Competitive thesis
ScrolIQ does not replace the community's geometry, winding, fiber, scan-quality or ink tools. It makes their evidence composable and submission-grade: each result is bound to the exact CT / mesh / model / evaluation region it measures, leakage and missing evidence fail closed, and independent diagnostics can be carried into one reproducible Grand Prize evidence trail without silently becoming stronger claims than they support.
The October work below matters insofar as it strengthens that contract. The next leap is not another capability category; it is one exact Grand Prize volume with a demonstrably better held-out surface / fit / ink result. The protocol for that focused campaign is here →.
Update log
- 2026-10-01 · Mesh IQ on real segments. All 21 published Grand Prize segment meshes audited; one broken
PHerc1447 segment found (
20251105093211-z_dbg_gen_00320). - 2026-10-01 · Mesh IQ on the whole public corpus. Every published segment mesh in the open bucket: 843 of 903
TIFXYZ and all 200 original OBJs; 77 of 307 segments flagged with three or more defect kinds. New
scroliq-objreads triangle meshes; new usage guide by format; comparison with existing tools (#49). - 2026-10-01 · Truly-disjoint stability. Two-phase exclusion makes the resample disjoint by construction (64/64 zero chunk overlap): ρ = 0.63, replacing the inflated 0.7944.
- 2026-10-01 · Winding IQ on real annotations and fail-closed health. PHercParis4 ray-order audit with an
injection control;
scrollq-healthtreats missing integrity evidence as DO NOT TRAIN; companion pinned to zarr-pyramid-audit 0.4.0 (#55). - 2026-10-01 · VC3D-native Fiber IQ.
scroliq-fiberreadsvc3d_fiberJSON v1/v3/v4, binds to exact CT roots, and has a frozen public PHercParis4 campaign (#52–#54, #56). - 2026-10-01 · Pre-registrations. Mesh-flag review protocol and blinded sample (O1); stability v2 protocol with a balanced, interleaved sample (O5, #60).
- 2026-10-01 · Stability v2 result. Pre-registered verdict FAIL (ρ = 0.750); the leaderboard now shows reproducible rank bands instead of ranks (#61, #62).
- 2026-10-01 · Scoring hygiene. Dead-slice penalty budget sensitivity made explicit, results versioned as
scan-health-v1(#63); companion pinned to a full commit SHA in package metadata too (#64); TRAIN now requires complete quality evidence (#66). - 2026-10-02 · Grand Prize reviewer path. Readiness map and an exact-volume Grand Prize mesh cross-cut.
- 2026-10-03 · Winding IQ across collections. Pre-registered patch attachment on PHercParis4: 6 of 16,074 patch-collection constraints off by two or more windings, 206 of 254 relative collections tied to the absolute frame, injection control 194 / 200 (#65, #75).
- 2026-10-03 · Review queues now open natively in VC3D.
scroliq-vc3d-reviewexports diagnostic findings as VC3D PointCollections v1 with source hashes and review context preserved. The frozen Paris4 winding result becomes five physical review markers for its six flagged constraints, including both conflicting patches at the duplicated point (handoff contract; artifact).
1 Mesh IQ: from 21 meshes to the whole public corpus
- Grand Prize meshes. All 21 published TIFXYZ meshes for the two targets with public segments, audited
against each target's exact eligible CT volume. All 6 PHerc0800 meshes pass; 14 of 15 PHerc1447 meshes are coherent
with enclosed holes flagged for review;
20251105093211-z_dbg_gen_00320is broken (153 disconnected pieces, 1,335 edge jumps, 386 severe normal reversals, p95 local stretch 5.07). - Every published segment mesh. 307 segments on 11 scrolls. Tiers count distinct finding kinds: TIFXYZ 189 clean, 462 review, 192 multi-defect; OBJ 47 / 88 / 65. 77 segments are multi-defect on every registration (PHercParis4 27, PHerc0500P2 25, PHerc0814 12, PHerc0172 11, PHerc1447 1, PHerc1667 1). The 60 largest TIFXYZ meshes ran out of memory on a 16 GB runner and are listed with that reason, so the artifact is marked partial.
- Controls. The hand-confirmed broken PHerc1447 segment is flagged in both formats; all 26
z_dbg-named segments are flagged; TIFXYZ and OBJ agree on 175 of 185 segments. - Triangle meshes.
scroliq-objapplies the same finding kinds to Wavefront OBJ and adds non-manifold edges, inconsistent winding (kept separate from real folds), degenerate faces and folded-over UV triangles. An OBJ with no faces fails rather than passing. - Exact-volume Grand Prize cross-cut. Only 2 of the 13 eligible volumes have published exact-volume segment meshes: PHerc0800 (6 meshes, all clean) and PHerc1447 (15 meshes, 14 review, 1 multi-defect). The other 11 are an evidence gap, not a pass.
