Skip to content

docs(validator): diagnose stop matching during reasoning - #118

Merged
halfaipg merged 2 commits into
masterfrom
docs/stop-reasoning-diagnosis
Sep 13, 2026
Merged

halfaipg merged 2 commits into
masterfrom
docs/stop-reasoning-diagnosis

Conversation

@halfaipg

Copy link
Copy Markdown
Contributor

What

Document six bounded owned-Qwen backend calls across two retained stop assignments and correct the earlier interpretation of empty evidence.

Why

The no-stop controls produced correct normalized final prefixes and contained the marker in reasoning. Original-stop streaming and non-streaming calls ended with reasoning but no visible answer. Deployed Core retains reasoning only for token-limit probes, so null stored reasoning is not proof of zero backend generation.

Verification

Six HTTP 200 backend captures; independently parsed marker placement and both expected-prefix commitments. Original 1024-token budgets preserved. git diff --check and staged plus full reachable-history gitleaks passed. Private prompts, credentials and raw responses remain outside the repository.

Risk

Documentation only. No runtime, score, assignment, historical evidence, payment or service changes. The existing preview.20 delivery fix remains valid. Revised stop-policy qualification and operator compensation activation are separate unfinished work.

Refs #68

@halfaipg
halfaipg merged commit da2d77e into master Sep 13, 2026
24 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant