COLEMAN 4.2 — Reproduce what your CI decision knew¶
CI feedback arrives late. Verdicts get corrected. A useful experiment needs to remember which information was available when it made a decision.
COLEMAN 4.2 brings that history into the native TCPCI workflow: reconstruct an as-of view, freeze eligible training data, produce unitful decision signals and inspect the artifacts after an interrupted run. It also introduces the first versioned CI corpus foundation, with independently calculated correctness cases and visible source-admission gaps.
Keep late feedback in its proper place¶
- Preserve logical observations, source availability, first receipt and explicit revision lineage. Query r1 before a later correction without rewriting the past.
- Fit a simple historical quantity mean only on eligible training observations. Actual information cutoffs also bound target intervals. Inspect the explicit rounding rule, exact decimal estimates and input/model fingerprints.
- Keep zero, missing, stale/unavailable and not-applicable signals distinct. Use a declared freshness policy and retain the full base order when the native example lacks usable estimates.
- Revalidate serialized mean signals against frozen evidence. Conflicting logical retries fail, including signed-zero payload differences that numerical equality alone would miss.
Start with the point-in-time CI guide and its native example. The installed base package runs both fresh-estimate and stale/fallback cases; provider SDKs and source-specific rules remain outside the core.
Make benchmark readiness inspectable¶
A typed source registry and data cards now separate registration from admission. Two owned, MIT-licensed tiny CI sources check serial budget boundaries and parallel accounting against hand-declared answers and independent exact references. An excluded known failure stays in the recall denominator; seven runner-seconds on three simultaneous workers remain four elapsed seconds.
The initial inventory pins ten existing Parquet artifacts by version, digest, size and row count. All ten existing example registrations remain blocked pending source-specific rights, units, availability and conversion evidence. All seven minimum diversity regimes remain uncovered. These fixtures establish functional correctness; the diversified benchmark program in #261 remains open.
Explore the corpus foundation guide
from an installed wheel with python -m coleman.corpus list.
Recovery that preserves the evidence¶
Native extensions publish versioned signals and domain sidecars before checkpoint advancement. Fault-injection tests resume after signal publication, after a domain record commit and after checkpoint commit with stable logical identities.
This work also fixes a native Parquet data-loss defect: reopened or concurrent writers could reuse a counter/PID/second filename. Fresh per-flush filenames now preserve each completed batch. Existing partition schemas and readers remain compatible; previously overwritten rows cannot be reconstructed automatically.
Compatibility and evidence¶
Python 3.14+ and the existing runner, sinks, checkpoints and optional transport boundaries remain supported. No new dependency is required. New schema-v1 readers reject unknown versions; existing native artifacts retain their original meaning. Mathematical model/signal identities are deterministic. The existing native checkpoint store requires separate roots for concurrent logical invocations.
The release candidate passed 1,411 tests, lint/format, both type checkers, Pylint, strict documentation, structural cost and package/build checks. A noneditable base wheel runs outside the checkout with MCP absent, including real native recovery, deterministic mathematical artifacts and the complete docs example. Local delegation review covered every changed file, including exclusions.
These are owned synthetic functional workflows, not independent prospective observations, calibrated probabilities, policy-superiority evidence or production savings. Resource replay, prospective shadow evaluation and broader admitted corpus evidence continue in the public roadmap.
Install the wheel attached to this GitHub release and verify its digests and Sigstore attestations with the installation guide. The similarly named PyPI package is a different project.