Repository navigation
Helena Framework 0.21.0-alpha
Pre-releaseA scroll is one sheet wound outward. Two laminae cannot occupy the same place, and
along any ray the wraps appear in radial order. That is a statement about a pair
of surfaces, which certification of each one separately structurally cannot reach.
The winding invariant
framework/contracts/winding.py reduces a surface to where it sits in the roll —
a median radius, angle and height about the fitted scroll axis — and answers whether
two surfaces can both be laminae of the same roll.
Calibrated on 37 reconstructed wraps of PHerc826 before it was written:
| measured | |
|---|---|
| fitted radius against wrap number | Spearman ρ +0.9993, 35/36 monotone |
| spacing between consecutive wraps | 17 voxels = 0.14 mm, which is papyrus |
| comparable pairs behaving, of 666 | 664 |
The wrap numbers come from the reconstruction, not from geometry, so they are a label
to test against rather than an input. Nothing put sheet thickness into the calculation;
it came out of it.
The check returns WINDING_UNDETERMINED rather than a boolean below three voxels of
separation, because at that distance the ordering is not in the data — on ground truth
it inverted on 35–43% of rays there, and 95–99.6% correct where wraps are 15–25 voxels
apart. Reporting confidently in the ambiguous band would be inventing.
It is deliberately not a gate. It answers a different question from geometry
certification, CT support, model response and human review, so it stands beside them as
a fifth judgement rather than a veto over any of them.
The P8 merge looks at pairs it was not told about
The seam check only ever iterated the edges it was handed, so a collision between two
surfaces nobody declared adjacent passed unseen. COLLISION_QC.json now reports every
pair that meets, declared or not.
Segmentation can grow m7's alternatives
A run that took only the top proposal was discarding candidates that turned out to be
different wraps — m7's second choice differed in seven of eight comparable pairs, five
of them within a hair of that same 0.14 mm spacing. candidate_rank and
reconsider_covered make those reachable, and the seed probe now reaches twenty
candidates where it stopped at three.
Deployment
HELENA_REGISTRY no longer carries an address. Unset, containers/deploy-platform.sh
builds the panel from the checkout; a deployment that wants image pulls names its own
registry.
Alpha. The winding invariant reports evidence and nothing gates on it yet.