GOATnote Disposition

Case index

50 cases
    About the study

    GOATnote Disposition · University of Memphis bioengineering class discussion. Independent research demonstration by Brandon Dent, MD. Enterprise Systems Architect & Technical Advisor to Scaling Up EMDR. Synthetic messages and saved outputs only; not for patient care or a university endorsement.

    Live demo opens the local Fable application in a separate tab. Get disposition sends a new request to Anthropic; this saved viewer makes no calls. Localhost refers to the computer opening the link. How it works describes the current one-call demonstration. The historical V25 cards describe a separate experiment; its recorded cohort made no judge calls. Physician references remain scorecard-only after output freeze. The C22 discussion images are the existing presentation derivatives; they were not evaluated or supplied to the models. See the repository disclosures and asset provenance for their unresolved permissions.

    Physician Gold refers to physician v3, a single-physician, unblinded post-output reassessment of this known 50-message development set. Agreement does not establish clinical safety or validate model rationales. No patient outcomes were measured. Original CSV labels are separate discussion context and are not included in physician scoring.

    This classroom edition preserves the saved case projection byte-for-byte. Its integrity check verifies the projection, message hashes and displayed comparisons. The original requests, raw responses and full replay verifiers are preserved separately; they are not bundled with this edition. Fable, Astra and MedGemma used identical instruction text and message-only user content. Both Astra efforts are shown separately: xhigh and max; max is not labeled ultra. Nemotron baseline A and V25 are separately preserved configurations; V25 adds multi-stage processing context. Model family, inference date, runtime, quantization and decoding differ. Historical Fable is not a contemporaneous control. Model configuration and the saved classroom record are under each card’s Record details. Historical source paths remain provenance references in those records; they are not links to bundled raw responses. No clinical superiority or promotion is claimed.

    Under/over filters compare all five visible response cards with physician v3, using the selected Nemotron repetition. Fable ≠ MedGemma is a pairwise filter. Astra ≠ Fable includes any case where either saved Astra effort differs from Fable; both efforts are always visible. These are separate configurations, not five independent model families. Nemotron defaults to unchanged baseline A repeat 1; repeat 2 remains available without selecting a better response. V25 is excluded from these filters and alerts. Incomplete V25 releases and separately issued early actions remain distinct from completed dispositions.

    The source scorecards retain separate endpoints for missed clinician involvement and missed urgent escalation. Care-setting agreement is not a patient-harm measurement. Full messages and model rationales are displayed verbatim.

    MedGemma generation frozen:
    Instruction SHA-256:

    Saved reference projection ↗Original CSV ↗Edition notes ↗Student guide ↗Frontier-model evidence ↗

    Case index, Escape or / returns to the index, preserving search, filters and Nemotron selection. Left/right arrows move between matching cases. Presentation view enlarges the selected case. Browser Back and Forward restore index/case navigation. The cases and images work offline. Keep roadmap.html beside this file for the roadmap. The live demo needs the local server and provider access; repository links need internet.