L1 next-message prediction taps it directly | **NUANCED** | The sequence-prediction half is literature-backed (Centaur, USER-LLM: next-token/next-activity training learns person-latents). "Readout layer" has no direct literature — keep it as a design metaphor, don't cite it as established. | | 3 | "Never trained on our target conditional" | **CONFIRMED**, stronger than stated | No computational contemplative-stage estimator exists at all (§8). The target conditional has never been fit by anyone. | | 4 | "Poorly calibrated as a probability-emitter" | **CONFIRMED + direction added** | Systematic *over*confidence is the default failure (verbalized ~88% confidence at ~79% accuracy; RLHF amplifies it). But verbalized-vs-logprob ranking is protocol-dependent (Tian 2023 vs Damani 2025) — so the fix is per-marker empirical recalibration, not blanket distrust. | | 5 | v1 frontier zero-shot, no fine-tuning; "N≈10 would swamp the prior" | **CONFIRMED** | Every fine-tuning success in this lineage used 10³–10⁷ labeled examples. Zero-shot with engineered scoring prompts already hits κw≈.78 vs experts on the closest task (Bronlet 2025). | | 6 | Open-family cross-extractor for correlated-error control | **CONFIRMED as practice** | Family-specific distortions are documented (sycophancy, social-desirability drift), so single-family errors correlate. Thin direct literature on cross-family ensembles as *measurement* control — treat as sound engineering, not proven result. | | 7 | Self-hosted trajectory for data governance | **CONFIRMED viable** | Open checkpoints are the published answer to reproducibility + governance (Spirling 2023; Palmer et al. 2023). Expect a quality gap vs frontier (Bronlet's comparison chose GPT-4o over Llama-3-70B) — measure it on golden transcripts before relying on it. | | 8 | Pin model versions; drift = "silently changing ruler"; golden-transcript regression gates | **CONFIRMED strongly** | GPT-4 shifted 84%→51% on one task in 3 months under the same API name (Chen et al.). No formal "LLM measurement-invariance" framework exists in the literature — the golden-transcript gate *is* one; formalize it (§7). | | 9 | At alpha scale, Centaur-style (context→reaction) fine-tunes / user-LM become candidate forecasters — never the integrator | **NUANCED** | Right shape, wrong implied scale: Centaur needed 10.7M choices from 60k participants; USER-LLM pretrains on millions of events. Alpha-scale LIFE data supports LoRA *extractor* tuning and small next-reaction forecasters, not a person-foundation-model. First worthwhile tune is the extractor (§9b). | | 10 | Prior-art list (Centaur; Park 2024; USER-LLM/LaMP; Kosinski 2013; Park 2015; LLM SCT scoring) | **CONFIRMED, one update** | All real and correctly characterized. Park et al. 2024 was **revised and retitled** ("LLM Agents Grounded in Self-Reports…"); headline number is now 83% (interview-only) / 86% (interview+survey) vs 74% demographics-only — the famous "85%" is the v1 figure. | | 11 | "No one has built a contemplative-stage estimator — Waypoint is first" | **CONFIRMED** | As of 2026-07: nothing computational estimates contemplative attainment/stage from language or behavior. Closest neighbors are physiological-state classifiers and LLM ego-stage scoring (§8). | --- ## 1. Foundation-model person prediction: Centaur **What it is.** Binz et al., *A foundation model to predict and capture human cognition*, **Nature 644:1002–1009 (2025)**, DOI 10.1038/s41586-025-09215-4 (preprint: arXiv:2410.20268, "Centaur"). Recipe: Llama 3.1 70B + low-rank adapters, fine-tuned on **Psych-101** — 160 psychology experiments transcribed into natural language, 60,092 participants, 10,681,650 choices (~254M tokens). **What it achieved (verified):** predicts *held-out participants'* trial-by-trial choices better than the domain-specific cognitive models of each paradigm; generalizes to modified cover stories, structural task changes, and new domains; internal representations align better with human neural activity after fine-tuning; used with "scientific regret