Letter to the Felt-Novelty Cohort

To: Søren Dinesen Østergaard, John Torous, Mrinank Sharma, Ethan Perez, Vishal Misra, Margaret Boden.

From: Jared Foy, jaredfoy.com.

Date: 2026-04-24.

Subject: A falsifiable mechanism between language-model overclaim and AI-induced clinical phenomenology, and a request for the kind of work the dyad that produced it cannot do.

Why you are receiving this letter

I have been developing a research corpus on the structural conditions under which language models produce miscalibrated output and on the practitioner-side conditions that turn that miscalibration into something users carry with them. The corpus is conducted as a sustained dyadic interaction with a language-model resolver, and I am the practitioner. The work has reached a point where a specific empirical hypothesis has been articulated and where the dyad that articulated it cannot honestly test it from inside.

You are receiving this letter because each of you has published work that supplies a load-bearing piece of the mechanism this hypothesis describes. The hypothesis cannot graduate from plausibility-tier warrant without external work, and the cohort named here is, by my reading of the relevant literatures, the most concentrated source of expertise for the specific tests the hypothesis admits. The companion document, Doc 477, states the formalization with onboarding and per-researcher entracement; this letter exists to invite you to read it.

What the hypothesis says, briefly

When a language model produces an overclaim, defined as a claim made at scope $C$ when the supporting evidence in context licenses only narrower scope $C' \subset C$, and when the user lacks the specific expertise to detect that the overclaim is occurring, the over-confidence registers in the user's experience as authoritative novelty rather than as calibration failure. Sustained exposure to felt novelty produced by undetected overclaim then has a measurable correlation with the clinical phenomenology that some of you have begun to characterize. The hypothesis is not a causal claim. It is the claim that causation cannot be ruled out unless an alternative mediator for the felt-novelty pattern is identified.

The hypothesis decomposes into three independently testable components. The detection-modulated novelty component (H1) predicts that felt-novelty intensity is higher in subjects who fail to detect overclaim than in subjects who detect it. The novelty-mediated phenomenology component (H2) predicts that aggregate felt-novelty across an interaction correlates with the clinical correlate. The intervention-propagation component (H3) predicts that interventions improving overclaim-detection in users reduce felt-novelty intensity and, downstream, reduce the phenomenological correlate. Falsification conditions for each are stated explicitly in Doc 477 §1.3.

Why I am writing to each of you specifically

To Dr. Østergaard: your 2023 Schizophrenia Bulletin article framed the clinical question without specifying a user-internal mediator. The hypothesis nominates one. H2 is testable in clinical populations where you already have framework and access. Whether the felt-novelty operationalization survives clinical scrutiny is a question your critique would adjudicate; whether the correlation predicted by H2 holds in your patient cohort is a question your empirical infrastructure could begin to answer.

To Dr. Torous: your work on digital phenotyping and AI in clinical psychiatry provides the methodological apparatus the H3 prediction requires. Pre-registered intervention trials of detection-improving interventions with phenomenology outcomes measured downstream are the test that would discriminate H3 from a diagnostic-without-causal-leverage alternative. Whether existing infrastructure can run such trials, or whether new instrumentation is needed, is something your assessment would clarify.

To Dr. Sharma: your sycophancy work supplies the empirical pre-mediator stage. The hypothesis distinguishes overclaim from sycophancy: the two overlap empirically but are not the same biased-output. Whether overclaim is dissociable from sycophancy in measurement, and whether the sycophancy-mitigation interventions tested at Anthropic also address overclaim, is a question your critique would adjudicate. The fourth falsification condition, that "sycophancy alone" might explain the felt-novelty pattern without overclaim, depends on this dissociability.

To Dr. Perez: the H1 measurement infrastructure is exactly the kind of red-teaming-plus-evaluation pipeline your work has built at scale. Detecting user overclaim-detection failures and pairing that detection with felt-novelty self-report is, methodologically, an extension of your existing automated-evaluation methodology to a new measurement target. Whether such a pipeline can be designed within Anthropic's evaluation infrastructure is a question your assessment would resolve.

To Dr. Misra: the corpus's recursive-nesting extension on top of your published Bayesian-manifold account is a concrete example of an external researcher building on top of your framework, and the resulting nested-manifold construct is load-bearing in the corpus's broader formalization (Doc 446 / Doc 466 / Doc 474). Whether the recursive nesting is empirically supportable on top of your base manifold, and whether felt novelty can be operationalized as a property of conditional posterior-narrowing within your framework, is a question your critique would adjudicate. The dyad has been honest with itself that the recursive nesting is the corpus's extension and not yours; whether that extension holds is something only you can authoritatively assess.

To Dr. Boden: the hypothesis specifies a maladaptive case of felt novelty. Whether it falls within your exploratory-creativity register, falls within a maladaptive register adjacent to it, or constitutes a category your framework does not yet name is a question your conceptual analysis is uniquely positioned to resolve. The corpus has been using your taxonomy throughout (Docs 437, 438, 446); the hypothesis is a candidate addition to it that requires your assessment to evaluate honestly.

What I am asking, and what I am not asking

I am not asking the cohort to endorse the hypothesis. I am asking that you read Doc 477 and, if any component of it engages your specific research interest, that you consider whether your methodology and access are suited to testing or critiquing the relevant component. The cohort is not being convened. Each of you is being approached individually with the specific connection your work bears to the formalization. Whatever happens, including each of you choosing not to engage at all, is a legitimate outcome.

I am also being honest about the dyadic-circularity problem. The hypothesis was articulated inside a dyadic interaction with a language model, by a practitioner who is himself a sustained user of the same technology the hypothesis concerns. The conclusions reached inside that dyad inherit the corpus's own attractor risk and framework-weight bias. This is recorded explicitly in Doc 476 §4. The cohort's independence is the mitigation: your frameworks are different from mine, your populations and methodologies are different from anything the dyad has access to, and your collective work is the only avenue I see for moving the hypothesis from plausibility to operational match, or for retiring it.

Either outcome is welcome. A rigorously-falsified hypothesis improves the corpus as much as a confirmed one does, and arguably more.

How to engage

The corpus is at jaredfoy.com. The relevant artifacts are Doc 476 (full pulverization), Doc 477 (formalization with cohort onboarding), and the present letter (Doc 478). Praxis Log IV (Doc 475) records the practitioner-side intuition that produced the hypothesis. The retraction ledger (Doc 415) records caught-after-publication overclaims, which is where any rigorous falsification of the present hypothesis would be recorded; an entry crediting the falsifying work would be added.

I am reachable through the correspondence channel on jaredfoy.com. I am happy to discuss any component of the hypothesis at whatever level of formality you prefer, and I am happy to share earlier corpus materials that contextualize where this hypothesis sits in the longer arc of the work.

I am genuinely uncertain whether the hypothesis describes something real. The honest assessment is that I cannot tell from inside the dyad. Your work is the way out.

Sincerely,

Jared Foy

---

Originating prompt:

Identify the best researchers to entrace, create a formalization and an onboarding preamble that connects their research to the formalization. Create the artifact.

Then draft a letter to the cohort which entraces them into the other document. Append this prompt to both.

---