Public CI Intelligence roadmap¶
COLEMAN 4.0 established versioned domain results, protected temporal evaluation, finite coverage and scenario-payoff decisions, independent solver benchmarks and a bounded application API. These foundations supplement its native TCPCI runner. A completed foundation is not evidence that the remaining roadmap is implemented.
4.1.0 delivers the bounded operational CI pilot in #257. Its claim is auditable accounting and honest evidence, including negative/inconclusive findings. It does not promise production savings from a synthetic example.
4.2.0 groups two reviewed public increments: point-in-time observations and versioned decision signals in #258, and the first corpus foundation in #261. The foundation includes typed manifests/data cards, two tiny known-answer sources, metadata inventory and installed native workflows. It does not complete source admission, corpus diversity or the full benchmark program; all seven required diversity regimes remain uncovered.
4.3.0 groups #259's exact CI resource accounting and cost-aware replay with
260's frozen shadow evaluation. Both reuse the native runner and checkpoints.¶
The shadow fixture demonstrates functional recovery, with zero collected forward builds and insufficient evidence for adoption.
4.4.0 groups the reviewed next two #261 slices: bounded native profiles and a reproducible inventory in #284, and licensed source-package materialization in #285. Typed integrity receipts, external caches, bounded archive validation and installed native recovery workflows reuse the existing runner/artifacts/checkpoints. All ten existing native registrations retain their admission blockers; all seven required diversity regimes remain uncovered.
4.5.0 publishes the reviewed RTPTorrent and LRTS source-assessment adapters, with explicit missing temporal evidence and reproducible admission blockers.
4.6.0 publishes controlled fault observations and pinned JUnit3 source capture, with raw evidence, interruption recovery and negative outcomes. The published wheel reproduces the Java pair and reopens its committed observations outside the checkout. This is source reproduction, not completed benchmark admission.
4.7.0 publishes the real LRTS ETL: explicit download, pinned nested-member conversion, analytical/native Parquet, conserved exclusions and actual native replay/checkpoint continuation. Native execution declares its post-hoc serial assumptions; protected historical admission and cross-corpus confirmation remain separate unfinished gates. Later steps cover admitted real-fault matrices, HCS/variant expansion and synthetic stress, comparable baselines/statistics, frozen confirmatory execution and the final evidence report. Compatible deliveries continue in later 4.x releases. #261 stays open until its admission, diversity, execution and evidence gates pass. Inaccessible or unlicensed sources remain visible blockers. Security fixes remain independently shippable.
Remaining corpus subtasks¶
The public #261 umbrella has native GitHub sub-issues. #301 and #302 shipped in 4.8; #303's larger LRTS stage views shipped in 4.9 through #318.
304's controlled Java matrix shipped in 4.10 through¶
#320.
305's controlled Python producer shipped in 4.11 through¶
#322. #306's constructed C source slice shipped in 4.12 through #324. #307's LibSSH HCS workflow shipped in 4.13 through #326. Seven of the thirteen public subtasks are delivered. #308–#313 remain open; the deterministic interaction-covering stress laboratory in #308 is next. Source conversion does not establish admission.
The published HCS workflow preserves raw pipeline/job correspondence, WTS/VTS Parquet, all 208 sparse cells and isolated multi-budget checkpoints. Its declared cost projection preserves original decimal observations. Installed validation checks 416 native rows independently against original source bytes; six metadata-only labels remain absent evidence. Admission and comparisons remain later gates.
| Stage | Subtasks | Dependency |
|---|---|---|
| RTPTorrent source path — delivered in 4.8 | #301 ZIP64 integrity, #302 real Parquet ETL/replay | #302 requires #301 |
| LRTS source expansion — delivered in 4.9 | #303 larger histories and explicit stage views | Existing reviewed main; temporal blockers remain explicit |
| Java source matrix — delivered in 4.10 | #304 controlled CLI/Codec faults | Existing JUnit producer; admission remains separate |
| Python source — delivered in 4.11 | #305 non-Java fault reproduction | Existing generic controlled observations; one fix, two correlated locale contexts |
| Seeded C source — delivered in 4.12 | #306 permitted Juliet slice | One synthetic template/function pair; six correlated inputs, no empirical admission |
| HCS source — delivered in 4.13 | #307 LibSSH WTS/VTS and variant accounting | Original pipeline/job identities; 13 known labels, seven with parsed tests; temporal admission remains blocked |
| Synthetic laboratory — next | #308 interaction-covering stress generator | Existing reviewed main; preserve tiny oracle/resource bounds and explicit synthetic dependence |
| Evidence integration | #309 admission and production reality | #302–#308 |
| Comparison protocol | #310 metric references/statistics, #311 comparable baselines | #310 requires #309; #311 requires #309 and #310 |
| Forward feasibility | #312 prospective source feasibility | #309 |
| Final execution/report | #313 confirmatory corpus | #309–#312 and the complete parent gates |
Work started with #301 and #302, one reviewed slice at a time. Their compatible source deliveries were collected in a release branch and published together in 4.8 through #316. No release number or subtask closure substitutes for an actual source-integrity, diversity, temporal-admission or scientific evidence gate.