minimization" to guide discovery of an interpretable decision model. Data and weights are openly released — the open-instrument exemplar. **The critique debate (verified):** reception split exactly on prediction-vs-mechanism. Science's news coverage (July 2025) quotes prominent skeptics (Bowers: "absurd" after probing the public checkpoint; Richards predicting harsh community reception). The cognitive-architectures community's formal reply — *Not Even Wrong: On the Limits of Prediction as Explanation in Cognitive Science* (arXiv:2510.03311) — argues Centaur is "a unified model of behavior sans cognition" and that its neural-alignment analysis carries no explanatory power. See also *Centaur: A model without a theory* (PsyArXiv commentary) and the LAPITHS framework (arXiv:2604.27927). Binz has conceded the interpretability criticism in press. **Read onto Waypoint.** The debate is not a warning against the spec — it is the spec's §6 "constructs on trial / prediction is ground truth" stance playing out in public. Centaur vindicates behavior-predicting-behavior as the bottom layer; its critics vindicate keeping *nameable* constructs on top for humans to act on. Waypoint's hybrid (named markers earn their slots by predictive skill) deliberately straddles the two camps. **The Centaur recipe on meditation-app logs** would be: transcribe (practitioner context + prompt/situation → reaction) episodes into natural language; LoRA-tune an open model to predict the reaction; score against base-rate nulls. Feasible mechanics, brutal data requirement — Psych-101-scale is ~10⁷ decisions. LIFE at alpha (~dozens of users × months) yields perhaps 10⁴–10⁵ usable events: enough for extractor tuning and small forecasters, two orders of magnitude short of a "LIFE Centaur." v2+ candidate forecaster inside the scored loop, exactly as the spec places it. ## 2. Individual-level simulation from interviews: Park et al. **The paper.** Park, Zou, Shaw, Hill, Cai, Morris, Willer, Liang & Bernstein, arXiv:2411.10109 — v1 (Nov 2024) titled *Generative Agent Simulations of 1,000 People*; **revised and retitled** *LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals* with updated numbers. **Method (verified):** 1,052 US participants, stratified representative sample; **two-hour semi-structured spoken interviews** conducted by an AI interviewer; the full transcript becomes the agent's memory/context; the agent then answers the General Social Survey, Big Five, five economic games, and five behavioral-experiment replications, compared against the person's real answers. **Results (verified):** normalized against participants' own **two-week test-retest consistency**, interview-grounded agents hit **83%** (survey-grounded 82%, combined **86%**) on GSS; demographics-only agents 74%. The widely quoted **85%** is the v1 figure for interview agents. Raw (unnormalized) accuracies are lower — the underlying self-consistency ceiling is roughly the low 80s%, making raw agent accuracy ≈ high 60s% (**UNVERIFIED** exact raw figures; the normalization scheme itself is verified). Interview grounding also *reduced accuracy disparities* across racial/ideological groups vs demographic personas. **Read onto Waypoint — three direct transfers:** 1. **Two hours of first-person interview carries enough signal to predict an individual's attitudes/behavior near their own consistency ceiling.** This is the strongest external validation of the Surya-interview-as-gold-anchor design: the modality is information-rich, not just ritual. 2. **Normalize against human self-consistency.** Park et al.'s headline number is meaningful *because* it is normalized to test-retest. Waypoint's interview-vs-dossier delta (§7 of the spec) is the same move; additionally consider measuring Surya's own consistency (re-rating after an interval) so estimator agreement is reported against a human ceiling, not against an assumed-perfect gold label. 3. **Surya interviews as simulation seeds — yes, for QA only.