What the prize actually requiresOfficial rules checked 2026-10-02; the Challenge page remains authoritative.
The 2027 Grand Prize requires a reproducible open-source pipeline, seamless VC3D integration, 100% recto unrolling, column-scale TIFXYZ meshes and images, legible ink, strict train/prediction separation, reproducibility records, false-positive mitigation, and held-out validation. The deadline is June 25, 2027.
Read the official Grand Prize rules →
ScrolIQ's role: make each stage auditable and volume-bound so weak evidence cannot silently become a readiness claim. It does not currently claim a qualifying full-scroll result.
Reviewer readiness matrix“Implemented” means the evidence mechanism exists and is committed. It does not mean the Grand Prize requirement is satisfied for a final scroll unless stated.
| Grand Prize requirement | Status | What ScrolIQ currently has | What still has to be shown |
|---|---|---|---|
| Permissive open-source code and reproducible pipeline | IMPLEMENTED CORE + PACKAGER | MIT-licensed public repository, deterministic CLI surfaces, exact-input provenance, immutable artifact campaigns, machine-readable reports, and a deterministic scroliq-package reviewer ZIP with SHA-256 inventory and self-verification. | A final end-to-end unrolling runner and a frozen reproduction package from a qualifying scroll still have to be demonstrated. |
| Use the exact eligible CT volume | IMPLEMENTED GUARDRAIL | scroliq-manifest and scrollq-grand-prize match exact volume IDs and reject same-scroll substitutions, including higher-resolution scans. | The final chosen target still needs a complete downstream evidence chain on that exact volume. |
| Seamless VC3D integration | PARTIAL + REAL HANDOFF | Native TIFXYZ, VC3D fiber JSON, winding PointCollections, and XYZ review coordinates are accepted or emitted across evidence layers. Mesh IQ now emits native VC3D PointCollections directly; the frozen PHerc0139 campaign produces 35 ranked review sites (15 edge jumps + top 20 normal reversals) on SHA-pinned public TIFXYZ bytes. | The PointCollections handoff removes coordinate translation, but a fully demonstrated in-VC3D end-to-end Grand Prize workflow still needs to be shown. |
| 100% recto surface unrolled | NOT SHOWN | Recto-coverage and geometry evidence tooling can describe coverage and reject cross-volume evidence. | A qualifying whole-scroll surface whose accepted meshes cover the required recto area, including required detached patches. |
| TIFXYZ meshes with low-distortion 2D parametrization, column-scale packaging | PARTIAL + MEASURED | scroliq-mesh audits native TIFXYZ structure/provenance and selected geometry checks. An exact-volume cross-cut finds published meshes for 2 / 13 eligible targets: PHerc0800 has 6 / 6 locally clean meshes; PHerc1447 has 15 / 15 with an enclosed-hole review finding, including 1 multi-defect positive-control segment. An independent same-byte cross-check against TIFXYZ Doctor agrees 10 / 10 on hole presence and 10 / 10 on single-component presence, including 7 / 7 hole-presence agreement on its PHerc0800 + PHerc1447 cases; that benchmark also exposed and verified the fix for a Mesh IQ isometry-normalization error. A separate binding-aware PHerc0139 dossier proves coordinate-exact identity between ScrolIQ and Windcheck on one published original, joining local Mesh IQ findings to an independent 3,333-contact nonlocal census without upgrading the evidence to semantic-exact or official-VC3D status. | These are fragment/local checks, not whole-scroll proof. Final per-column meshes, physical sheet identity, sequential naming, flattening quality evidence, and full-scroll union coverage remain required. TIFXYZ Doctor benchmark roles are not geometry ground truth. |
| Visible text and ≥70% preserved-character legibility for counted columns | AUDIT CONTRACT READY | scroliq-legibility now binds a reviewer ledger to exact submitted render/mesh hashes, requires one visible-character reading plus pixel bounds and explicit interpolated: false for each counted character, keeps uncertain preserved characters in the denominator, applies 70% per counted column, and requires documented Challenge acknowledgement for exclusions. | A qualifying full-scroll render set still needs real letter-by-letter reviewer evidence. The software audit checks the ledger; it does not decide papyrological correctness. |
| No train/prediction overlap | AUDIT EXISTS | scroliq-ink-audit fails closed on declared spatial overlap and records split, checkpoint, seed, and held-out provenance. | The final ink model must provide complete train/evaluation region declarations that pass the audit. |
