B04 / QDM Readiness Scoreboard¶

Current increment: 29 (Realistic-density routed-block external validity: ExactVectorCropSource + Ibex gate) Current HEAD: d9d0f8a (post-Inc28) + streaming architecture (RFC 0008) + Inc16–29 proof modules Last updated: 2026-09-10

This is the canonical live scoreboard required by the B04 architecture brief (§18/§22). Historical per-increment notes remain in proof_artifacts/README_B04_PATCH.md.

Status summary¶

EXACT-VECTOR PROOF PATH:            PASS
STREAMING FOUNDATION:               PASS
CERTIFIED-HALO FOUNDATION:          PARTIAL (brackets exist, tight closure open)
PLUGIN FOUNDATION:                  PASS
SOURCE-NATIVE VERIFIER FOUNDATION:  PASS (interfaces + reference shell)
CONTINUOUS CERTIFICATION:           PARTIAL (bridge characterized, interval eval OPEN)
PRE-FORWARD SCREENING:              PASS (fail-closed, certified decisions only)
VERIFICATION EXECUTION SEMANTICS:   R17 HARDENED (per-verifier sessions, true subdivision, dual ledgers)
REAL-DENSITY SCREENING VALIDITY:    MEASURED (Inc29 Ibex routed crops, MODERATE verdict)
FULL-CHIP DENSE ALLOCATION:         ABSENT (MetricOnlyTileSink path verified)
ACTIVE-WORK INSTRUMENTATION:        PASS (WorkAccounting driven by run_streaming)
LARGE-LAYOUT MEMORY SCALING:        PASS (memmap in/out, O(tile+batch) verified)
LARGE-LAYOUT SPEED CROSSOVER:       MEASURED (benchmark_b04_inc21/22/23/24/26/28)
REAL GDS/OAS SEMANTICS:             PASS (KLayout-gated semantic tests)
SIMULATOR INTEROPERABILITY:         PASS (backend contract + reference Hopkins)
PHYSICS DOMAIN VALIDATION:          OPEN (synthetic Hopkins, not foundry-calibrated)
INDUSTRIAL VALUE:                   PROMISING
QDM READINESS:                      CONDITIONAL

What is proved (with evidence)¶

Exact-vector proof path¶

EXACT_VECTOR_PIXEL_CENTER_INDICATOR is enforced by contract (ExactVectorMaskContract.__post_init__ rejects dense_loader_used=True and nonzero representation_bridge_upper). The dense/PIL loader is absent from the proof path. Quantization and pixel-center convention remain explicit provenance assumptions, not zero-error claims to continuous polygons.

Streaming foundation¶

TileSource (Tensor, Memmap, VectorLayout), TileSink (Tensor, Memmap, MetricOnly), TileScheduler with verifier-driven refinement, and run_streaming pipeline with O(tile area + active batch) peak memory. Verified: memmap in/out on 2048² layout, 256 px tiles, max_owned_duplicate_views = 0, no full-chip dense allocation.

Certified halo (bracket, not tight closure)¶

verify/halo.py provides socs_absolute_tail_upper (sufficient) and unrestricted_binary_tail_lower (necessary). On the 72×72 ArF K=24 snapshot, ε=2.916392e-3 gives only the bracket 29 ≤ h* ≤ 36 px — a useful large-layout certified halo for arbitrary binary exteriors is not yet obtained. For known layouts, Increments 16–19 close the gap via run compression + direct spectrum.

Global finite spectral statistic (Inc 26)¶

global_finite_spectral_summary streams the complete finite layout into arbitrary-frequency Fourier moments with spatial_halo_error_upper = 0 for the declared finite plane-wave quadrature model. This eliminates the spatial halo as an error source for that model — the proof obligation moves to pupil/source quadrature.

Continuous mask lift + fixed-source Arb pupil (Inc 27)¶

continuous_square_mask.py gives the exact lift M̂_c(f) = p² sinc_π(p fₓ) sinc_π(p f_y) M_point(f) with representation_bridge_upper = 0 for the declared square-aperture semantics. arb_fixed_source_pupil.py certifies the one-source-point coherent field as a rigorous complex ball (python-flint acb.integral, hard pupil mapped to a fixed rectangle). Independent proof backend; source-plane quadrature and partial-coherence enclosure remain open.

