COLEMAN 4.14 — Reproduce the CI decisions that are hardest to trust¶

What happens when failures disappear for a long interval, a test changes its name, feedback arrives late, or a suite grows to 50,000 tests? COLEMAN 4.14 adds an owned deterministic CI stress laboratory so those questions can be examined with explicit truth, availability and accounting.

A small design with inspectable coverage¶

Eleven declared factors cover scale, failure regimes, costs, churn, drift, flakiness, feedback, budgets, context, variants and modeled worker counts. Thirty-two deterministic cases cover every two-factor level interaction. Each row has a private coverage witness: inclusion-minimal, without a claim of globally optimal size. Seven named adverse sequences add temporal mechanisms. Nothing automatically launches the full matrix or a Cartesian explosion.

Reuse the platform you already have¶

Generate one case into canonical truth JSONL and native Parquet, verify its bytes by regeneration, then use the existing API, real Random/UCB policies, result sinks and checkpoints. Immutable complete-last exports support identical concurrent retry. Native recovery handles interruption between result flush and checkpoint commit, without inventing another experiment framework.

Point-in-time feedback pages keep the established history bounds. Current truth stays separate from decision-visible context, unknown feedback stays null and late corrections preserve earlier knowledge. Unsupported native feedback is explicitly incompatible instead of receiving a misleading policy score.

Evidence with limits you can see¶

The qualification includes a fresh installed wheel outside the checkout, an actual 50,000-test native history, exact tiny references, independent interaction enumeration, mutation/leakage guards and real process-interruption recovery. Synthetic cases remain constructed, dependent evidence. Four-worker schedules are modeled post-hoc replay, not observed parallel CI speedups. Seeds are not independent projects, and no algorithm superiority or corpus admission is claimed.

Start with the CI stress laboratory guide. Existing historical, controlled-source and HCS workflows retain their contracts. The next #261 gate integrates source admission and production-reality evidence; the benchmark program remains open until its complete scientific criteria pass.