ScrolIQ

(1) Which scroll data did you work on?

ScrolIQ was run on the public scroll-volume corpus rather than a hand-picked showcase:

• 64 / 64 listed volcomp CT volumes in volumes.txt, decoded from the public dl.ash2txt.org store with the real volcomp decoder. The frozen September campaign uses 24 requested L0 chunks per volume from a 5×5×5 candidate grid.
• Open-data label and segment coverage joined from the companion S3 inventory. The September inventory found 70 ink-detection roots associated with PHercParis4; the exact-volume join leaves 14 top-quartile campaign entries with no matched ink labels and marks them label_next.
• All 13 then-current 2027 Grand Prize target volumes, matched by exact prize-listed volume ID. The qualifier explicitly excludes same-scroll higher-resolution substitutes; for PHerc1203, the 2.403 µm scan is recorded as excluded while the prize-listed 9.362 µm volume is used.
• All 22 First Letters prize scans, matched by the exact scan IDs on the prize page, plus the released m7 surface prediction made on each of those scans (public vesuvius-challenge-open-data bucket), read at level 0 alongside the masked CT.
• The Challenge's open S3 bucket in its OME-Zarr v2 layout: stored chunk sizes sampled across all 65 indexed volumes and checked exhaustively for the 23 prize-eligible volumes, plus four rescans registered to their original scan by the published transform.json.
• PHerc1667 as a known-success control for the spatial scan metrics, so target observations can be compared against a scroll that has already been completely virtually unwrapped and read.

(2) How does it increase the probability of reading the scrolls?

The scarce resources in virtual unwrapping are expert attention, annotation time, and compute. ScrolIQ makes the scan-quality part of that allocation decision measurable instead of implicit. It does not claim to know which scroll contains the easiest text; it tells the team what the CT data actually supports before expensive downstream work begins.

• Real voxels, not metadata proxies. Each score is built from decoded level-0 data: signal presence, gradient/texture energy, dynamic range, saturation, and a strict dead-slice scan.
• Sampling provenance travels with every score. Requested, decoded, missing, sparse/background, transport-failure, and decode-failure states are recorded so an apparently precise score cannot hide an incomplete read.
• Label coverage is joined to scan health. High-health exact-volume entries with no matched ink labels are surfaced as label_next candidates, turning an invisible allocation gap into an inspectable queue.
• Spatial evidence replaces one-number blindness. scroliq-scan-map retains the L0 voxel bounding box of each observation, while scroliq-passport can attach exact-volume winding, native TIFXYZ mesh, and ink-validation evidence without upgrading still-unmeasured surface, fiber, or spiral stages from unknown.
• Integrity and quality can be checked together. scrollq-health combines this triage layer with zarr-pyramid-audit, so structurally defective data can be rejected before a quality score is trusted.

(3) What does it enable that was not possible before?

