Skip to content

Record DiscoverPhysics partial parity metadata (depends on #9) - #41

Closed
audreina17 wants to merge 6 commits into
harbor-framework:mainfrom
audreina17:discoverphysics-parity-metadata
Closed

audreina17 wants to merge 6 commits into
harbor-framework:mainfrom
audreina17:discoverphysics-parity-metadata

Conversation

@audreina17

Copy link
Copy Markdown
Collaborator

Dependency

Depends on #9. Merge #9 first; keep this PR draft until its base has been reconciled with the merged adapter.

The branch is based on Randl’s PR #9 head, 40bc08e6250211a618daa8db5ae12eca71f7ee51, with one metadata commit, c0af4634edac1503434b8df4817c9a0895e466bc, on top. GitHub cannot select a fork-owned branch as the base of a PR in this repository, so the temporary base is main and the full diff also includes #9. This is a dependency recorded in Git ancestry and PR text, not a GitHub-native stack.

Review only the follow-up: three-file metadata diff. It changes only src/discoverphysics/README.md, adapter_metadata.json, and parity_experiment.json.

The migrated adapter still reports parity as pending and lacks parity_experiment.json, although the historical partial evidence is now uploaded in HF PR #270. This is the parity-metadata follow-up to adapter PR #9.

Record two separate reviewer configurations: Gravity stream-v1 (2/22 tasks) and the complete durable-v3 subset (6/44 tasks: Yukawa, Oscillator, Hubble). Each has three repetitions per side, raw repetition scores, mean ± sample SEM, source pins, and evidence links. The README retains world-level results and Ether's incomplete strict result plus separately qualified salvage. adapter_metadata.json records the corresponding scopes and sampling rates. The README also links the author's 22/22 oracle upload, HF PR #271.

Coulomb and DeepSeek were excluded retrospectively; failures and truncated responses within the selected configurations remain represented. This records partial historical evidence and does not claim full-adapter acceptance or migrated-head equivalence. The reviewer-runtime patch/fork is deferred, so original_parity_repo remains explicitly empty. Standard CLI-agent compatibility and the existing security/fidelity findings remain outstanding. No runtime, grading, generated task, or raw evidence files change, and no model runs were launched.

Validation: adapter structural validator passes 31 checks with zero errors or warnings. Both experiments' arrays were checked against the six per-side result collections; means, sample SEMs, range overlap, counts, sampling rates, and parity-summary parsing agree. git diff --check passes.

Prepared with Codex assistance.

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.

2 participants