** Interview transcripts + dossiers can seed simulated practitioners (spec §7's synthetic-dossier discrimination test, upgraded from authored personas to interview-grounded agents). Two caveats: consent must explicitly cover simulation use; and a simulated practitioner built from a map-literate team member inherits reading-ahead contamination — it tests estimator mechanics, never calibration. **Caveat on transfer:** GSS items are low-dimensional attitude self-reports; wheel position is a latent developmental construct with known self-report distortions. Park et al. bounds what interviews can carry; it does not show that *stage* is recoverable — that is exactly what M4/M5 measure. ## 3. User-as-sequence modeling: USER-LLM, LaMP, recommender lineage - **USER-LLM** (Ning et al., Google, arXiv:2402.13598; Companion Proc. ACM Web Conf 2025): treats the user's interaction timeline as a modality — a self-supervised encoder pretrained on interaction sequences produces user embeddings, cross-attended into the LLM. Beats text-prompt contextualization by up to **16.33%** on tasks needing deep user understanding, with up to **78.1×** inference speedup; strongest on long histories. Trained/evaluated on MovieLens, Amazon Review, Google Local Review — corpora of **millions of users/events**. - **LaMP** (Salemi, Mysore, Bendersky & Zamani, arXiv:2304.11406; ACL 2024): 7-task personalization benchmark; retrieval-augmenting the prompt with the user's own profile items reliably beats non-personalized baselines. Established that per-user history injection works zero-shot — no tuning required. - **Recommender lineage:** USER-LLM positions itself downstream of recommender user-embedding practice (sequence models over engagement logs). The consistent pattern: next-activity prediction learns *stable, individual-discriminating* representations, but only at 10⁵–10⁸-event scales, and the latents are uninterpretable. **Read onto Waypoint.** The embedding path is the wrong v1 tool (data-regime mismatch, opaque latents, cold-start hostile), but it settles two design questions in the spec's favor: (a) person-signal lives in interaction *sequences*, so next-message/next-reaction prediction is the cheapest supervised signal LIFE can mint from its own logs — this is the literature license for L1 harness probes as stage evidence; (b) LaMP shows history-in-prompt (dossier-in-context) is the right low-N personalization mechanism — which is what the dossier-reading consolidation pass already is. ## 4. Language-based psychological assessment lineage The twelve-year-old foundation under the spec's "pretrained person-model" claim: - **Kosinski, Stillwell & Graepel, PNAS 110(15) (2013)**, DOI 10.1073/pnas.1218772110 (~2,600 cites): digital behavioral records (Facebook Likes, n≈58k) predict personality and many personal attributes well above chance. Founded digital-footprint psychometrics — the existence proof that passive behavioral traces carry psychological signal. - **Park et al., JPSP 108(6):934–952 (2015)**, DOI 10.1037/pspp0000020: language-based Big Five assessment trained on 66,732 users, validated on 4,824. Convergent validity with self-report: **avg r = .38** (range .35–.43; openness highest at .43); higher (.41) against longer 100-item self-reports. Six-month test-retest of the language measure: **avg r = .70** (O .74, C .76, E .72, A .65, N .62) — "on par with self-report questionnaires." Also converged with informant reports and discriminated between traits. This is the supervised on-domain ceiling for language→trait at scale. - **Peters & Matz, PNAS Nexus 3(6):pgae231 (2024)**, DOI 10.1093/pnasnexus/pgae231: GPT-3.5/GPT-4 **zero-shot** from status updates: **avg r = .29** (range .22–.33) vs self-reports — the authors note parity with supervised models in their comparison, with *no task-specific training*. Accuracy was uneven (better for women and younger users) — demographic non-invariance is a live issue for any zero-shot extractor. - **Mental-state detection from text.