COLEMAN 4.3 — Account for CI work. Audit every recommendation.¶
COLEMAN 4.3 makes CI test-prioritization experiments easier to audit. Account for required runner work, replay cost-aware choices and evaluate frozen recommendations alongside an unchanged CI suite—with explicit units, recoverable artifacts and clear evidence requirements.
The product remains test-case prioritization and selection for continuous integration. Reusable typed contracts make the audit easier to inspect and reuse.
Know where the budget went¶
New exact resource ledgers distinguish estimates, reservations, actual occupied runner milliseconds, unused release, dispatch and shared setup. Actual overruns remain visible and stop new analytical reservations; they are never clipped into feasibility. Unknown costs stay unknown, zero stays zero and negative comparison savings remain valid findings.
Historical replay compares frozen policies under equal budgets through the native COLEMAN runner. Reports separate modeled subset work from full-suite execution, shared costs and unavailable latency. A shorter serial duration sum is not presented as observed wall-clock saving.
Observe recommendations without changing required tests¶
A prospective shadow protocol freezes candidate/model/features, source identity, temporal windows, trial registry, thresholds and evidence requirements before feedback. Typed source events preserve upstream chronology and first validated application receipt. Operational health uses an explicit local monotonic clock; research conclusions remain separate from connectivity.
Native extension and monitor hooks seal recommendations before outcome evaluation, persist the original source contribution, and publish audit records before native checkpoint advancement. Duplicates are idempotent. Conflicts, stale sources, unsupported corrections and resource/quality breaches remain explicit. Atomic journal frames recover committed events even if a checkpoint-reference write was interrupted.
Start with a small reproducible CI workflow¶
Use Python 3.14 and the verified release wheel attached below. Copy an example into a separate workspace and run it twice to exercise checkpoint recovery:
Choose examples/prospective-shadow/ci.py
for the complete native shadow workflow, or
examples/resource-replay/ci.py
for historical resource replay. The base package is sufficient; neither example
needs MCP, hosted deployment, credentials or externally triggered CI jobs.
Read the complete shadow guide and resource guide for units, invariants, complete API examples and compatibility limits.
Evidence you can inspect, limits you should keep¶
Recovery checks inject failures around actual source, artifact, result and native checkpoint commits. Golden fixtures, unknown-version rejection, signed savings, unit checks, chronology, bounded retention, concurrent publication, metamorphic independent-arrival permutations and installed-package native examples accompany the implementation. The integrated implementation passed 1,675 tests with the repository's branch-coverage gate, plus typing, lint, structural, documentation, package/install and security checks before publication.
The shipped shadow source is an owned synthetic functional fixture. Two builds on one day do not satisfy its frozen seven-day requirement: the result is inconclusive, with zero collected prospective builds. Repeated seeds and job retries create no independent observations. The guide includes a forward-run source/authorization checklist and an empty report template; it does not fabricate a live study or establish production savings, causal value or algorithm superiority.
Compatibility and what follows¶
This is an additive 4.x release. Existing native scheduling outputs, v3.2 artifacts, domain records, runner/checkpoints and the application API retain their contracts. New resource and shadow JSON artifacts require explicit schema version 1 and reject unknown versions. Legacy rows cannot be relabeled as prospective evidence; there is no implicit migration. Resource v1 does not support refunds, top-ups or retrospective occupancy corrections.
The diverse-corpus program in #261 remains open. Source admission, secure materialization, broader regimes, baselines and frozen cross-corpus evidence will follow in reviewed vertical slices. No broad external-validity claim is made by this release.