• A reproducible corpus-wide scan-health survey. One command scores the public volume list with real decoded voxels and emits machine-readable per-volume components plus sampling provenance.
• Uncertainty-aware prioritization. The repository now includes a deliberately disjoint resampling test, per-chunk score distributions, and a public stability gate. A misleadingly stable 3×3×3 result was superseded after we proved it was re-reading the same sparse shards.
• Exact-volume Grand Prize triage. scrollq-grand-prize compares all 13 target volumes on a weight-free Pareto frontier over scan-health score and existing segment count, while pinning the exact eligible volume and recording prohibited same-scroll higher-resolution alternatives.
• Exact-scan First Letters triage. scrollq-grand-prize --prize first-letters qualifies all 22 First Letters scans. Each prize manifest declares its own required bootstrap assets, so the 9 eligible scrolls without a lasagna prediction are not silently dropped. Run on both disjoint resample runs, it shows which frontier members are robust and which are sampling noise.
• Exact-scan surface-prediction CT support. scroliq-support measures how much of a released surface prediction lies on real CT material (prediction voxel > 127 where masked CT is exactly 0 counts as phantom), on the exact eligible scan, with a bootstrap 95% interval. A prediction made on a different scan of the same scroll fails closed. External surveys are now imported by a script pinned to the source commit instead of by hand.
• Open-bucket access and integrity. A dependency-light OME-Zarr v2 reader for the bucket VC3D streams from, and scroliq-chunk-audit, which checks every stored chunk object against the size its .zarray declares without downloading chunks.
• A pre-registered test of the scan metrics. scroliq-pairs asks whether intensity-invariant metrics rate the documented-better scan protocol higher on registered rescans. The hypothesis and decision rule were committed before any real-data metric was computed.
• Prize manifests derived, not hand-copied. scroliq-manifest derives the Grand Prize and First Letters target lists from the pinned official eligibility list and bucket index, and tests fail if the hand-entered manifests drift from them.
• Fiber trace continuity. scroliq-fiber lists trace-break candidates in ordered fiber traces (gaps relative to each trace's median step, sharp turns), records the input's SHA-256, and fails on any parse error. It does not prove fiber or sheet identity, so the passport's fiber stage stays unknown.
• Evidence passports instead of synthetic readiness scores. scroliq-passport organizes data, scan, label, winding, native TIFXYZ mesh, and ink-validation evidence around the Challenge bottlenecks and leaves unmeasured stages unknown.
• Submission-lineage validation. scroliq-provenance checks exact-volume lineage, training/prediction separation, public dataset/checkpoint provenance, stochastic seeds, experiment runs, mesh/render traceability, scale-bar declarations, banner coverage, and public held-out validation against known ground truth; same-volume training/validation overlap fails closed.
• Winding-input diagnostics. scroliq-winding audits PointCollections v1 inputs, SHA-256 provenance, role semantics, numeric sanity, volume binding, and axial annotation coverage before winding constraints are trusted as geometry evidence.
• Native Mesh IQ. scroliq-mesh audits TIFXYZ structure and hashes, mask-aware connectivity and enclosed gaps, bbox/scale consistency, local edge jumps, severe normal reversals, symmetric quad-area distortion, and per-triangle Jacobian singular-value isometry (which catches area-preserving anisotropic stretch that an area check misses). It can also validate an official VC3D vc_tifxyz_selfcross census and a Villa vesuvius.surface_preflight CT-support report against the exact surface and volume being audited, failing closed on stale or mismatched evidence. None of these local checks proves sheet identity.
• Ink falsification discipline. scroliq-ink-audit checks exact-volume experiment binding, checkpoint identity, train/evaluation separation, held-out declarations, run provenance, and controls including normal offsets, adjacent winding, geometry perturbation, and an independent checkpoint. Missing controls remain visibly partial rather than being inferred.
• Held-out ink validation. scroliq-ink-validate scores a prediction and its falsification controls (normal offsets, adjacent winding) against the same known ink labels and validation mask, and emits a hash-pinnable report for the provenance graph. It measures ink-signal recovery, not reading. See ink-validation.md. Provenance schema v4 binds each held-out validation to its exact report: checkpoint digest, evaluated-array digest, model window and named falsification controls must match, or the gate fails closed.
• Certified held-out geometry. scrollq-geometry-probe freezes a blind fit/held-out regional split for the Grand Prize probe and certifies a predeclared minimum spatial gap between fit and held-out boxes, so held-out spiral-fit residuals cannot quietly leak from nearby fit inputs. See the blind probe protocol.

(4) What evidence have you provided?

2027 Grand Prize bridge

ScrolIQ is not an unrolling or ink-detection submission by itself. Its role is to make the path to a qualifying full-scroll submission harder to fool: choose exact eligible data, preserve spatial evidence, expose missing constraints, bind artifacts to the correct CT root, and make provenance fail closed before a result is packaged.

The current Grand Prize tooling therefore favors auditability over a single winner score. The primary target qualifier publishes a Pareto frontier instead of hiding trade-offs inside arbitrary weights; scroliq-recto-coverage requires every declared main/attached/detached component to be unrolled and enforces the strict <10% disconnected-outer-patch exception; provenance schema v3 additionally requires the ledger's mesh-ID set to match the submitted mesh set exactly; passports keep unmeasured stages unknown; provenance validation rejects wrong-volume lineage, training/prediction overlap, and training/held-out overlap while requiring public known-ground-truth validation; mesh and ink evidence remain explicitly partial unless their own direct diagnostics pass.

Honest scope

The 0–100 ScrolIQ score is scan-health triage only. It is not readability, surface quality, ink quality, or Grand Prize readiness. It does not infer text from CT texture. A high score does not mean a scroll will read; a low score does not prove it will not. The score weights are visible in src/scrollq/score.py, and the September stability gate is published as failed. Surface-prediction CT support is likewise a geometry-prior sanity check: it says how much of a released prediction sits on scanned material, not whether that surface carries ink or text. And because the pre-registered protocol-pair test came out negative, scores from different scans of the same scroll are not comparable.

This distinction is part of the contribution: every later ScrolIQ layer is designed to keep measured, partial, and unknown states separate instead of converting missing evidence into confidence.

One suite, three layers

Team

Solo human contributor: Sven. Muse (AI assistant) was used for implementation, analysis, documentation, and review support. Code is MIT licensed; campaign outputs and the exact evidence behind the published claims are committed in the repository.