COLEMAN 4.7 — From LRTS source data to runnable Parquet¶

Download a public CI benchmark, inspect its converted Parquet files and run a declared simulation through COLEMAN's existing TCPCI runner. This release closes that path for a pinned, bounded Karaf slice of LRTS.

The recipe downloads the verified raw-report archive and metadata, selects eight source builds and writes three useful tables: original method observations, source-build metadata and native-compatible input. A completion manifest binds source revisions, data terms, implementation and file digests. Interrupted work can reuse verified inputs; repeating a completed run resumes its native checkpoint.

Try the complete workflow¶

Use Python 3.14+ and an installed wheel built from this release. Cache and data directories must be outside Git; Docker and an optional MCP SDK are unnecessary.

python -m coleman.transports.lrts_etl \
  --cache /tmp/coleman-lrts-source-cache \
  --output /tmp/coleman-lrts-data \
  --fetch --run

python -m coleman.transports.lrts_etl \
  --cache /tmp/coleman-lrts-source-cache \
  --output /tmp/coleman-lrts-data \
  --run

The first command explicitly permits pinned public GitHub/Zenodo downloads. The second works offline from the verified cache. The JSON summary gives the actual Parquet, completion and native-result paths.

The selected real reports contain 6,876 method executions. All original rows remain in observations.parquet; 6,567 complete reported outcomes/durations enter native input, including 29 reported failed executions. The 309 not-run methods remain visible with explicit exclusions. Unknown outcomes and unavailable durations have independent exclusion reasons; zero remains observed zero. Metadata accounting retains the eight selected rows and 57,429 outside the slice.

Default Random and UCB (c=1) runs use RNFail, seed 42 and a 50% ratio budget. Eight rounds and two policies produce 16 native result rows. Native checkpoints, results, RunSpec and artifact manifests retain their existing behavior. The ETL receipt is linked in the native manifest; no second experiment runner is added.

Reproducible inputs and honest assumptions¶

The complete guide documents the exact eight source builds, component/data terms, units, table schemas, exclusions and installed-package reproduction. Third-party rows and archives stay in external storage. COLEMAN distributes recipes and owned tiny fixtures.

Source and report digests are verified before conversion. A separate bounded ZIP selection recipe supports the real archive directory while retaining the old complete-package v1 limits and ZIP64 rejection. Unsafe paths, aliased entries, unsupported types, contradictory metadata and changed bytes fail closed. New ETL/selection schemas require version 1; unknown/future versions reject. Existing Parquet/results/checkpoints keep their meaning without implicit migration.

Analytical records preserve raw identities, source timestamps and unavailable values. The native projection declares post-hoc serial simulation: scheduled labels order rounds, reported method seconds are added as modeled costs, feedback uses native budget-censored RNFail rewards (zero nondetection reward also for unscheduled cases), and matching labels/occurrences are an assumed correspondence. Zero reward does not assert a passing source verdict. Legacy full-label receipts remain readable but reject new exports/execution; opt-in correction creates new identities without modifying old artifacts. These assumptions do not become observed CI chronology, runner occupancy, parallel latency or verified stable test identities.

Admission remains blocked_source_evidence, and generalization remains insufficient_evidence. This release makes a real bounded source usable for inspection and that declared simulation. It does not establish protected temporal validity, policy superiority, independent faults or prospective savings.

Validation and next steps¶

Owned golden fixtures, old-artifact reads, round trips, corruption checks, deterministic identities/bytes, concurrent publication, interrupted export, missingness/nonfinite cases, source-order guards and native checkpoint recovery are exercised. An installed wheel outside the checkout converts the actual source slice and executes/reopens actual native runs. Acquisition is separately checked against the public source with an initially empty cache.

The diversified benchmark program in #261 remains open. Larger slices, other sources/languages, HCS and synthetic stress, complete baseline matrices and protected confirmation/reporting continue as sequential work. The seven required diversity regimes remain uncovered; the slice is useful data and functional execution evidence, not completed corpus admission.