Scale-first work avoidance (Inc 28)¶

screening.py adds a fail-closed pre-forward screen: tiles may be skipped only with a certified SCREENED_OUT decision carrying an exact fill value, and only when the screen explicitly certifies any attached verification plugins. WorkAccounting is driven by the actual run_streaming loop (unique active area, screened-out area, read-window/forward/screen counters, accounted_pct closure). verify_layout() surfaces VerificationResult.work_accounting. The Inc28 scale-first ladder (512²–16384², dense_full/tiled_raster/b04_vector/b04_selective) measures wall time, peak RSS and crossover fields with zero dense-allocation events on the selective path.

Realistic-density routed-block validity (Inc 29)¶

On the pinned public PDB Sky130HD Ibex routed GDS (15515 cells, 0.579 utilization) via deterministic nested center crops, exact-empty-context screening still discharges real work at every scale: active fraction 0.3125 (4096²) → 0.5391 (8192²) → 0.5117 (16384²), with a stable 0.570 screen-only diagnostic at 32768²; peak RSS stays flat (~1.4 GB) while the dense baseline already needs 2.0 GB at 4096²; zero dense allocations on the selective path; correctness witness max_abs_error = 0.0. Verdict: MODERATE_ON_REALISTIC_DENSITY — empty screening is retained, and the next work-avoidance mainline adds nonempty-tile (sensitivity-collar) screening. Honest negative: no wall-time crossover at real density in the tested range (selective is slower than dense/tiled per common-size comparison); the B04 claim at realistic density is certified work avoidance and bounded memory, not speed.

Semantic hardening (Inc 22)¶

PhysicalInstanceKey distinguishes source repetition from read repetition. KLayoutAlignedRunSource stamps physical:<sha256> owner ids. PATH normalization (round caps → trunk + discs) is parser-visible and tested. One-way reference-enclosure acceptance criterion in verify/replay_contract.py.

Raster-bridge characterization (Inc 23)¶

On the 512² fixture, the dense loader's 6226-pixel symmetric difference is entirely false positives (zero false negatives), mutual Chebyshev radius 1 px, Hausdorff upper 11.31 nm @ 8 nm pixels. The module does not assert loader/exact-vector equivalence.

Optical raster-to-contour bridge (Inc 24)¶

SOCS modewise intensity bridge Σ w_j(2A_jB_j+B_j²), delta_I/κ contour transfer, ProcessWindowBridgeBudget additive composition. Periodic 128 px Hopkins tile diagnostic: max intensity delta ~0.17, EPE ~24 nm. Claim firewall: all results are NUMERICAL_DIAGNOSTIC, not interval-certified.

Proof-carrying verification replay chain¶

Increment 11–15 frozen artifacts replay in CI: 164 active cells, 148 seeded, ALL_COMPONENTS_COVERED, Hausdorff uppers 0.976/1.056/2.032 nm. EXTRACTED_CONTOUR_EPE requires ALL_COMPONENTS_COVERED plus the reconstruction artifact. epe_max_nm semantics untouched.

Remaining blockers to QDM READY¶

Pupil/source quadrature remainder: interval-certified evaluation of the frozen Hopkins/SOCS operator on the compact optical frequency domain (Increment 26 defines the finite moment vector; Increment 27 certifies the single-fixed-source coherent field ball — composing it over a certified source-plane quadrature is the next proof obligation).
Certified oriented-slab spatial extraction: wire the seed/slab theorem's η_sp into TileVerificationResult.spatial_extraction_error.
Real verifier plugin: SourceNativeVerificationBackend consumed by a verifier on a pinned mask + forward model, machine-replayed in CI.
Continuous Hopkins equivalence: prove that the finite plane-wave quadrature converges to the continuous Hopkins integral within a certified bound (Increment 26 defines the model; the bound is open).
Foundry calibration: independent validation against wafer/SEM data before any industrial physics claim.

What is explicitly NOT proved¶

zero error to arbitrary continuous polygon/wafer models
foundry-calibrated optical accuracy
practical full-chip speedup over dense raster at production scale
certified process-window on real layouts
complete GDS/OASIS semantic coverage