Research log / RL-009

RL-009 Continuation guards and stall stop

Evidence grade DBC

Field Value
Date 2026-09-12
Harness version 2026-09-12-live-editing-82026-09-12-code-progress-11
Models qwen3:4b-thinking (observed); local model for the live check
Evidence grade D (journal) + B (one live reproduction) + C (9 regression cases)
Pre-registered no (backfilled 2026-09-13)
Source record CODE-PROGRESS.md

Observation

Journal, 2026-09-12, qwen3:4b-thinking:

Time Message Outcome
10:20 create a website about dogs, multiple pages, tailwind incomplete draft: <!DOCTYPE html><html lang="en">
10:23 continue incomplete draft
10:25 keep going error: spent its reply budget on reasoning
10:51 continue error: spent its reply budget on reasoning
11:09 continue incomplete draft shorter than before
11:14 continue error: spent its reply budget on reasoning

4 of 6 continuation attempts failed; one "success" shrank the draft.

Hypotheses (reconstructed)

Intervention

Multi-page briefs get the single-file limitation before generation; continuation cannot shrink or repeat; incomplete edits cannot replace complete drafts; two-stall stop per revision/model; +1024 output tokens for thinking: always, bounded by available context.

Method

Nine regression cases replaying the reported failures. One disposable live run.

Results

Threats to validity

n = 1 live; length and delimiter checks are not semantic correctness; the stall counter does not survive restart.

Conclusion

H1, H2 supported (C). H3 not supported in real use: budget exhaustion recurred after the change (D).

Follow-ups

Retry once with low reasoning effort on continuations; classify the failure explicitly; recommend a non-reasoning coder for drafts.

Reproduce

npm test