The Dataset Viewer has been disabled on this dataset.

chamber-bench-reference-trajectories

The reference and probe evidence for CHAMBER-Bench v1.0, released with the fsafaei/concerto repository: the episodes that pin down who the partners are and the exploratory runs that sized the campaigns.

What is in this dataset

  • partner-fingerprints/ — for every partner-set member, a verifiable probe bundle (the preregistered probe episodes it was fingerprinted on) plus a per-set fingerprints.json summary; also the cocarry_partners-v2-dropped-candidates/ record documenting the four learned candidates rejected by the committed competence floor. A fingerprint is a deterministic behavioural summary of a partner measured on fixed probe episodes — it identifies a partner without revealing its parameters.
  • power-pilots/ — the power-pilot bundles: the run_purpose: power exploratory runs that sized the campaign sample budgets before preregistration. They are committed evidence, deliberately excluded from every leaderboard manifest — pilots inform design, they never contribute to a reported number.
  • manifest.json, SHA256SUMS.txt — file list and digests; check a download with sha256sum -c SHA256SUMS.txt.

How to verify

Every probe bundle and power-pilot bundle is an ADR-028 result bundle; from a repository checkout:

uv run chamber-eval verify <bundle-directory>

The evaluation protocol these bundles feed, with worked examples, is documented in docs/explanation/evaluation-protocol.md. The admission-report archives that decide which tasks count are released alongside the rows they support, in chamber-bench-leaderboard-bundles.

Provenance

Generated in simulation by the repository's tooling under tag-locked preregistrations, July 2026. Probe episodes for private partner members were produced with the maintainer-held seed; the episodes reveal behaviour, not parameters. No personal data of any kind.

Intended use

Auditing partner identity and competence floors; regression-testing partner behaviour against committed fingerprints; auditing how the campaign sample sizes were chosen. Training on probe episodes and then submitting to the leaderboard defeats the unseen-partner condition and is out of scope.

Limitations

Probe coverage follows the preregistered probe grid; power-pilot coverage exists for the runs the v1.0 campaigns actually used to size their budgets.

Licence and citation

Apache-2.0, same as the repository. Cite via the repository's CITATION.cff (DOI 10.5281/zenodo.20128468). Maintainer contact and deprecation policy: MAINTENANCE.md in the repository.

Downloads last month
69