Why this matters for the Grand PrizeOfficial rules checked 2026-10-02; the Challenge page remains authoritative.
The 2027 Grand Prize requires reviewers to reproduce a complete pipeline, inspect TIFXYZ meshes, verify generated images, and trust that the submitted result is supported by the CT data. Silent all-zero or malformed pyramids can invalidate that chain without producing an obvious reader error.
Official Grand Prize requirements →
Where the integrity gate fitsThe right-hand column is intentionally blunt about what belongs elsewhere in the winning pipeline.
| Grand Prize requirement | zarr-pyramid-audit contribution | Status | Outside this tool |
|---|---|---|---|
| Fully reproducible public pipeline | MIT-licensed CLI, deterministic reports, schemas, fixtures, manifests, and committed campaign artifacts make integrity checks replayable. | IMPLEMENTED | The final Grand Prize pipeline still needs its own end-to-end runner, Docker image, and system requirements. |
| Reliable CT / Zarr inputs | zpa-audit, zpa-gate, and chunk-content probes distinguish malformed metadata, missing levels, header-only levels, unknown evidence, and sampled present-but-empty content. A Grand Prize exact-volume preflight matches all 13 prize-listed CT roots into the frozen S3 audit: 13 / 13 have all six levels present, no chunkless or unknown-presence level, and zero header findings. Separately, a 64-volume live volcomp campaign checked 2,008 shard indexes and independently confirmed 331 CRC32C mismatches as writer non-conformance; the indexes still parse and chunks decode, so the condition is retained as a medium warning rather than mislabeled voxel corruption. | IMPLEMENTED + MEASURED | This is dated structural/input-format evidence, not semantic voxel validation, scan quality, surface identification, or unrolling correctness. |
| VC3D integration | Machine-readable reports and TIFXYZ audits are designed for modular integration; review coordinates and standard formats can feed downstream tooling. | PARTIAL | A seamless in-VC3D user workflow is not demonstrated here. |
| 100% recto unrolled | TIFXYZ files can be structurally audited before acceptance into a larger coverage ledger. | SUPPORTING | Segmentation, sheet identity, flattening, and proof of whole-scroll recto coverage. |
| Per-column TIFXYZ submission meshes | zpa-tifxyz checks native TIFXYZ structure/provenance and optional content-tier consistency. | AUDIT EXISTS | Column boundaries, low-distortion isometry quality, sequential naming, and final package completeness. |
| Visible and legible ink | Integrity gating can prevent known-bad input volumes from entering training or inference. | OUT OF SCOPE | Ink detection, legibility, papyrological evaluation, and line counts. |
| No train/prediction overlap | Can ensure the underlying stores and surface files used by an external split are the expected artifacts. | OUT OF SCOPE | Spatial split definition and leakage auditing belong in ScrolIQ / the ink pipeline. |
| False-positive mitigation | Fail-closed tri-state evidence prevents timeouts/403s/unknown listings from being misreported as absence; direct chunk reads test whether “present” data actually contain voxels. | IMPLEMENTED FOR DATA | Ink hallucination controls and held-out text validation. |
| Held-out validation | Offline fixtures, positive controls, and live known-clean / known-defective controls validate integrity logic. | DOMAIN-SPECIFIC | The Challenge's held-out ink/ground-truth validation must be satisfied by the unrolling/ink method itself. |
Controls that reduce reviewer risk
Unknown is not absent
Timeouts, permission failures, rate limits, and unsupported listings remain UNKNOWN; they cannot silently become a missing-data finding.
Headers are not content
The fast audit checks metadata and chunk presence; the sampled content probe separately decodes real voxels to test a different silent-corruption class.
Fail closed before publish
zpa-gate converts severity policy into a non-zero CI exit code so known defect classes can block a publication step.
Surface files get their own audit
TIFXYZ checks are separate from Zarr checks because a valid CT store does not imply a valid submitted surface.
Claims stay dated
Live-store findings are tied to dated artifacts so “still broken” is never presented as an eternal fact.
Shard checksums are independently verified
Live volcomp CRC32C mismatches were reproduced with a separate bit-by-bit implementation over the exact fetched index bytes, preventing a parser assumption from being promoted into a corruption claim.
Companion boundary
ScrolIQ owns downstream quality/evidence judgments; this repository supplies integrity primitives and does not infer readability.
What a winning pipeline should do with this
Return to the live integrity dashboard → · October evidence plan → · Source repository →