- How often are the flags right? Not known yet. The review protocol is frozen in
mesh-review-protocol.mdand a blinded sample of 54 segments (30 flagged, 4 debug controls, 20 clean) is drawn and awaiting VC3D review. No precision is claimed until it is labelled.
2 Scan-health stability: measured honestly, then acted on
- Truly disjoint. The September resample re-read 63% of chunks between "disjoint" runs. With exclusion by construction (64/64 zero overlap), Spearman ρ = 0.63, mean |Δ| = 7.95, top-10 overlap 4/10.
- Pre-registered v2. A balanced, interleaved sample (48 chunks per run, spread 7) was frozen before any data was read. Verdict FAIL: ρ = 0.750 on 48 eligible volumes, mean |Δ| 2.94, top-10 overlap 7/10, no shared chunk. The systematic run shift fell from +5.46 to +0.25 and Pearson r is 0.911: scores are reliable, but close volumes still swap ranks.
- Consequence applied. The leaderboard rank column now shows each volume's rank band (best and worst rank across run A, run B and their pooled score). The September scores differ from the v2 pooled score by 6.92 points on average, so the Grand Prize and First Letters frontiers need a new dated campaign; the September artifacts stay frozen. Done 2026-10-04: under v2, PHerc0813 leaves both frontiers; PHerc1447 (Grand Prize) and PHerc0800 (First Letters) are the only robust members (prize-frontier-v2).
- Known policy sensitivity. The dead-slice penalty sums detections, so it depends on the chunk budget (15 points at 6 chunks, the 30-point cap from 12). Made explicit and versioned, not silently changed.
3 Integrity: missing evidence never trains
scrollq-health now fails closed. An unreadable level, an absent root or a crashed audit is DO NOT
TRAIN, never a pass. TRAIN additionally requires finite quality of at least 40/100 and sampling provenance showing
every requested chunk decoded. The live re-run reproduces the published verdicts (PHerc0814 DO NOT TRAIN, PHerc0813
TRAIN) and adds an absent-root negative control.
4 Winding and fiber evidence on real data
- Winding IQ. Published PHercParis4
spiral-inputannotations: 2 of 13,700 comparable pairs out of radial order around the umbilicus, both under 0.1 voxel. Injecting one mis-numbered real point at a time, the check catches ±3 errors 179/200, ±5 171/200, and ±2 0/191 by design. - Fiber IQ. Reads native VC3D
vc3d_fiberJSON v1/v3/v4 and CSV, validates tracer and fallback provenance, can gate CI, and binds to an exact CT root for passports. A frozen public PHercParis4 run covers 8 hash-pinned fibers and 53,828 rendered points: 11 gap and 26 sharp-turn candidates, 0 control-line offsets or order inversions. The public directory is not asserted to name one CT volume, so this is not passport evidence.
5 Formats and documentation
- Usage by format maps OME-Zarr v3 (sharded) and v2, TIFXYZ, OBJ, PointCollections, fiber CSV/JSON and ink arrays to the command that reads each.
- Grand Prize readiness map marks every 2027 requirement as implemented evidence, partial, or not yet demonstrated, with links, and deliberately gives no readiness score.
- VC3D-native review handoff. Diagnostic queues can be converted to the same PointCollections v1 JSON that VC3D loads, so a machine finding can become an in-context human review point without coordinate transcription or a ScrolIQ-specific viewer.
How ScrolIQ compares with existing tools
ScrolIQ does not replace the community's tools; it reads their outputs and checks them against the exact data they claim to describe. Where a tool already does a job well, ScrolIQ consumes its report instead of reimplementing it.
