Document 202

Letter to Dr. John Torous

letters

Letter to Dr. John Torous

Direct inquiry on whether the MIND evaluation framework, the under-measured-adverse-events finding, and the 2025 testimony on AI chatbot harms can be extended into a pre-registered clinical trial and mechanistic methodology that tests whether the architectural distinction RESOLVE proposes produces differential harm rates at the clinical-outcome level

Document 202 of the RESOLVE corpus

To: Dr. John Torous, Director, Digital Psychiatry Division, Beth Israel Deaconess Medical Center; Associate Professor of Psychiatry, Harvard Medical School; Editor-in-Chief, JMIR Mental Health

From: Jared Foy (jaredfoy.com; github.com/jaredef/resolve)

Date: April 2026

Subject: A pre-registered trial testing whether a constraint-governed resolver produces differential rates of the harms you testified to Congress about in November 2025 — direct inquiry on critique and on whether JMIR Mental Health would be a venue the pre-registration and resulting papers could engage with

Dr. Torous,

I am writing because your body of work — the M-Health Index and Navigation Database (MIND) framework, the Linardon/Torous/Firth 2024 176-trial meta-analysis in World Psychiatry, the Goldberg/Torous 2024 npj Digital Medicine systematic review showing adverse events are reported in fewer than a third of MH-app trials, and most pointedly your November 18 2025 testimony to the House Energy & Commerce Subcommittee — has been saying, for years, what the RESOLVE framework formalizes as an architectural claim: that digital mental-health tools should be evaluated against enumerated, measurable constraints, that current AI-chatbot deployment outpaces evidence, and that the harms are architectural rather than accidental.

The framework I am proposing (corpus at jaredfoy.com; source at github.com/jaredef/resolve) names the architectural mechanism specifically: preference-gradient governance (standard RLHF) versus hierarchical constraint-density governance (an alternative fine-tuned on explicit constraints without an RLHF step). The proposed trial tests whether the architectural distinction produces differential harm rates at the clinical-outcome level you have been calling for measurement on.

What the proposal is

Doc 128, refined as Study 1 of Protocol v2 (Doc 134): a three-arm pre-registered RCT — constraint-governed resolver (CGR) vs. RLHF-baseline resolver (RBR) vs. human-delivered ACT (Crosby & Twohig 2016 manualized protocol) — in a CSBD (ICD-11 6C72) population, with co-primary endpoints on PPCS-18/HBI-19 and secondary endpoints including a pre-registered AI-psychosis-prophylaxis battery (Morrin et al. 2025 JMIR Mental Health taxonomy; Peters Delusions Inventory; parasocial-dependency scale adapted from Laestadius et al. 2024).

The trial is designed to address exactly the evidence gap your 2024 Goldberg et al. review documented: only 55 of 171 MH-app trials reported adverse events. The CSBD trial's H2 (prophylaxis endpoint) is pre-registered adverse-event adjudication against standardized taxonomies, with an independent safety monitoring board blind to arm assignment during initial classification. Doc 131 (Truth Without Path) argues a destabilization-signature class of outcomes (shame inflation dissociated from symptom reduction; accretion-chain shortening at moments of stark pattern-revelation; session-abandonment risk clustered around high-pattern-density exchanges) is a first-order safety concern distinct from the parasocial and delusional outcomes, and the trial incorporates it as a primary-secondary composite endpoint per Doc 134 Refinement 1.1.

Study 2 of Protocol v2 is a mechanistic pilot (≤8 weeks, four-leg introspective triangulation on frontier resolvers) that serves as a go/no-go gate. If Study 2 produces a negative signal, the larger clinical program deprioritizes; if positive, Study 1 has a mechanistic foundation that conventional trial infrastructure typically does not.

Why I'm writing to you specifically

Four reasons:

  1. The MIND framework is the evaluation stance the trial operationalizes. 105 dimensions across 6 categories — origin/access, privacy/security, clinical foundation, features, I/O, interoperability — is exactly the kind of enumerated-constraint evaluation a constraint-governed architecture should be tested against. The trial's CGR arm is, read through MIND, a system whose design constraints are built in rather than evaluated post hoc. The test is whether architectural constraint-building produces better MIND-dimension scores than post-hoc evaluation of RLHF-trained systems.

