COLEMAN 4.15 — Know what your benchmark can support¶

A converted dataset is a starting point. Which units does it report? Which jobs share a build? Was a label available when the decision was made? COLEMAN 4.15 adds an auditable source-admission workflow that makes those questions explicit before you compare CI prioritization policies.

The published wheel was qualified outside checkout: all 306 SDK resources match the reviewed source and tag, and 61 source bindings reproduce the same versioned report. Both distributions have verified GitHub workflow signatures. Actual Karaf agents reach checkpoint step 8 and resume published 4.14 results without changing their protected bytes.

One inventory, inspectable evidence¶

Reuse the existing corpus registry, data cards and source completions. Audit retained observations, execution groups, duplicates, outcome missingness and duration distributions without changing the source's units or old results. Generate deterministic versioned JSON and a human report with 35 production-reality axes, including stages, variants, availability, retries and partial outcomes. Every fact keeps its evidence status and limitation.

Preserve the original producer¶

Controlled functional qualification verifies the original raw packets with an explicit adapter bound to the original installed COLEMAN wheel. Package/version, SDK bytes, source context and normalized observations must agree. Readable proof metadata cannot replace that verification. The domain remains independent of CLI, MCP and capture transports; offline auditing does not execute the subject tests.

A result can say “insufficient evidence”¶

Negative and inconclusive functional observations remain valid. Constructed C examples stay separate from real faults, and all 39 declared stress cases are required before covering the broad synthetic mechanism regime. Neither category supplies missing historical CI clocks or independent projects. Uncovered diversity cells remain uncovered; no policy advantage or production saving is implied.

Resume without rewriting evidence¶

Existing source receipts, native results and learning checkpoints retain their contracts. New schema versions reject missing, coerced and unknown versions. Immutable complete-last publication supports interruption, concurrent identical attempts and idempotent retry. Reading a report is inspection; explicit validation rechecks all source bindings before reuse.

Start with the source-admission guide for complete CI examples, original-package bindings, units, compatibility and recovery. The broader #261 benchmark program still requires statistical references, comparable baselines, prospective feasibility and a frozen confirmatory report.