COLEMAN 4.18 — Trace the data. Bound the claim.¶

Bring original industrial test results into reproducible Parquet, and compare bounded CI metrics without assuming normal differences. This release strengthens the path from source records to an inspectable decision: explicit quantities, preserved abnormal outcomes, conservative uncertainty and reliable recovery.

Explore original industrial CI results¶

The Westermo adapter fetches a pinned, licensed original release only when requested. It verifies the archive and decoded file, then writes separate typed observation and session tables. Offline retries preserve completed bytes; interruption cannot publish a false completion receipt.

The installed-package qualification checks all 1,063,915 original rows and exports 1,064 observations from sixteen frozen sessions. Pass, fail, invalid and unmappable outcomes remain distinct. An independent reference checks every selected field, and different Polars thread settings produce identical export bytes. No third-party dataset is bundled in the package.

Session duration retains its documented seconds. Per-test duration retains its original integer and an explicitly unverified unit. Shifted source dates remain naive labels. These distinctions help you inspect the data without accidentally claiming physical runner savings or verified historical availability.

Compare bounded metrics with conservative uncertainty¶

Use the versioned BoundedMeanProtocol for paired means with known metric limits and justified independent inference units. The complete native CI example adds --bounded to the existing Random/UCB runner, repetition records, protected access audit and checkpoints.

The method provides full-family Holm adjustment, simultaneous confidence bounds and explicit practical equivalence. Rational calculations and outward rounding keep numerical approximations conservative. Missing pairs remain incomplete evidence; observed zero and valid signed effects retain their meaning.

Keep existing artifacts and workflows¶

This is an additive release. Existing Student-t contracts, golden artifacts and default native examples retain their behavior. No old result acquires inferred bounds, new units or a different statistical interpretation. Unknown versions reject; preserve recorded producer/runtime configurations when reading old source exports. Older SDKs reject the new inference method explicitly.

Build or install the 4.18 package before using the new adapter and protocol. The guides include complete installed-package commands, source attribution, recovery, version compatibility and evidence limits. The base package remains independent of optional CLI/MCP transports.

Evidence that stays honest¶

The native example retains sixteen build/seed observations from one related family and reports insufficient evidence. Repeated seeds do not create independent projects. Protected reuse is recorded and cannot create a fresh confirmatory claim.

The source conversion is analytical access, not historical benchmark admission. Four mandatory historical/HCS diversity regimes still lack qualified evidence. The final protected cross-corpus program and its source qualification continue under #313 and #261. This release makes no algorithm-superiority or universal-generalization claim.