# Frozen v1 protocol — 2026-10-27

Recorded before the first benchmark run. 20 seeds: 101–120. 1200 ordinary predictions per seed and preset. Presets: directional, rotations, identical, rapid, expensive. Default observation SD 0.12 per output coordinate; query price 0.025 per extra action; hidden switch probability 0.035 per ordinary round, 0.30 for rapid. Expensive price 0.5. All seeds retained, no parameter search. UI runs use 600 rounds and seed 42; those are demonstrations, not the aggregate benchmark.

Four records trained independently with 512 noisy samples each, normalized LMS. Training IDs are supplied, evaluation IDs are hidden. Bank is frozen and identical across arms. Training noise SD 0.02. Random global input/output rotations prevent a fixed axis from always being special. In identical control all ground-truth operators coincide; separate calibration noise can leave tiny learned differences.

Compare passive feedback, fixed probe, random probe, active probe. Bayesian state belief, known noise and known switch probability are shared assumptions. At each round: transition prior; observe ordinary x; optionally query; predict using posterior-weighted record mixture; score against noiseless expected output; then receive noisy ordinary feedback. Every query is an extra noisy observation; state cannot switch within this pair of actions. A query's benefit is allowed on the current ordinary prediction, never on its own probe outcome. The fixed probe is selected from calibration with uniform weights; active selection uses current weights. Both use 16 unit directions on [0,π). Random selection uses this same bank. Query gate is current-output model variance > price. It does not use the true rule or future observation.

Primary loss = (sum of per-round, per-coordinate squared prediction error + price × query count) / ordinary rounds. Also report raw MSE, query count, total actions, MAP routing accuracy (diagnostic only; meaningless for identical rules), actual switches, training error and candidate evaluations. CPU time and storage are not matched. Query price models an imposed cost/latency preference, not measured seconds. All methods receive ordinary feedback, including the passive baseline.

Predeclared descriptive success criterion for directional preset: active mean charged loss at least 5% below EACH of passive, fixed and random, with paired wins over each in ≥15/20 seeds. Failing closes the claimed advantage for this preset/protocol. Report mean, sample SD and paired differences; this is not a significance test or proof of general superiority. Rotation control checks exact direction-independent disagreement algebraically. Identical control should buy essentially no queries at default price. Rapid/expensive cases measure boundaries; no success requirement. Stronger online model learning, unknown hazards, drift, probe back-action and natural data are outside v1.