snowy localized panel holdout

Labnote 015: snowy localized panel holdout🔗

Question🔗

Does deterministic localized candidate bc155807a28aee1a2d7d10fb18a96b24d40b9fe300fc3e01650f70e5d61cc649 generalize to a fresh balanced holdout captured during the snowy Claw of Jormag encounter?

Method🔗

• Holdout: 16 newly captured, structurally masked observations balanced across

Inventory-only, Hero-plus-Inventory, Hero-only, and closed-panel states.

• Labels: bound from operator-staged states before transfer; recognition did

not run during capture.

• Holdout digest:

dbb09c405d8da0492a44f8755ff6bebbc349128191a235f4e4d25790171257a1.

• Threshold: candidate's frozen 0.65 threshold.
• Training and calibration: disabled.
• Gate: at least 87.5% aggregate exact accuracy, at least 75% exact accuracy in

every state, and no class above a 12.5% aggregate error rate.

• Accelerator: RTX 4070 through a non-preemptive accelerator lease.
• Internal execution record retained; identifier withheld.

Result🔗

Aggregate exact accuracy was 14/16 (87.5%). Inventory classification was perfect with 8 true positives, 8 true negatives, and no errors. Hero had 6 true positives, 8 true negatives, 2 misses, and no false positives. Its 12.5% class error rate met the ceiling, but the 50% Hero-only state accuracy failed the frozen 75% state floor. The promotion gate therefore failed.

The evaluator performed no training or calibration, emitted no sample-level predictions, completed on CUDA in 3.6 seconds with exit code 0 and no stderr, and released the accelerator lease normally.

Decision🔗

Do not promote the candidate. The snowy holdout is spent and must not be used for tuning or repeat evaluation.

Inventory position generalization is now supported by fresh evidence. The next development direction is Hero-only evidence across independently collected bright and snowy environments. Any successor must pass the development gate and a different fresh holdout.