Approved app revision: https://github.com/Anti-Conjugates/AntiConjugate-Devin/pull/3/commits/8e75242ae8081888475b5ab021adabbce007a10e
Clinical release stays blocked. Public Claude stays disabled. No clinical benchmark or calibration claims.
[Written by Devin](https://app.devin.ai/sessions/c3488f5989f34b69a72dc1168d61087b)
Check a claim opens with three one-click rules-only runs, shows a seven-stage
stepper fed by the streamed trace, names a rejected draft as rejected instead
of 'not enough evidence', and can rerun the same scope with the label withheld
or a fault injected and show both results side by side.
How it works gains a who-sees-what panel with the real payload shapes for
Claude, the local tools and the verifier. Evals replaces the bar chart with
per-scope grids for the scripted strategies and the fault tests. index.html
gets a description, Open Graph tags and a favicon. The API answer text and
one test regex are reworded in plain language; no behaviour change.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Resolve the runtime-skills path from the test file, not an absolute path
- Remove Protocol view; rename navigation; rewrite UI text with the plain-copy skill
- Report Anthropic stop_reason refusal as CLAUDE_REFUSED (fail closed, no rules-only fallback)
- Rewrite planner and draft system prompts in plain language; the claim id is the question under audit
- Evals view renders evals/results.json (source withholding, 48 fault tests, scripted strategies, Claude grid)
- Exploratory ESM-2 antibody sequence map in ADC table from public Thera-SAbDab sequences
- scripts/eval.ts, scripts/embed_antibodies.py, Dockerfile for a Hugging Face Space, docs
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Baseline: research workbench after redesign workflow
