Research log / RL-011

RL-011 Verified website builds in chat

Evidence grade BC

Field Value
Date 2026-09-12
Harness version 2026-09-12-verified-code-13
Models qwen2.5-coder:7b, 8,192-token Ollama context
Evidence grade B (one live workflow) + C (browser fixtures, scripted chat)
Pre-registered no (backfilled 2026-09-13)
Source record VERIFIED-CHAT.md, docs/verified-chat/*

Observation

RL-010's verifier and repair loop ran only from evaluation scripts; chat users received unverified drafts.

Hypotheses (reconstructed)

Intervention

WebsitePlan.ts selects from a fixed check catalogue; VerifiedWebsiteBuild.ts coordinates draft, repair and rollback; WebsiteVerifier.ts runs a fresh isolated renderer per candidate; regression on follow-up restores the previous draft; repair prompts include the actual card DOM.

Method

npm run check:website:live: real renderer, preload, host, verifier; disposable state; asserts final checks pass.

Results

One live run: 10/10 supported checks passed (responsive Tailwind dog site, three named breeds, live search, expandable FAQ, local email confirmation). The harness repaired the search label and full displayed-name matching. Tests 644.

Threats to validity

n = 1; creation only, not live follow-ups; the check catalogue bounds what "passing" means.

Conclusion

H1 supported at grade B for one creation workflow.

Follow-ups

Follow-up edits and non-card apps (→ RL-012).

Reproduce

npm run check:website
npm run check:website:live