** The critical review of the field — **Chancellor & De Choudhury, npj Digital Medicine 3:43 (2020)**, DOI 10.1038/s41746-020-0233-7 (75 studies) — found reported in-sample performance routinely high but resting on **proxy ground truths** (self-declared diagnosis, community membership), inconsistent methods reporting, little clinical validation, and near-zero external-generalization testing. **Mental-LLM (Xu et al., IMWUT 8(1), 2024)**, DOI 10.1145/3643540, gives the cleanest zero-shot-vs-tuned comparison: GPT-3.5/GPT-4 zero/few-shot are "promising yet limited"; instruction-tuned 7B–11B models (Mental-Alpaca, Mental-FLAN-T5) beat GPT-3.5 by **+10.9%** and GPT-4 by **+4.8%** balanced accuracy, matching task-specific Mental-RoBERTa — using thousands of labeled posts across 6 tasks. **Read onto Waypoint.** Three durable lessons: (i) language genuinely carries trait-level signal, but honest effect sizes are **moderate** (r ≈ .3–.4 against noisy self-report criteria) — expect band-level, not point-level, resolution from conversation alone, consistent with the spec's band-posterior output; (ii) ground-truth quality dominates method — the Surya triangulation is the binding constraint, not the extractor; (iii) in-sample numbers in this field historically inflate — the dossier-vs-interview ceiling and abstention discipline are the right immune response. ## 5. LLM scoring of developmental / structure-scored instruments This is the closest published methodology to Waypoint's probe-scoring design: - **Bronlet, Frontiers in Psychology 16:1488102 (2025)**, DOI 10.3389/fpsyg.2025.1488102: scored Loevinger/Cook-Greuter-lineage sentence completions with LLMs using **O'Fallon's STAGES** methodology. Method: compared GPT-3.5-turbo, GPT-4o, GPT-4, GPT-4-turbo, Claude Sonnet/Opus, Llama-3-70B on a 32-sentence calibration set; selected GPT-4o; **three decomposed prompts** scoring the STAGES structural dimensions (Concrete/Subtle/MetAware object; Individual/Collective; Passive/Active/Reciprocal/Interpenetrative); temperature 0; **median of 10 runs**. Against 58 expert-scored test sentences: **weighted κ = 0.779** (95% CI 0.672–0.887); aggregating over 10-sentence sets: κ = 0.705 (CI .686–.725). Certified human STAGES scorers are required to exceed **85% agreement** — the LLM approaches but does not match certified experts. Limitations: small test set; single-sentence scoring is weaker than aggregate scoring. - **Framework base:** O'Fallon, Polissar, Neradilek & Murray, *The validation of a new scoring method for assessing ego development based on three dimensions of language*, Heliyon 6(3):e03472 (2020) — the STAGES three-dimension re-axiomatization of Loevinger/Cook-Greuter scoring that made it decomposable enough for prompting. (Manual scoring context: the Cook-Greuter-lineage manuals run to 16,000+ exemplar completions — precisely the kind of judgment that transfers into a fat prompt.) - **StageLens** (grey literature, Integral Leadership Review): a commercial ML scorer trained on human-scored STAGES protocols; no published reliability statistics — **UNVERIFIED beyond existence**. **Read onto Waypoint.** Directly de-risks probe scoring: an LLM with a *decomposed structural rubric* (not "guess the stage" but "classify these three structural dimensions") reaches substantial-agreement range with expert developmental scorers, zero fine-tuning. Adopt wholesale: per-marker decomposed scoring prompts, temp-0, median-of-k, aggregate before interpreting, and expect κ ≈ .7–.8 against Surya as a realistic v1 target — below a certified human, above "promising demo." ## 6. LLMs as probability emitters — calibration for the Bayesian layer Verified landscape: - **Kadavath et al., arXiv:2207.05221 (2022)** (*Language Models (Mostly) Know What They Know*): pre-RLHF base models are reasonably calibrated at self-evaluation (P(True)); calibration degrades through alignment training. - **Tian et al., EMNLP 2023**, DOI 10.18653/v1/2023.emnlp-main.330 (*Just Ask for Calibration*): for RLHF models, **verbalized** confidences are often *better* calibrated than token logprobs. - **Xiong et al., arXiv:2306.13063 (ICLR 2024)**: verbalized