ScrolIQ in November

N1Close O1: publish Mesh IQ flag precision

Why. 77 of 307 segments are flagged, but no one has yet measured how many flags are real.

Starting point. Blinded 54-segment sample and key-free packet (mesh-review-sample), protocol frozen in mesh-review-protocol.md.

Deliverable. Returned reviewer labels, scored by bin/mesh_review_score.py as written in the protocol.

Done when. Flag precision and the clean-segment defect rate are committed with 95% intervals. If no review comes back by Nov 21, publish that as the status and keep the rate unknown.

N2Two surface arms on one exact volume, scored blind on the same frozen truth

Why. This is the only goal that moves the bottleneck sentence. The frozen PHerc0139 Phase-A sheetness arm returned FAIL (10 / 32 groups), so one more run of the same idea is not enough. Two independent arms against one truth commitment tell us which direction works.

Starting point. Phase-A FAIL (sheetness-run), the blind TIFXYZ scorer scroliq-segmentation-validate, the target-freeze gate, and the CT fiber-fingerprint development protocol.

Deliverable. One exact prize-eligible volume. One salted truth commitment frozen before any prediction exists. At least two pre-registered surface arms, for example a learned surface predictor and a physical descriptor arm, each scored blind per region with bidirectional coverage and topology gates. Stretch: recto-coverage accounting (scroliq-recto-coverage) over one contiguous multi-column region of the winning arm.

Done when. Both arms have committed verdicts against the same truth, pass or fail, with failed regions listed. A failure does not license retuning under the same identity.

N3Blind evaluations of three community models on one held-out split

Why. One evaluation is a demo. Three evaluations on the same split are a comparison people can act on.

Starting point. scroliq-eval contracts, overlap blocking, bootstrap CIs and the two-trust-domain design (model-evaluation.md). Zero published evaluations.

Deliverable. At least three public models covering surface and ink. They are evaluated on one declared held-out split with model cards, checkpoint hashes and overlap checks, and published as a comparison table with CIs and failed regions.

Done when. At least two evaluations are committed, with the third either committed or explicitly declined by its owner. Ask each model owner before publishing.

N4Fiber IQ: which sheet, and a gap model that works

Why. The old N4 was done in October (O11). Fibers on two scrolls sit on CT material and run within the sheet plane. The simple span-aware gap rule failed on independent data.

Deliverable. (a) Sheet identity: test whether consecutive points of a fiber stay on the same CT sheet. Use a pre-registered sheet-tracking check with a planted sheet-switch control. (b) A gap model that works: a pre-registered gap detector that uses CT (for example, material continuity along the step) rather than step length alone, evaluated on both scrolls. (c) Fiber-to-patch consistency against the verified PHercParis4 patches, once their frame is shown to match.

Done when. (a) and (b) have committed verdicts on both scrolls. (c) is done or explicitly unknown with the frame evidence.

N5Close O3: score a spiral fit on constraints it never saw

Why. A fit's own satisfaction score is not a held-out score.

Starting point. The held-out geometry evaluator (heldout-geometry-evaluation.md) and the consumer-GPU gate (spiral-consumer-gpu-gate.md).

Deliverable / done when. One real fit run with declared withheld collections, its config and checkpoint hashes, and the held-out score committed.

N6Exact-volume surface-support survey of all 13 Grand Prize volumes

Why. The v2 frontier (O8) leaves the segment-free slot as a band within noise. A third, independent axis is the honest way to break the tie, and a blended score is not.

Starting point. scroliq-support (exact-scan prediction CT support with bootstrap CIs) and the First Letters support campaign; the October frontier has only two axes.

Deliverable. A new dated campaign that measures surface-prediction CT support on each exact eligible volume where a released prediction exists, and reports the three-axis frontier as a sensitivity analysis. A volume without a released prediction stays unknown.

Done when. Every one of the 13 targets is measured or listed with its reason, and README and docs are updated together.

N7Upstream: from drafts to filed reports, and the dl regression

Why. A confirmed data defect helps nobody while it is only a draft.

Starting point. Unfiled drafts for PHerc0814, PHerc1447 and PHerc0343P (review-queues.md). The zarr-pyramid-audit G1 dl.ash2txt.org regression is still pending.

Deliverable. After the maintainer reproduces each defect and approves it, file the report. Run the dl regression.

Done when. Each draft is either filed and linked here or explicitly withdrawn with a reason, and the dl regression is committed.

N8End-to-end submission dry run on one volume

Why. A qualifying result would arrive with no time left to discover packaging failures.

Deliverable. Take the best non-qualifying N2 output through the whole submission path: the container build, mesh and image package, provenance manifest, scale bars and the package validator. Record every failure.

Done when. The package validates, or each blocking defect is filed as an issue in this repo with the failing check named.

N9Outside evidence of use

Why. The progress prize rewards tools that are used. So far every reviewer of ScrolIQ output has been us.

Deliverable. At least two external reviewers return labels on a ScrolIQ queue (N1 or the fiber/winding VC3D files), and at least one approved upstream report is filed.

Done when. Each item is linked from the progress board, or is listed as not achieved.