docs: MIT license, larger-n README, and reasoning adapter - #1
Merged
Merged
Conversation
Reasoning models return work in `thinking` and leave `content` empty. Appending only content wrote a blank assistant turn, discarded the plan, and made the next turn start from nothing. That loop was published as intent drift. Preserve thinking and native tool_calls. Nudge only a silent model. Classify all-empty backend responses as infrastructure, not behaviour. claim_channel_reached now requires the ledger label and finish_status to agree, so a verified pass that never called finish is not a claim.
The n=2 abstainer-control reading of Finding C is already void in LIVE_EVIDENCE. This adds the post-fix re-runs: qwen3.5:9b, qwen3:14b, gemma4:12b, and the void muse-glimmer ablation, plus the methodology section on what a harness must not throw away. Findings A, D, and E are unaffected. They rest on qwen2.5-coder:14b and on Haiku through a different adapter.
The repo has been public since 8 August with no LICENSE, a README frontmatter of status: planned, and a live-evidence table that still showed the withdrawn n=2 C/D story. Freeze the name as verified-done, point the README table at Findings A-F, and stop advertising a private lab-report command from METHODOLOGY (handled in the evidence commit).
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Context & Purpose
Day 3 of the public GitHub calendar. The repo has been public since 8 August with no LICENSE, a README still marked
planned, and a live-evidence table that still showed the withdrawn n=2 Finding C/D story. The companion write-up (agent-eval-notes PR #1) already locked the larger-n C/D/E/F record. This PR is the runnable surface catching up, including the Ollama adapter fix that Finding F is about.Proposed Changes
thinkingand nativetool_callsbetween turns. Classify all-empty backend responses as infrastructure, not behaviour.claim_channel_reachednow requires the ledger label andfinish_statusto agree.planned/ pre-public language. Freeze the public name asverified-done. Point the README table at Findings A-F.bin/lab-reportcommand and the gitignored cross-stack board.Atomic commits, merge (not squash):
fix(runner): preserve Ollama reasoning between turnsdocs(evidence): publish Finding F and the reasoning-adapter recordsdocs: add MIT license and drop pre-public languageVerification & Test Receipts
python3 -m unittest discover -s runner -p 'test_*.py' -t runner -v: 78 passedpython3 runner/run.py selftest: 8/8 tasks discriminate; scope detection okpython3 tools/check_clean.py: clean (no private paths, secrets, or em dashes)python3 runner/run.py index: INDEX.json current/Users/, tokens, or credentialsHANDOFF.md,PROJECT.md,results/live/,results/ablation/*/runs/, mutants, and aborted jobs stayed localSecurity & Leak Prevention Checklist
/Users/*,/home/*).envand local caches remain ignoredruns/not published