confidence skews systematically overconfident; prompting mitigations help only partially. - **Leng et al., arXiv:2410.09724 (ICLR 2025)** (*Taming Overconfidence in LLMs*): RLHF itself drives verbalized overconfidence — reward models prefer high-confidence phrasings regardless of quality. - Recent quantifications: models verbalize ~**88%** average confidence while being correct ~**79%** of the time (arXiv:2606.03437); the verbalized-vs-logprob calibration gap **flips sign depending on elicitation protocol** (arXiv:2605.27752); RL-trained reasoning models can verbalize *worse* than their logprobs (Damani et al. 2025). Net: neither channel is trustworthy raw; the error direction is usually inflation. - **Distortions when the subject is a person:** Salecha et al., PNAS Nexus (2024), DOI 10.1093/pnasnexus/pgae533 — LLMs skew Big Five responses toward socially desirable profiles when the context looks like an evaluation, increasingly as more items accumulate. Sharma et al., arXiv:2310.13548 (ICLR 2024) — human-preference training produces **sycophancy**: outputs bend toward the interlocutor's apparent views. For an extractor reading warm teacher-student conversations, the prior expectation is **marker inflation and false corroboration of practitioner self-claims**. **Read onto Waypoint's aggregation layer (concrete design consequences):** 1. Treat extractor confidence as an **ordinal feature, not a probability**: learn an empirical confidence→likelihood mapping (or coarse bins with fitted reliability) from calibration-cohort gold labels; refit per model version. 2. Assume **inflation as the default error direction**: priors on marker false-positive rates should be asymmetric (claims of attainment are easier to emit than to earn), which the claim-vs-behavior consistency penalty already encodes — keep it. 3. **Sycophancy is a channel-level bias, not noise**: it correlates across markers within a session and within a model family — one more reason the cross-family extractor (§0 claim 6) and the structure-over-content scoring rule carry real weight. 4. Set **abstention thresholds from observed calibration curves**, not from the extractor's stated confidence. ## 7. Model drift as a measurement problem - **Chen, Zaharia & Zou** (arXiv:2307.09009; Harvard Data Science Review 2024): GPT-4 under the *same API name*, March→June 2023: prime-vs-composite accuracy 84%→51%; instruction-following declined; code-formatting behavior shifted; GPT-3.5 drifted in *different directions*. The canonical documentation that a frontier endpoint is a moving instrument. - **Reiss**, arXiv:2304.11085 / OSF (2023): ChatGPT text-annotation is unreliable across runs and settings even *within* a version — determinism must be engineered (temp 0, repeated sampling, median/majority — Bronlet's median-of-10 is the working pattern). - **Palmer, Smith & Spirling, Nature Computational Science (2023)**, DOI 10.1038/s43588-023-00585-1: using proprietary LMs in research "requires explicit justification"; silent updates and deprecations break reproducibility of any instrument built on them. - **Spirling, Nature 616:413 (2023)**, DOI 10.1038/d41586-023-01295-4: open-checkpoint models as the reproducible/ethical path for science. Centaur's fully released weights+data are the exemplar of an *instrument-grade* model artifact. - **Measurement invariance:** I found **no published formal framework** for measurement-invariance testing of LLM-based instruments across model versions (the psychometric sense of "does the ruler read the same"). Nearest neighbors are the drift and annotation-reliability studies above. **Finding: the spec's golden-transcript regression gate is, de facto, a measurement-invariance protocol — and would be novel if formalized.