Publish exhaustive audited DSPy and calibration benchmark results

[CompleteDotTech]

Scope

Publishes the completed offline audit of the captured DSPy and Jev runs from #28 and #29. The earlier publication workflow did not run because the hosted #29 analysis failed on a numerical near-tie. This branch reconstructs the audit from the original captures and publishes the result with explicit provenance; it makes zero new inference or paid API calls.

Evidence and findings

The committed experiments/dspy-calibration-20260918/results/ directory contains 519 files (92.1 MB), including raw call journals, failed pilots, proposal traces, fitted calibrators, complete metrics, and hash manifests. Original 36 Actions artifact ZIPs were independently checked against GitHub SHA-256 digests before extraction. The package verifies eight formulations, 40 searches, 792 primary metric rows, seven multi-call workflows (378 forecast and 126 policy rows), 432 repeatability rows, and 528 graph replays. Original failed workflow status is retained in offline-audit-status.json; no old job is represented as green.

The results summary reports improvements, regressions, and null effects. It separates calibration NLL, Brier, ECE and accuracy; excludes 12 uncertain entity fixtures only from binary-label metrics while retaining their raw diagnostics; and documents one scalar-temperature floating-point near-tie without counting it as a semantic gain. These are retrospective exploratory results on previously evaluated public datasets, not an untouched external benchmark.

Validation

• 515 source and 518 published files pass hash-bound verification.
• 21 focused tests, optimizer inventory verification, actionlint, and whitespace checks pass.
• All 31 workflows are manual-only; the publication workflow verifies committed results and does not push.
• The 34 compressed journals containing 27,842 records and 477 uncompressed result files were scanned for common credential/header markers with zero matches. This heuristic is not a guarantee about arbitrary natural-language content.
• Largest committed file is 9.2 MB; no Git LFS dependency.

The full evidence remains in Git after Actions artifacts expire.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-28T19:34:59.357669Z	389e200	Draft marked ready
🔒 Security Review	✅ Completed 2026-09-28T19:36:03.290547Z	389e200	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.