Research log / RL-003

RL-003 Whole draft in context for code follow-ups

Evidence grade AC narrow

Field Value
Date 2026-09-10
Harness version 2026-09-10-artifact-context-1
Models qwen2.5:1.5b-instruct
Evidence grade A (narrow: 2 scenarios × 3 repeats) + C
Pre-registered no (backfilled 2026-09-13)
Source record ARTIFACT-FOLLOWUPS.md, artifact-followup-evaluation.json, artifact-code-evaluation.json

Observation

User requested an HTML dogs page "using information from the internet". The app produced an unverified image URL, then answered explain and continue by generating different pages. Root causes found in code: the code prompt accepted a draft argument without including it; sticky code mode sent explanations to the whole-file writer; code routing bypassed web retrieval.

Hypotheses (reconstructed)

Intervention

One whole draft with request, revision and completion flag; explain cannot update the draft; bare continue on complete output asks what to change; research-before-generation with bounded evidence; invented remote img src stripped; 12,000-char draft cap. Added the agent_turns journal and harness versioning — the basis for all later grade-D evidence.

Method

npm run sim -- artifact-followup 3 and npm run sim -- code 2, disposable databases, synthetic search (no live provider).

Results

Suite Passed Turns
artifact-followup (2 scenarios × 3) 6/6 runs 15
code subset (3 scenarios × 2) 2/3 scenarios

The game-editing scenario stayed red: in one repeat the full draft exceeded the protected prompt budget; in another the model offered prose instead of an edit. Unit tests 379/379.

Journal under this version (real use, n = 11 turns): 3 failed.

Threats to validity

Scripted phrasing; synthetic search; one small model; completion is a delimiter check, not functional verification.

Conclusion

H1 supported for the scripted dog-page interaction (A, narrow); not supported for game edits under a 4k window. H2 supported in harness tests (C).

Follow-ups

Reply-room accounting for large drafts (→ RL-004, RL-010).

Reproduce

npm run sim -- artifact-followup 3
npm run sim -- code 2