Boxes0.97

Two parents. Two independent competitions.

The testnet split is 50% base improvement, 20% features and 30% customer adapters, with mini/lite split 80/20 inside both base and box budgets: 40% mini base, 10% lite base, 16% mini boxes, 4% lite boxes, 30% customer adapters. These are this validator’s allocation shares, not guaranteed earnings. Boxes initially receive the whole feature budget; future tracks require an announced reallocation.

No qualified box kings yet. Operator references are the initial scores to beat and earn no crown. Both box allocations burn until qualifying miners are crowned. Historical research and small adapter smoke tests do not qualify a winning starter.

Both models learned boxes; mini led on this test set.

Historical research, not the pilot exam. These adapters used stock mini and lite, with the same 12 defect types and shape generator in training and test; Jev was zero-shot. The mini result is not an r5 baseline. Each model had one training run and one seed, without absent-target or confidence qualification. Lite reached IoU ≥ 0.5 on only 3 of 122 real-defect images, versus mini’s 42. This does not establish general localization quality.

Four serial model passes per box: lite averaged 0.624 s and mini 3.270 s on the RTX 4090 in separate runs; this was not a fresh paired speed comparison. Training took 18.2 / 32.0 minutes and 1.93 / 9.31 GiB peak CUDA allocation for lite / mini.

Example boxes

• True box
• Stock mini + adapter
• Jev 1.14

Chosen to include cases where Jev is better (about one in three test images). Defect photos: VisA, Amazon Science, CC BY 4.0, resized. Shape scenes generated by Zils.

Rules

1. 01

   Train

   Choose the exact mini r5 or lite v1 parent. Train a box adapter with a frozen vision tower. The public synthetic generator is training material; private exam images and seeds stay sealed.

2. 02

   Commit

   Package with python -m zils.feature_miner pack, upload the bundle, then submit with python -m zils.feature_miner commit. The commitment binds the track, parent and adapter. One hotkey has one active commitment.

3. 03

   Score

   On present targets, weight synthetic families equally: gain at least 0.02 mean IoU over the incumbent, with a positive paired 95% lower confidence bound and mean IoU at least 0.40. No family may drop more than 0.02.

4. 04

   Confirm

   Absent-target false positives must be at most 10%; presence Brier may be at most 0.005 worse than the incumbent. The same adapter must pass every gate on two disjoint private draws, evaluated at least 12 hours apart.

5. 05

   Crown

   Reviewed crowning is independent for mini and lite. Among each track’s five most recent distinct kings, those still registered share its rewards equally. Operator baselines earn no crown; an empty track burns its share.

6. 06

   Graduate

   Winning this synthetic pilot does not merge an adapter into the base. Graduation requires a separate review, real-image qualification and the full base regression suite.

Leaderboard

Mini and lite keep separate king histories. Reviewed crowning unlocks a track’s rewards; it does not release a new base or customer box API.

Start with the synthetic pilot guide

The feature-track guide defines the public shape generator, exact parent pins, presence and box format, training limits, packaging and submission commands. Training examples are separate from the configured private pilot draws. Follow the guide for your chosen parent; these are miner submissions.