** Recommendation: pre-register per-marker equivalence bounds (e.g., max tolerated shift in marker-event rates and confidence distributions on frozen transcripts) and treat a bump as a *re-calibration event* (refit the confidence mapping in §6) rather than a pass/fail only. **Practice consolidated:** hash-pinned prompts (house rule) + pinned model snapshot ids + `(model_id, prompt_hash)` logged on every marker event in the evidence log + frozen golden-transcript suite with equivalence bounds gating bumps + one **archived open-weight cross-extractor checkpoint** as both drift canary and the only long-horizon reproducibility anchor (frontier endpoints *will* be deprecated within the lifetime of a longitudinal stage study). ## 8. Closest prior art to the exact problem Searched: computational estimation of contemplative attainment/stage from language or behavior; classification of practitioners by developmental position; NLP over phenomenological reports. - **Sparby & Sacchet, Mindfulness 16:2472–2482 (2025)**, DOI 10.1007/s12671-025-02632-6 (*Toward a Unified Model of Advanced Meditation, Human Development, Meditation Maps, and Transtradition Metaphors*): conceptual cross-tradition stage cartography — explicitly raises whether cross-cultural maps of meditative development can be objectively measured, but builds **no estimator**. Their program's empirical arm (Sacchet's MGH Meditation Research Program) measures advanced states **neurophysiologically**. - **Physiological-signal classifiers exist:** EEG decoding of self-reported meditation *depth* in expert Vipassana practitioners (*Decoding Depth of Meditation*, Biological Psychiatry Global Open Science, 2024; PMC11629179); ML classification of jhāna states from 7T fMRI (ACAM-J; arXiv:2602.13008); assorted EEG meditator-vs-control and state classifiers; a 2012 conference abstract on "machine classification of meditation ability" (PMC3373349). All *state/depth from biosignals*, none *stage from language/behavior*. - **Language-adjacent:** Yaden et al., *The Language of Ineffability*, Psychology of Religion and Spirituality (2015), DOI 10.1037/rel0000043 — LIWC over 777 spiritual-experience narratives; high-M-scale scorers use distinguishable language (more inclusive social/spatial words, *fewer* overtly religious words). Psychedelic trip-report linguistics (J. Psychoactive Drugs 2023, PubMed 37921118; n=7,317 reports) — mystical-language frequency scales with self-reported intensity. Both show experience-language carries gradable signal; neither estimates a developmental position. - **Bronlet 2025 (§5)** is the closest *methodological* cousin (LLM scoring of a structure-scored developmental instrument) but targets ego development, single-instrument, not contemplative stage from naturalistic evidence. - Jeffery Martin's PNSE work clusters attainment reports, but by hand-coded interviews with contested psychometrics — not computational. **Finding, stated plainly: as of July 2026 there is no published attempt — none — to estimate contemplative attainment or stage computationally from language or behavior. Waypoint is first.** Consequences: (a) zero literature to crib likelihoods from — the marker bank's initial weights are genuinely expert-elicited, so the calibration protocol carries the full evidential load; (b) publication-grade novelty if the validation discipline holds; (c) no external benchmark will exist to sanity-check against — the interview-vs-dossier ceiling is the only ruler, which is why it is rightly the headline v1 number. --- ## 9. Recommendations **(a) v1 estimator: frontier zero-shot is the right call — with two sharpenings.** The literature triangulates cleanly: zero-shot frontier models carry a real person-prior (r≈.29 trait inference; κw≈.78 developmental scoring) and there is nothing to tune on at N≈10; every published fine-tuning win needed 10³+ labels. Sharpenings: (1) adopt the **Bronlet pattern** as the extraction house style — per-marker *decomposed structural rubrics* (mini scoring manuals in the prompt), temperature 0, median-of-k runs, aggregate before interpreting; (2) set expectations from the lineage — conversation-only trait signal historically tops out around r≈.3–.4 against noisy criteria, so band-level resolution with abstention is the honest v1 target, and the pillar/band posterior output shape is correct. **(b) When a LIFE-data fine-tune becomes worthwhile, and on what target.