Repository navigation
docs: reconcile harness adapter schemas and JEV contract against 0.15.29 - #662
Conversation
…custom-profile scope
tommy-ca
left a comment
There was a problem hiding this comment.
Thermo-Nuclear / Code Quality review — exact docs PR diff at d6352cb
Scope: Only the six added/modified files in PR #662. The JEV P0/P1 findings #658/#659/#660/#663/#664 concern pre-existing runtime files and are tracked separately; they are not introduced by this PR, consistent with the upstream Thermo-Nuclear review scope.
Correctness / security diff audit: README and porting-guide links point to the new source-pinned architecture document. The new document distinguishes eight descriptive profiles, six built-in host targets, candidate versus packaged states, four OMP scenario-pinned capabilities, model planning versus host-native execution, and JEV transport mock versus real provider proof. The document explicitly flags protocol mismatch #664, raw-query egress #658 and synthetic evidence #659. No runtime code, feature gate, provider key, or deployment behavior changes in this diff.
Code-quality / 'code judo': Adding one tiny source-derived docs contract suite is preferable to copying status tables/role registries or inventing a schema compiler. The added CI step reuses existing pytest and doesn't add dependencies, a scheduler, second router or provider SDK. No changed source file crosses the Thermos 1,000-line threshold. The new scope changes test_shadow_routing.py from outside-gate to offline; corresponding check-selection test was correctly reconciled.
Observed verification: Portable offline workflow #39483367641 passed at exact PR head d6352cb. Earlier run #39482879748 failed one test-tier expectation on the first CI expansion; fixed and re-run. This check proves hermetic tests, not authenticated JEV transport, live host execution, or independent model-family judging.
Acceptance / gate: Documentation-only diff has no observed new P0/P1 defect. Keep this as a draft until independent reviewer sign-off and the declared docs acceptance workflow is complete. Do not merge safety/runtime corrections under this docs PR. Swarm/Arena/Interrogate are requested future native runs, not claims made by this single-reviewer source audit. Parent #657 maintains those unresolved gates.
|
Independent Verification Verdict: Verified against head
|
Purpose
Reconcile user-facing harness and JEV design documentation with the current source at the audited 2026-10-20 trunk
dea54d5(canonical upstream 0.15.29). Source audit, not a claim of runtime or authenticated JEV verification.Changes
docs/architecture/harness-adapters-and-jev-2026-10-20.md: architecture/ownership, all 12 schema contracts, eight profile vs six built-in matrix, model/role resolver, package projector/driver/receipt planes, Jev neutral decision objects and present System Oneurllibtransport versus proposed SDK.docs/guide/12-porting.mdwith a link to current snapshot and tag older five-host acceptance narrative historical; updateREADME.mdnavigation.Evidence / outstanding controls
pstack.package.json, all eight profiles, schema/driver/core files, living JEV spec, relevant existing tests. No local pytest, native-host runtime, independent agent swarm or keyed JEV smoke has run in this docs-only PR. Do not count source inspection as live PASS.Requested review (Thermo / Swarm / Arena / Interrogate)
python3 scripts/project-package.py --check, docs tests,pytesttargeted decision/adapters tests, then only operator-authorized native/paid runtime smoke. Record SHA and separate BLOCKED from PASS.No code or configured routing behavior changed; keep as draft until docs verification and reviewers accept it.
Exact-head test reconciliation — 2026-10-20
d6352cb5940b8f319d025caa54db1088266a7c0c, offline workflow run #39483367641: SUCCESS. Addstests/test_architecture_documentation.pyand offline JEV/advisory/rollback eval suite to.github/workflows/portability-offline.yml.tests/test_shadow_routing.pyasoutside_gateeven though it was newly included in CI. Test expectation changed toofflinein the same PR; next run passed.questions+answersobjects do not match code'squestions:[]andresults:[]mocks. Other opt-in blockers remain [JEV][P0] Never transmit raw task queries or call provider on bypass routes #658/[JEV][P1] Gate advisory promotion on persisted pre-existing calibrated evidence #659/[JEV][P1] Reject nonfinite confidence, incomplete response identity and unbounded paid calls #660/[JEV][P1] Fail closed when decision-routing registry drifts from applied canonical pin #663. No native-host/real-provider proof and no independent agent reviews. Keep PR draft pending review.