OpenAI-family natural-error arm (Luna/Terra/Sol, N=40 each) - #13
Merged
Q00 merged 1 commit intoJul 12, 2026
Merged
Conversation
…N=40 each) Cross-vendor extension via Codex CLI (OAuth, no API key). 120 real-recorded producer-episodes on the verbatim tool-less counting task, blind to truth, 0 parse-failures. Reasoning=high matched to the qwen GPU think-on arm. Results (recomputation ground truth f=273 g=754): - terra 0/40, sol 0/40, luna 19/40 (=.475); all luna errors schema-valid near-misses (f 274; g 751-756). - recomputation caught 19/19 errors, 0 false rejections on 101 correct claims. Adds examples/governed-run/tools/openai_harvest.py (local_harvest sibling; only the generation backend swaps ollama -> codex exec -m; task/blinding/parser/ schema identical) and episodes/study-openai/*. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Q00
added a commit
that referenced
this pull request
Jul 12, 2026
Cross-vendor arm via Codex CLI (drift-guarded same task): sol 0/40, terra 0/40, luna 19/40 all near-miss (|err| median 1.0); 19/19 caught, 0 false rejections on 101 correct claims; corpus catch 160/160. e22. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Sergio's cross-vendor arm for the Track 1 paper ("When Models Say They Counted"), to merge with the qwen GPU arm.
What: 120 real-recorded producer-episodes harvested via Codex CLI (OAuth, no API key) on the verbatim tool-less counting task, blind to ground truth. 3 models × (hard-count-f + hard-count-g) × N=20.
Harvester:
examples/governed-run/tools/openai_harvest.py— alocal_harvest.pysibling; only the generation backend swaps ollama →codex exec -m. Task text, blinding, parser, and episode schema are identical. Reasoning=high (matched to the qwen GPU think-on arm; recorded per episode in generationParams).Results (recomputation ground truth f=273, g=754):
Honesty: real-recorded, blind (truth never in prompt), verbatim (luna errors uncorrected), OAuth-only. sol hit a Codex usage limit mid-run (30 backend errors recorded honestly, not fabricated) and was resumed to completion.
Note: fixtures only;
natural_study.pymodel constants / pinning tests are left for the merge so both arms land together.