QDM / B04 Plugin Readiness Checklist¶

Status after RFC 0008 foundation + B04 Increment 14 integration (2026-09-10). This checklist is the honest scoreboard required by the B04 architecture brief (§23).

Foundation statuses¶

STREAMING FOUNDATION:            PASS
CERTIFIED-HALO FOUNDATION:       PASS
PLUGIN FOUNDATION:               PASS
SOURCE-NATIVE VERIFIER FOUNDATION: PASS (interfaces + reference shell only)
CONTINUOUS-METRIC FIREWALL:      PASS
B04 ACTUAL-MODEL READINESS:      READY
QDM READINESS:                   CONDITIONAL

Evidence map¶

Requirement	Evidence
A. big-layout streaming under budget	`tests/test_streaming/test_pipeline.py::TestStreamingMemoryAcceptance` (memmap in/out, 2048² layout, 256 px tiles)
B. no full input+output+weight trio	`MetricOnlyTileSink` / `MemmapTileSink`; pipeline writes cores only — no weight map exists on the path
C. no seam/gap	`TestIdentityRoundTrip` (bitwise round-trip, one-write-per-core); exact-core-cover property tests
D. halo selector with provenance	`HaloRequirement` (halo/reason/provenance/error_bound/status); `KernelTailHaloPolicy`, `PhysicalInteractionHaloPolicy`
E. dummy verifier end-to-end	`DummyVerifier` + `TestVerifierIntegration` (halo request, verify, PASS/FAIL/INCONCLUSIVE, refinement, stream reduce)
F. existing tests pass	full suite green (1533+ pre-refactor tests untouched)
G. old vs streaming benchmark	`benchmarks/benchmark_streaming.py`
H. source-native foundation	`verify/source_snapshot.py`, `verify/source_native.py` (opt-in `source_native_full` backend, no top-K dependency)
I. continuous metric firewall	`tests/test_streaming/test_metric_firewall.py`; `epe_max_nm` semantics unchanged
J. budget composition	`StreamingVerificationReducer.error_budget` + `test_error_budget_composition`
K. coverage semantics	`PARTIAL_COVER != PASS` pinned in reducer tests and `verify/spatial.reduce_coverage`
B04 actual model	`tests/test_verify/` replays the frozen Increment 11–14 artifacts (164 cells, 148 seeded, `ALL_COMPONENTS_COVERED`, Hausdorff uppers 0.976/1.056/2.032 nm)
Certified-halo brackets	`verify/halo.py` + Increment 15 artifact: on the 72×72 ArF K=24 snapshot, ε=2.916392e-3 gives only `29 ≤ h* ≤ 36` px (PARTIAL). The unrestricted-binary lower bound proves no smaller halo can be certified — a useful large-layout certified halo for arbitrary binary exteriors is not yet obtained and is honestly reported as `PARTIAL_NONTRIVIAL_SUFFICIENT_HALO_NOT_OBTAINED`

Why QDM READINESS is CONDITIONAL¶

No QDM mathematics is implemented (explicitly out of scope this round).
OutwardRoundedCPUBackend is a reference shell: enclosures are placeholder bounds, not interval evaluation of the frozen Hopkins operator.
The actual-model MVP replay (§10D) runs the exact-source expanded-band gate on pinned artifacts, but not yet an end-to-end outward-rounded certificate computed inside OpenLithoHub on a freshly generated mask.

Remaining blockers to QDM READY¶

Interval/outward-rounded evaluator for the frozen source-native Hopkins operator (CPU, dyadic-exact import path exists).
Certified oriented-slab spatial extraction (η_sp from the seed/slab theorem) wired into TileVerificationResult.spatial_extraction_error.
SourceNativeVerificationBackend consumed by a real verifier plugin on a pinned mask + pinned forward model, machine-replayed in CI.
Increment 15 no-go: for arbitrary binary exterior perturbations, the sufficient halo proof only closes at the full half-tile radius. A useful certified halo needs either an exterior regularity/interface class (polygon/TV/edge-density control) or a known-layout streaming tail oracle. No exponential spatial decay is assumed.