docs/release_notes.md

Release Notes

scientific-agent-protocol-traces-v2.1.0

• Replaced mixed native/v1 records with typed, normalized Parquet tables.
• Added stable problem, trajectory, event, agent, hypothesis, intervention, and label IDs.
• Disambiguated SciBench source-ID collisions by problem-text hash.
• Collapsed exact MaScQA duplicates and retained the decision in the crosswalk.
• Separated benchmark slice, actor model, evaluator model, compute site, backend, and release status.
• Added source-run versus publication-target coverage fields and release gates.
• Imported frozen Aurora and CRUX classic runs with per-file manifest checksums.
• Added generated benchmark accuracy and LAB-Bench filter-audit documentation.
• Bound all 40 classic comparison cells to frozen source manifests.
• Added provenance-checked token and request-submission measurements for 39,918 Direct Baseline, Single-Agent Iterative, and Broadcast Deliberation trajectories across Gemma 4 31B and GPT-OSS 120B, plus direct per-problem elapsed time for all 53,224 matched trajectories. Planner–Executor–Reviewer token and request fields remain null because cumulative main-model and resetting memory-summary counters are not exactly separable after the fact.
• Kept matched four-protocol token oracles null whenever a successful protocol lacks token measurements, rather than treating unmeasured PER as absent.
• Normalized released cost-manifest problem and row identities to public trajectory keys while preserving their original source identifiers.
• For 11 trajectories in eight documented OmniMath2 partial reruns, aligned labels with the selected complete rerun: four final-answer correctness values and 11 first-pass correctness values changed. Final-answer labels are corroborated by the complete result and trace; first-pass labels come from the selected complete metric file. The released source manifests preserve both file identities, checksums, and the earlier values.
• Suppressed 9,500 inherited cost rows from supplemental GPT-OSS and Gemma-3 slices that lack validated per-trajectory accounting and cost manifests; their released values remain null.
• Corrected SciBench broadcast row-level backend and source-commit provenance from its frozen base/recovery source manifest.
• Rebuilt composite licensing, Croissant metadata, checksums, and manifest.

Source HF revision: dada613c7589209a862d91c7b3f0129c99d532e0.