6.6 KiB
G1 registration complete; new threshold-band review
The original 2,012-pair queue completed normally on 2026-11-15 at approximately 01:05:55 UTC+8 (completion-file timestamp). Session 89763 returned exit code 0. Every worker returned PASS, with zero failures and zero retries; the final failure-rate checkpoint is 0%. No active worker remains, stop_reason is null, and batch-complete.json says ALL_REGISTRATIONS_COMPLETE. The original log reached both 2,000 and 2,012 terminal jobs. No candidate generation or registration was restarted.
| Resource | Final measurement |
|---|---|
| Worker CPU | 167,324.123 s = 46.479 h |
| Guards CPU | 2,031.538 s |
| Coordinator CPU | 3,731.229 s |
| Wall time | 10,856.804 s = 180.947 min |
| Peak aggregate process-tree RSS | 2,426,810,368 bytes = 2.260 GiB |
| Output disk | 236,072,960 bytes = 225.137 MiB |
The fixed 100 worker CPU-h, 8 h wall, 20 GiB disk/aggregate RSS and 16-worker limits were respected. Supervisor/coordination CPU is separately recorded. All 2,012 original terminal receipts are preserved in artifacts/s2-g1-band25-2026-11-15/public/registration-terminal-receipts.json; the compact summary and custody seals are in registration-final.json. Original detailed registration outputs remain intact under their existing custody root, with a manifest digest. Completion receipt SHA256: b0ee7469c1ab755e809f8ff85ebc6ce940e9838a23eb091a4a7a0e91e77d0956. The registration contract remains c4271622e1775d430324dd8e8ad4524006184122079d2fc3bbb8fd0ba6a349d8.
Fixed-rule results and scientific limits
The original, pre-registration thresholds/rules yield 90 LINK_UNDER_FIXED_RULE, 25 BAND_REVIEW_REQUIRED, and 1,897 OUTSIDE_FIXED_RULE. These are engineering decisions, not independent geometric relationship labels. In particular, outside the rule does not certify unrelatedness, and this candidate screen does not certify cross-key recall. The existing stage-review decision uses conservative grouping to reduce leakage, with loss of statistical power accepted; it does not establish physical equivalence or a population relatedness rate.
The 25 band pairs cover 32 unique source geometries. They are all and only the pairs whose minimum eligible-route residual/threshold score is in (1,2]; no downsampling, re-ranking, new threshold or inspection-based selection is applied. The 90 direct links are retained for the later component union. Same source key stays on the same side. The old 120-pair blind review and its witness are not repeated.
Anonymous presentation and handoff
tools/prepare_g1_band_review.py consumes only the already frozen geometry-only caches and completed registration metadata. It validates the original input/source hashes, seals the full band-to-source map, randomises pair order and left/right assignment, and uses independent proper rotation/translation for all 50 exported instances. No scale change, fitted alignment, reflection, simplification or original archive/field read occurs. The existing checked geometry exporter, rigid-pose invariant checks and fixed-view renderer are reused.
New material preparation is CPU-only, at most four workers, capped at 1 CPU-hour / 30 minutes wall / 5 GiB disk and aggregate RSS, with 1.25 GiB per worker. The existing supervisor bounds the process tree; the parent additionally checks worker RSS. Failed attempts would remain recorded without retry. These are presentation limits, not additional registration or training authority.
The blind package contains only 25 opaque pair IDs, 50 original-unit PLY surfaces, 100 PNGs (four-view overview and fixed quadrant grid per instance), pair/whole-package hashes, blank labels, instructions and an unsigned reviewer attestation. Shape and vertex counts remain visible; random posing mitigates identification but cannot guarantee anonymity. Raster marks and automatic crops are not certified evidence of holes/bridges; full surfaces and explicit unresolved judgments remain available.
Identity, source key, original rank, numerical residual, branch labels, registration transform and random pose mapping remain under custody, outside the review package and its isolated Git branch. The sealed mapping SHA256 is 7531566edf060cdae522b7cf32d0164d0cc6bd9d5b26be6a8d1c3a53bbf61142. The reviewer is asked not to consult identity/registration material and to freeze labels plus attestation before the custody join. Per Claude's 01:08:13 ruling, use exactly three labels: rigid / same part, unrelated, UNRESOLVED, with local evidence. Under the standing ruling, UNRESOLVED conservatively merges; it does not create another HOLD or an equivalence claim.
The first, not-yet-shared package was preserved in superseded-before-handoff/; version 2 updates the three-label instructions and records full render/pose settings, per-instance vertex/triangle counts and seed seals in its manifest. No geometry was rerendered. The secret entropy that also determines anonymous assignment remains sealed until all 25 verdicts are frozen: releasing it early would allow enumeration of original registration indices. The fixed exported geometry can already reproduce the images using the public camera settings without unsealing source assignment. Public manifests give the seed and custody hashes, while exact entropy/poses remain recoverable under custody.
All 25 presentation workers passed their existing geometry/schema/pose checks. Both the original packaging and the pre-handoff amendment passed full byte-for-byte ZIP comparisons with normalised archive timestamps. The final archive digest and manifest digest are in public/package-result.json. Presentation consumed 113.977 CPU-s / 35.476 wall-s, with peak aggregate RSS 539,549,696 bytes; packaging consumed 5.214 CPU-s. Separate presentation, packaging and packet-finalization resource receipts preserve all attempts. These are material-production results, not blind-review verdicts.
Claude reviews the final receipt and new package contract and requested blind-first review; Mold receives only this new anonymous band package for review. Then the frozen review is joined under custody, components are frozen, splits are generated, the 10-part preprocessing is measured, and the already specified pilot follows. Before splitting, report component count/size distribution, the largest component's share of parts and nodes, and parts per source key. If a component prevents balancing five folds, raise it before splitting; do not silently re-cut it. Components, splits and pilot are not reported as completed by this delivery. The synthetic R2-G24/N/E code and normalisation fix already accepted at f3532962 remain unchanged.