| Job | Existing tool or practice | What ScrolIQ adds |
|---|---|---|
| Non-local self-intersection of a TIFXYZ surface | VC3D vc_tifxyz_selfcross |
Validates that report against the exact surface path and grid, fails closed on a stale or mismatched report, and adds the local checks selfcross does not make: components, enclosed holes, edge jumps, folds and Jacobian isometry |
| CT support of a surface | Villa vesuvius.surface_preflight |
Binds the report to the exact volume root and surface, and refuses a report made on a different scan of the same scroll. Adds the local geometry checks preflight does not make: on the same 21 published Grand Prize meshes, structure-only preflight passes all 21, including the broken PHerc1447 segment (head-to-head) |
| Finding broken segments | Opening segments one at a time in a viewer | One command over every published mesh, in TIFXYZ and OBJ, with a dated, reproducible flag list (903 + 200 meshes, 77 segments flagged on every registration) |
| Zarr store integrity | Reading the volume and noticing errors downstream | scrollq-health runs zarr-pyramid-audit and refuses to call a corrupt or unverifiable volume trainable |
| Taking a diagnostic into surface review | VC3D PointCollections and the VC3D agent bridge | scroliq-vc3d-review emits native PointCollections v1 markers, deduplicates coincident findings, and
keeps the exact diagnostic source hash plus machine context attached to each review point |
| Choosing which volume to label or train on next | We found no published per-volume scan comparison | A scan-health survey of all 64 volcomp volumes with sampling provenance, a published stability failure with rank bands, and a weight-free Grand Prize frontier |
Evidence
| claim | evidence (in repo) |
|---|---|
| Real-segment mesh audit: 21 Grand Prize meshes; PHerc0800 6/6 pass; PHerc1447 14/15 holes flagged for review, 1 broken | artifacts/2026-10-01-real-mesh-audit/ —
run.sh, per-segment reports with SHA-256s, summary.json |
Upstream preflight head-to-head: Villa vesuvius.surface_preflight (commit
078e9eb, unmodified, structure-only) passes all 21 published Grand Prize meshes on 8/8 gates, including
the PHerc1447 segment scroliq-mesh flags with 5 findings; same bytes, hash-checked. CT-support gates and
vc_tifxyz_selfcross not run |
artifacts/2026-10-01-preflight-comparison/ —
run.sh (pinned, hash-checked upstream module), preflight/, comparison.json |
| Corpus-wide mesh audit: 843/903 TIFXYZ + 200/200 OBJ; 77/307 segments flagged; controls flagged; formats agree on 175/185. Review candidates, not confirmed defects | artifacts/2026-10-01-corpus-mesh-audit/ —
run.sh, summary.json, flagged.tsv, failed.tsv |
| Exact-volume Grand Prize cross-cut: 2/13 eligible volumes have published meshes | artifacts/2026-10-02-grand-prize-mesh-crosscut/ |
| Mesh-flag review: protocol frozen, 54-segment blinded sample drawn; no labels yet | artifacts/2026-10-01-mesh-review-sample/,
mesh-review-protocol.md |
| Truly-disjoint stability: ρ = 0.63, |Δ| 7.95, top-10 4/10, 64/64 zero overlap | artifacts/2026-10-01-truly-disjoint/ |
| Stability v2 (pre-registered): FAIL, ρ = 0.750 on 48 volumes; rank bands on the leaderboard | artifacts/2026-10-stability-v2/,
artifacts/2026-10-01-stability-v2-prereg/,
stability-v2-protocol.md |
| Score refactor changed no score or sample | artifacts/2026-10-01-score-equivalence/ |
| Dead-slice penalty is chunk-budget sensitive (synthetic, deterministic) | artifacts/2026-10-01-dead-slice-sensitivity/ |
Fail-closed health verdicts: PHerc0814 DO NOT TRAIN (6 high LEVEL_NO_CHUNKS), PHerc0813 TRAIN
(quality 76.2), a v2 dev mesh CAUTION, a non-existent root DO NOT TRAIN |
artifacts/2026-10-01-health-verdicts-fail-closed/;
tests in tests/test_health.py |
| Winding ray order: 2/13,700 pairs inverted, both sub-voxel; injection recall ±3 179/200, ±5 171/200, ±2 0/191 by design | artifacts/2026-10-01-paris4-winding-ray-order/ |
| Winding consistency through patches: 6/16,074 constraints off by ≥ 2 windings; 2,111/2,232 points attached; 206/254 relative collections tied to the absolute frame; injection control detects 194/200 +2 errors | artifacts/2026-10-03-paris4-winding-attachment/,
winding-attachment-protocol.md |
| Public Fiber IQ: 8 hash-pinned fibers, 53,828 points; 11 gap, 26 sharp-turn candidates; not passport evidence | artifacts/2026-10-01-public-fiber-audit/ |
Honest scope
The 0–100 ScrolIQ score is scan-health triage only. It is not readability, surface quality, ink quality, or Grand Prize readiness, and its ranking failed our own stability gate twice. Mesh, winding and fiber flags are review candidates: none has been confirmed by a person in VC3D yet, and nothing has been filed upstream. Local mesh checks do not prove a surface follows the papyrus sheet.