ScrolIQ is not counting commits toward a prize. It is tracking the evidence required to turn an exact Herculaneum CT volume into a complete, reproducible, VC3D-compatible, held-out-validated virtual unrolling. Every stage stays measured, partial, unknown, or blocked until the evidence exists.
No exact 2027 Grand Prize volume currently has a ScrolIQ-backed whole-scroll recto segmentation/unroll. Detached or broken patches are not yet closed out either.
Native TIFXYZ/OBJ auditing, isometry checks, self-cross binding, flattening comparison, package hashing, and column traceability exist. A complete sequential column set does not.
Leakage audits, held-out ink scoring, falsification controls, legibility interfaces, and submission-image traceability exist. No full-scroll sequence of qualifying readable column renders exists yet.
scroliq-eval provides model/dataset/result contracts, checkpoint and code hashing, overlap blocking, fixed seeds, region-level failures, and deterministic 95% bootstrap CIs. scroliq-segmentation-validate now adds blind TIFXYZ surface scoring with a salted frozen-truth commitment.
ScrolIQ reads native TIFXYZ, VC3D fiber JSON, PointCollections, upstream self-cross reports, and emits review queues in native coordinates. Winding attachment and ray-order findings can now be exported as VC3D PointCollections with source hashes and finding context. End-to-end automated unrolling inside VC3D is not yet demonstrated.
The repo is MIT licensed; CI builds the wheel and reviewer container; exact-source hashes, create-only artifacts, provenance manifests, images, scale-bar rules, and package validation are explicit. The missing piece is reproducing a qualifying full-scroll result.
Pre-registered and replicated: PHercParis4 and PHerc0139 fibers sit on dense CT material and run within the sheet plane (median about 3.4° out of plane vs 30° random). Wrong-frame controls passed.
Re-derived from the pre-registered v2 scores, PHerc1447 is the only robust Grand Prize frontier member and PHerc0800 the only robust First Letters member. The segment-free slot is a band within noise.
Pre-registered and SUPPORTED with controls passing: 101 / 502 fallback spans vs 2 / 5,001 native spans carry a gap. Only 2 of 388 cues sit in native trace spans.
136 / 136 fibers, 388 gaps and 519 sharp turns. All 388 gaps fall in fibers that contain fallback-interpolated spans; the 57 fully native fibers have none. Not yet a causal claim.
scroliq-eval provides the fail-closed envelope; scroliq-segmentation-validate adds bidirectional TIFXYZ coverage, exact-distance diagnostics, topology gates, and a salted commitment that freezes private truth before scoring.
scroliq-eval is now an installed evaluation harness, not scaffolding: versioned model/data/results contracts, checkpoint and code hashing, overlap blocking, deterministic bootstrap CIs, explicit failed regions, and a two-trust-domain blind-evaluation design.The highest-leverage next move is therefore not another dashboard metric. It is an exact-volume physical surface campaign that produces held-out surface predictions, validates them against truth without leakage, and then scales the successful geometry into complete recto coverage.
The repository Dockerfile is exercised by CI with revision labels, non-root execution, installed CLI smoke tests, and pinned dependencies.
Dockerfile →The current strategy is explicitly to produce one independently reproducible held-out result on one exact prize-eligible volume before broadening claims.
Grand Prize proof campaign →Model cards, dataset manifests, per-region outputs, leakage blocking, checkpoint hashes, CIs, and the trusted/untrusted evaluation boundary are documented.
Model evaluation protocol →Ray-order and patch-attachment results are frozen with commands, input identities, controls, and review queues.
Winding attachment artifacts →The PHerc0139 reference campaign freezes geometry-only probes and source provenance before any sheetness result is read.
Sheetness preregistration →zarr-pyramid-audit's byte-verification artifact demonstrates that the live shard CRC mismatches are not an implementation-specific false alarm.
CRC byte verification →Grand Prize package, image, legibility, and provenance contracts define how an eventual qualifying result stays tied to exact source geometry and CT data.
Reviewer map →MIT-licensed implementation, tests, dated evidence, command-line interfaces, and current campaign state live in one public repository.
ScrolIQ on GitHub →ScrolIQ delegates Zarr structural integrity to zarr-pyramid-audit instead of silently assuming that published volume bytes are trustworthy.
zarr-pyramid-audit →