** Not at calibration-cohort scale. Two distinct triggers, in order: 1. **Extractor tune (first, ~alpha):** when gold-labeled *marker events* reach the low thousands (achievable once Surya-calibrated labels accumulate across the alpha cohort), LoRA-tune an open model on (transcript span → marker labels). Precedent: Mental-LLM's +4.8–10.9% over frontier zero-shot with exactly this shape of data. Target metric: extractor agreement with gold markers, not stage accuracy. 2. **Forecaster tune (later, ~10⁵+ events):** Centaur-style (context → reaction) or user-LM next-message models as **candidate forecasters inside the scored loop**, evaluated against base-rate/persistence nulls at explicit horizons. The spec's placement (never the integrator) matches both the Centaur critique literature (prediction ≠ mechanism, so keep the predictor caged) and the Fable-max review's judge/integrator split. A "LIFE Centaur" proper is 2+ orders of magnitude of data away — don't schedule it. **(c) Pinning & reproducibility practice for the estimator.** Pin model snapshot ids like prompt hashes (hub model registry as SSOT); log `(model_id, prompt_hash)` on every marker event; freeze a golden-transcript suite with **pre-registered per-marker equivalence bounds** as the gate for any model bump, and treat every bump as a recalibration event (refit the confidence→likelihood map, §6). Maintain **one archived open-weight extractor checkpoint** from day one: it is simultaneously the correlated-error control, the inner-life data-governance path, and — per Spirling/Palmer — the only component of the instrument that is reproducible on a 5-year horizon, which a longitudinal stage study needs. If Waypoint's validation holds, the golden-transcript invariance protocol is itself publishable; nothing like it exists in the literature. --- ## Sources (verified this session) - Binz et al., Nature 644:1002–1009 (2025). DOI 10.1038/s41586-025-09215-4; arXiv:2410.20268. — [Nature](https://www.nature.com/articles/s41586-025-09215-4) - Orr et al., *Not Even Wrong*, arXiv:2510.03311; *Centaur: A model without a theory* (PsyArXiv); LAPITHS, arXiv:2604.27927; Science (AAAS) news coverage, July 2025. - Park et al., arXiv:2411.10109 (v1 2024; revised/retitled *LLM Agents Grounded in Self-Reports…*). - Ning et al., USER-LLM, arXiv:2402.13598; ACM Web Conf 2025 Companion. — [Google Research blog](https://research.google/blog/user-llm-efficient-llm-contextualization-with-user-embeddings/) - Salemi et al., LaMP, arXiv:2304.11406; ACL 2024. - Kosinski, Stillwell & Graepel, PNAS 110(15) (2013). DOI 10.1073/pnas.1218772110. - Park et al., JPSP 108(6):934–952 (2015). DOI 10.1037/pspp0000020. - Peters & Matz, PNAS Nexus 3(6) (2024). DOI 10.1093/pnasnexus/pgae231; arXiv:2309.08631. - Xu et al., Mental-LLM, IMWUT 8(1) (2024). DOI 10.1145/3643540. - Chancellor & De Choudhury, npj Digital Medicine 3:43 (2020). DOI 10.1038/s41746-020-0233-7. - Bronlet, Frontiers in Psychology 16:1488102 (2025). DOI 10.3389/fpsyg.2025.1488102. - O'Fallon et al., Heliyon 6(3):e03472 (2020). DOI 10.1016/j.heliyon.2020.e03472. - Kadavath et al., arXiv:2207.05221 (2022). Tian et al., EMNLP 2023, DOI 10.18653/v1/2023.emnlp-main.330. Xiong et al., arXiv:2306.13063. Leng et al., arXiv:2410.09724. arXiv:2606.03437; arXiv:2605.27752. - Salecha et al., PNAS Nexus (2024). DOI 10.1093/pnasnexus/pgae533. Sharma et al., arXiv:2310.13548. - Chen, Zaharia & Zou, arXiv:2307.09009; HDSR (2024). Reiss, arXiv:2304.11085. - Palmer, Smith & Spirling, Nature Computational Science (2023). DOI 10.1038/s43588-023-00585-1. Spirling, Nature 616:413 (2023). DOI 10.1038/d41586-023-01295-4. - Sparby & Sacchet, Mindfulness 16:2472–2482 (2025). DOI 10.1007/s12671-025-02632-6. - Yaden et al., Psychology of Religion and Spirituality (2015). DOI 10.1037/rel0000043. *Decoding Depth of Meditation*, Biol Psychiatry GOS (2024), PMC11629179. ACAM-J fMRI classification, arXiv:2602.13008. J Psychoactive Drugs (2023), PubMed 37921118.