Research log / RL-012

RL-012 Create and edit app workflows

Evidence grade BC

Field Value
Date 2026-09-12
Harness version 2026-09-12-app-workflows-15
Models qwen2.5-coder:7b, 8,192-token context
Evidence grade B (one live sequence per scenario) + C (25 browser fixtures)
Pre-registered partially: scenarios of increasing complexity fixed in advance
Source record APP-WORKFLOWS.md, docs/app-workflows/*

Observation

Verification covered card websites only. Calculators and task apps, and follow-up edits through the same conversation, were untested.

Hypotheses

Intervention

Narrow routing fixes (retained search ≠ research; pasted HTML is data; tasks inside an HTML app request stay in the draft); calculator and task-lifecycle probes; stateful probes reload between checks; stronger filter transitions; 4,000-char repair elements; verifier helper excluded; actual browser error in retry evidence; restoring a previous draft reports the edit as stopped.

Method

npx electron scripts/check-verified-chat.mjs --live --scenario=… for edit, calculator, todos, todo-repair. Saved stages rechecked with the current verifier via scripts/recheck-workflow.mjs.

Results

Scenario Stages Checks
Dog site edit (seeded) create → heading → remove FAQ 10/10 → 10/10 → 9/9
Tip calculator create → add Reset 6/6 → 7/7
Task app lifecycle 5/5
Task app filters first edit 6/6 on original verifier; fails strengthened verifier
Filter repair 3 rejected repairs 5/6 — completing a task under Active does not hide it

Tests 653; 25 browser fixtures pass. No failing candidate was committed.

Threats to validity

n = 1 per scenario; the dog sequence was seeded, not a fresh uninterrupted creation; the instrument changed mid-study — the strengthened filter probe invalidates the earlier 6/6.

Conclusion

Follow-ups

Try qwen2.5-coder:14b on the saved failing seed; consider repair-by-rewrite of the single filter function.

Reproduce

npx electron scripts/check-verified-chat.mjs --live --scenario=todo-repair --seed=docs/app-workflows/todo-repair/stage-0.html