4.8.0 delivers #301's reviewed ZIP64 selection and complete RTPTorrent archive/member integrity through #314, plus #302's full analytical/native Parquet workflow, explicit eight-source-job simulation and checkpoint recovery through #315. The signed published wheel was exercised outside checkout with the actual source cache, native runs and checkpoint continuation. Remaining source admission, diversity and confirmatory work continues through the public subtasks. #303's new profiles preserve all raw/processed stages and use exact context/node bindings, while retaining explicit missing temporal evidence. Remaining expansion follows through the sequential public source subtasks.
4.9.0 publishes the stage-view contracts and three frozen presets: ActiveMQ, Log4j Ubuntu and Log4j Windows. All raw observations/processed stages remain analytical evidence; exact context/node bindings select one native view. The Log4j views share original source identities. Installed real-source workflows exercise native replay and checkpoints while keeping all temporal blockers. Existing single-stage LRTS producers, recipes and artifacts retain their semantics.
4.10.0 publishes #304's bounded Java fault matrix on the existing source producer: CLI-13, CLI-51 and an explicitly partial CODEC-65 fixture. Typed plans, raw/context verification, honest incomplete indices and offline recovery preserve all negative outcomes. Source qualification collected 714 role observations / 357 same-test pairs across two codebases. Controlled source collection remains separate from admission and policy comparison in #309 onward; the remaining source subtasks follow sequentially.
4.11.0 publishes #305's controlled Python capture. An unchanged PySnooper fixture reproduces a Unicode exception/fix under ASCII; the generic result remains inconclusive. Both revisions pass in the predeclared UTF-8 contrast. Whole-wheel component terms, actual runtime provenance, installed process interruption and offline recovery remain explicit. These two contexts share one curated fix, and source admission remains in #309.
4.12.0 publishes constructed C source capture with individually pinned permitted Juliet inputs, GCC/ABI provenance, a frozen arithmetic oracle and six correlated boundary-input pairs from one 2017 synthetic template. The qualification retains five both-pass and one output-assertion failure/pass pair; it does not establish six independent faults. The actual signed wheel was qualified outside checkout with source collection, offline replay and real process-interruption recovery. Shared transport guards and strict JUnit v2 ownership retain cleanup after output relocation; an actual published 4.11 v1 marker was also checked without rewriting its bytes. Existing generic datasets and native learning checkpoints keep their contracts. This source delivery does not admit a new empirical diversity regime or complete #261.
4.13.0 publishes #307's original-data HCS workflow. The frozen sixteen-pipeline slice retains 856 test executions, explicitly excludes 29 other-attempt records and preserves 208 cells across thirteen known job labels. Seven labels have parsed tests; six remain metadata-only, without inferred test semantics or fabricated native executions. Independent original-byte checks validate all cells and 240 WTS / 176 VTS result rows. Native checkpoint scopes separate exact budgets, reject ambiguous nested legacy paths and verify result ownership before reconciling uncommitted tails. The actual signed published wheel retains the previous source workflows and resumes completed published-4.12 single-budget results without changing their bytes; ownership metadata adoption is explicit. This is a post-hoc test-only simulation, not protected admission or evidence of a policy advantage. #308 is the next bounded implementation.
The existing public adapter issues now have consciously revised generic CI specifications in M4 — Point-in-time CI adapters and prospective evaluation:
- #258: point-in-time CI observations and versioned decision signals; implemented for 4.2 through existing temporal contracts and native extensions.
- #259: deterministic CI resource accounting and cost-aware replay; implemented for 4.3 on top of #258 and existing allocation/evaluation/benchmark foundations.
- #260: prospective shadow evaluation of frozen CI candidates; implemented for 4.3 on top of #259 and protected evaluation, without fabricating prospective observations.
Continue the remaining public corpus phases sequentially in later 4.x releases, with reusable core contracts and CI-centered fixtures. Provider dependencies stay outside the core. Generic mechanisms, native CI integration, functional recovery and actual prospective evidence each need explicit validation; offline replay must never masquerade as a forward observation.
Compatible features use subsequent 4.x minor versions; compatible fixes use 4.x patch versions. A major version requires an actual breaking public contract, not a large issue count. Release branches may group related reviewed PRs where needed, while security fixes remain independently shippable. Only implemented, reviewed, tested behavior belongs in published release notes.
Every slice reuses existing runner, results, artifact and checkpoint interfaces; includes installed-package examples and independent correctness evidence; passes local review and repository checks; and receives PR review before integration. No source admission, performance result, prospective observation, independence assumption or commercial value is inferred from a planned feature.