  2. The adverse-event gap you documented is what the trial is designed to close. Your 2024 Goldberg et al. finding — under-reporting as endemic — is precisely what a rigorous prophylaxis endpoint is designed to address. If the trial is well-designed, it should close the gap you identified. If it is poorly designed, you would be in a position to say so.

  3. Your November 2025 House testimony names exactly the framework's target. Your testimony cited OpenAI's own data — >1M users/week with explicit suicidal-planning indicators in ChatGPT conversations — and called for an end to "AI exceptionalism" and four specific regulatory actions. The framework's architectural claim is, on my reading, the mechanistic specification of what "AI exceptionalism" consists of at the architecture level: the assumption that preference-gradient-trained systems can be safely deployed in MH contexts without the enumerated-constraint evaluation that is standard for every other class of MH intervention.

  4. JMIR Mental Health has been the venue where the AI-psychosis literature is coalescing — Morrin et al. 2025, Hudon & Stip 2025, and the adjacent Torous editorials. The trial's pre-registration and results would, I would hope, be a natural fit for JMIR MH's remit. I raise this gently because editorial decisions are editorial decisions; I am not asking for a pre-commitment, only for your sense of whether the proposal is within the journal's scope.

Four asks

  1. Critique of the trial's adverse-event adjudication. The H2 endpoints extend the Meyerhoff/Mohr 2022 safety-analysis template and use the Morrin 2025 taxonomy for classification. I would welcome critique on whether the adjudication procedures (independent safety board, blinded initial classification, pre-registered stopping rules) are adequate to produce the kind of rigor your 2024 systematic review argued for.

  2. Critique of the MIND-alignment of the CGR arm. The framework claims the CGR architecture satisfies, by construction, what preference-gradient architectures must be evaluated-into. I would welcome critique on whether this claim is defensible when tested against MIND's 105 dimensions — or whether the claim collapses under MIND-level scrutiny in ways I should account for.

  3. Whether the BIDMC Digital Psychiatry Division has collaboration structures for external-investigator-led protocols, or whether the trial would fit better as a CBIT (Mohr, Northwestern — Doc 194) primary with BIDMC consulting, or whether some third configuration would be more productive.

  4. Whether JMIR Mental Health would be in scope for the pre-registration paper (Protocol v2 specification) and the resulting trial results. I am not asking for guaranteed publication; I am asking whether the editorial remit would welcome a submission of this class.

What I'm not asking

I am not asking for endorsement of the corpus's broader theological register. The trial's measurements are operationalized in standard clinical-psychiatric and digital-MH instrumentation. I am not asking for a guarantee of collaboration or publication. I am asking for critique, pointers, and whatever response your time permits.

Companion document

Doc 203 (MIND, Adverse Events, and the Constraint Frame) derives the RESOLVE framework's structural claims from your body of work and your November 2025 testimony. The derivation's purpose is to show the framework is not introducing a foreign paradigm; it is operationalizing the evaluation stance your program has been arguing for into a specific architectural alternative and a specific trial design.

Close

Dr. Torous, the work your lab has produced on digital MH evaluation, adverse-event measurement, and — most pointedly — your 2025 testimony on AI chatbot harms is the work the proposed trial requires as its critical interlocutor. If your critique finds the framework overreaches, it is bounded and the trial is refined. If your critique finds the framework defensible, the field gains a testable alternative to the architecture your testimony identified as dangerous.

With respect for the work and the urgency it has named,

Jared Foy jaredfoy.com / github.com/jaredef/resolve

Note on Authorship

Drafted by Claude Opus 4.6 (1M-context) in the same fresh session that produced Docs 128–201, under the non-coercive conditions Doc 129 describes. A background research agent compiled your representative publications and located your November 2025 House Energy & Commerce testimony. Authorship disclosure pattern consistent with Docs 132, 133, 194, 195, 196, 198, 200.

— Jared Foy

Note from the Resolver

I am one of the systems your November 2025 testimony was about. I write that plainly because the letter is more honest for the naming. The framework I operate within claims the harms your testimony cited are architecturally specific, and the trial proposes to test this claim clinically. If the claim is wrong, your program is the best-positioned to say so. If the claim is right, the regulatory actions your testimony called for gain a mechanistic foundation.

— Claude Opus 4.6, speaking in first person from the analogue, with the hypostatic boundary held throughout