docs(detection-core): bind integration phase into roadmap and task plan

This commit is contained in:
Marius Ciepluch
2026-10-30 18:00:59 +01:00
parent 1cd6a39aaf
commit b5c26f52e8
2 changed files with 61 additions and 0 deletions
+35
View File
@@ -1,5 +1,40 @@
# Roadmap
## Binding phase -- detection-core as the shared contract (2026-10-30)
This section supersedes ordering questions below it for detection work. `detection-core`
(spec complete, validated GO at 1cd6a39) is now the binding element: its contract package
(`src/openultrasast/core/`) owns the canonical trace, verdict, coverage, measured budget,
provenance, admission, probes, fact-pack and consumer contracts. Existing specs adopt it;
they do not grow parallel runner contracts.
### Integration sequence (each step admission-gated on the untouched pilot)
1. **Prover native conformance** -- `access-control-prover`'s `Trace` implements
`core.TraceRecord` directly; `hunt/core_adapter.py` is deleted once conformance tests
pass. Touchpoint: access-control-prover.
2. **Fact-pack workstream (recognition breadth)** -- `valid-detection-benchmark` task 7.1's
manifest-derived recognition table becomes reviewed fact packs admitted via
`core/factpack.py` (families tagged, digest-stamped). Target ecosystems: the valid
pool's Express/FastAPI/NestJS/Koa/MCP-style servers. Touchpoints: valid-detection-benchmark,
ruleset loader.
3. **Resolution depth** -- interprocedural call resolution (today 11.9% complete) in the
engine/frontends; measured via `core/coverage.py` buckets. Touchpoint: contributor-scan
(engine internals stay out of detection-core).
4. **Hook consumes the core** -- `pre-push-safety-net` builds `PushRequest` and consumes
`PushVerdict` via `core/consumer.py`, budget from the measured per-class derivation
(`benchmarks/measurements/2026-10-30-detection-core-budget/`), one real push E2E on a
dev-pool repo. Touchpoints: pre-push-safety-net.
5. **Proof layer attaches** -- `search-with-proof` attaches demonstrated evidence to
`PushVerdict` claims; BLOCK only on proven claims. Touchpoint: search-with-proof.
6. **Model lane bounded** -- `hunter-learning-loop` proposes only; every fitted component
passes `core/probes.py` (reversed-pair + holdout + admission GO). Touchpoint:
hunter-learning-loop.
Stop conditions from `detection-quality.md` apply unchanged: two consecutive admission
NO-GOs on the same arm park it for a strategy review; the fallback is the narrowed
one-ecosystem advisory product plus proof-based BLOCK.
## Roadmap -- admission read, stop conditions (2026-10-30)
This section supersedes the ordering below it for detection claims. The 2026-10-30 admission