| Public ML datasets/checkpoints with required licensing | NOT YET PACKAGED | Provenance fields exist for linking checkpoints and evidence runs. | Every training dataset, pseudo-label stage, checkpoint, and license required by the rules must be published and indexed. |
| Fixed random seeds and public experiment tracking | PARTIAL | ScrolIQ audits seed/checkpoint declarations and keeps deterministic analysis artifacts. | Final training and inference runs need complete public experiment-tracking records, not just metadata checks. |
| Fully automated unrolling; ≤8 documented human hours | ACCOUNTING CONTRACT READY | The deterministic package builder requires a machine-readable human-input ledger, rejects totals above eight hours, and requires the ledger to reconcile exactly to provenance. | An actual end-to-end automated unrolling run must still produce a qualifying surface while staying within that documented budget. |
| Submission images, 1 cm scale bars, full-scroll numbered banner | GENERATOR + PROOF READY | scroliq-submission-image deterministically adds a physically derived 1 cm scale bar, emits hash-pinned scale proofs, builds a sequential numbered banner, and schema-v6 provenance binds those proof sidecars to the exact render/banner bytes. The final packager includes them. | The same machinery still has to be run on every qualifying full-scroll column render. |
| Docker image and system requirements | PACKAGE CONTRACT READY | scroliq-package requires non-empty system-requirements documentation and a copy-paste reproduction command using the exact digest-pinned Docker image already bound in provenance. | The final unrolling pipeline's Docker image still has to be built, published, and demonstrated on the documented hardware. |
| False-positive mitigation | PROTOCOL READY | The Grand Prize probe protocol requires correct-surface checks, normal offsets, adjacent winding, perturbed geometry, and an independent checkpoint/fold. | Measured results from those controls on the final ink candidate. |
| Held-out validation on known ground truth | PROTOCOL READY | Held-out geometry and provenance contracts are predeclared; failed regions remain in denominators. | Published quantitative held-out results for the final method, including ink validation where trained models are used. |
Evidence already worth a reviewer’s time
Exact-volume discipline
Grand Prize manifests are keyed to the exact official volume IDs; same-scroll substitutions are not accepted as evidence.
Blind geometry protocol
First-wave targets use deterministic fit/held-out regions so geometry can be tested before committing to whole-scroll processing.
Mesh evidence contracts
TIFXYZ audits separate structural validity from stronger claims such as correct sheet identity.
Ink falsification
Ink evidence is treated as something to falsify, not something to “score up” until it looks convincing.
VC3D-native inputs
Fiber and winding audits preserve exact-file provenance while reading formats already used by the Challenge tooling.
Unknown stays unknown
Passports exclude cross-volume artifacts and keep unmeasured downstream stages explicit rather than inferring readiness from scan quality.
Shortest path from diagnostics to a qualifying package
- Choose one exact eligible volume using blind geometry evidence, not a blended “best scroll” score. Keep failed regions in the denominator.
- Close the surface chain first: winding → spiral/geometry → TIFXYZ → recto coverage, with exact-volume bindings at every handoff.
- Make VC3D the reviewer path, not an export target. The final workflow should open, inspect, and reproduce the same evidence in VC3D with minimal translation.
- Only then spend large ink-model compute. Require train/prediction exclusion certificates, fixed seeds, public tracking, and falsification controls before counting any letters.
- Automate submission packaging early. Generate sequential column meshes, matching images, 1 cm scale bars, and the full-scroll numbered banner from the same manifest.
- Ship the reproduction environment before the result is “perfect.” Docker, system requirements, licenses, exact commands, and human-time accounting should be continuously testable.
Current claim boundary
What this page does not claim: that ScrolIQ predicts readability, that a high scan-health score implies a good surface, that a passing mesh audit proves sheet identity, or that an ink prediction is real because its provenance is clean. Those are separate empirical questions and remain separate on purpose.
Return to the live ScrolIQ dashboard → · October evidence plan → · Source repository →