From d289a46293e0b422c195746ad8b310d4b50ded23 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Thu, 23 Jul 2026 08:12:22 -0500 Subject: protocol: freeze calibration-split oral-B-v2 recovery --- ORAL_B_V2_CALIBRATED_RECOVERY.md | 95 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 95 insertions(+) create mode 100644 ORAL_B_V2_CALIBRATED_RECOVERY.md (limited to 'ORAL_B_V2_CALIBRATED_RECOVERY.md') diff --git a/ORAL_B_V2_CALIBRATED_RECOVERY.md b/ORAL_B_V2_CALIBRATED_RECOVERY.md new file mode 100644 index 0000000..c2187fc --- /dev/null +++ b/ORAL_B_V2_CALIBRATED_RECOVERY.md @@ -0,0 +1,95 @@ +# Oral-B-v2 calibrated recovery: label-free psychometric assay + +This is a third and final development branch before the still-untouched +confirmation task seeds `30--35`. It preserves three prior records: the old +temporal-difference R2 failed its biological prediction gates; the first v2 +development grid failed from a cold-start; and the dense-velocity, +fixed-target recovery passed every requirement on task seeds 23 and 24 but +failed only the class-balance requirement on seed 25. Seed 25 achieved 100% +ordinary task success, 100% terminal outcome decoding, and passed the other +17/18 frozen checks, but also succeeded on 98.96% of fixed target-ladder +trials. Its failure gate remains at +`results/bci_v2_recovery_dev_gate.json`. + +The algorithm is now frozen without change: forward rate `0.1`, gamma `0.8`, +critic rate `0.03`, velocity reward scale `1.0`, terminal reward `1.0`, +eligibility decay `0.8`, and all dynamics, local updates, warmup, six +conditions, costs, and full-target performance tests from +`ORAL_B_V2_RECOVERY.md`. + +Only the psychometric assay changes. A fixed absolute target cannot be a +model-independent difficulty because independently trained policies have +different cursor scales. The calibrated assay uses a separate, outcome-free +split: + +1. Freeze the trained model and disable all learning, critic feedback, and + terminal outcome feedback. +2. Run 512 calibration episodes with unreachable target `2.1`, so every + trajectory lasts 28 steps. +3. Compute each episode's maximum cursor and take its fixed quantiles + `{0.20, 0.35, 0.50, 0.65, 0.80}` as five target levels. +4. Discard the calibration trajectories. On independent environment seeds, + run 128 full-horizon episodes at every calibrated target under intact, + acute-critic-lesion, and acute-outcome-lesion readouts. + +No calibration success/failure label exists or is consulted, no evaluation +outcome selects or reweights a target, and calibration/evaluation seeds are +disjoint. This is analogous to calibrating a psychometric stimulus range on +neural output, then measuring rewarded versus censored outcomes on new +trials. Calibration examples, state transitions, selected target values, and +both seed sets are reported. + +## Frozen development validation + +R1 uses previously untouched task seeds `26--28`, model seed 0, full +performance seed `540000 + task_seed`, calibration seed +`550000 + task_seed`, and target-assay seed +`560000 + 1000 task_seed + k` for zero-based quantile index `k`. There is no +hyperparameter selection and exactly three records. + +The 18 per-seed gates are exactly those in the failed fixed-target recovery: +final success at least 70%; gain at least 10 points; fixed-role gap at least +20 points; oracle deficit at most 10 points; plasticity lesion no more than +half gain; role cosine at least `0.80`; pooled outcome fraction in +`[0.15,0.85]`; terminal balanced accuracy at least 75%; role separation and +acute outcome-lesion drop each at least `0.20`; critic expectedness at least +`0.05` and critic/value correlation at least `0.95`; residual--soma +correlation at most `0.10`; raw-minus-residual gap at least `0.20`; +surrounding accuracy at least 52%; decoder correlation at least `0.02`; sign +inversion at least `0.01`; and velocity advantage at least `0.05`. + +Any failure closes the branch and leaves confirmation untouched. Development +cannot change the score. + +## Frozen untouched confirmation + +Only a complete R1 pass permits task seeds `30--35` and model seeds `0--4`, +exactly 30 records with no further choice. Performance seed is +`590000 + task_seed`, calibration seed `600000 + task_seed`, and independent +target-assay seed `610000 + 1000 task_seed + k`. The task seed remains the +uncertainty unit: average the five models per task, then use one-sided 95% +Student-t bounds with five degrees of freedom. + +All R2 thresholds remain exactly those frozen for the fixed-target recovery: + +- learning: mean final at least 70%, every task mean and lower bound 60%; + mean gain at least 10 points and lower bound 5; fixed-role gap mean 20 and + lower bound 10; oracle deficit mean at most 10 and upper bound 20; + nonnegative plasticity half-margin lower bound; role cosine mean `0.80` and + every record `0.70`; +- innovation: residual--soma mean at most `0.10` and upper bound `0.12`; + raw-minus-residual mean at least `0.20` and lower bound `0.15`; + surrounding accuracy mean 52% and lower bound 50%; decoder correlation mean + `0.05` and nonnegative lower bound; positive sign in 25/30 and every task; + velocity advantage mean `0.05` and nonnegative lower bound; +- outcome: every pooled fraction in `[0.10,0.90]`; terminal accuracy mean 80% + and lower bound 75%; separation mean `0.20` and lower bound `0.15`; acute + outcome-lesion drop mean `0.20` and lower bound `0.15`; critic expectedness + mean `0.05` and lower bound `0.02`; every critic/value correlation `0.95`. + +Records require clean tracked source, exact filenames, complete cost audit, +and SHA-256 binding to every transitive runner/analyzer, dynamics and metrics +module, this protocol, D4, all prior failure gates, and the R1 gate. A complete +R2 pass raises the formal milestone from 7 to 8 and permits only a separately +frozen oral-A-v2 protocol. It does not erase prior failures, open the old +oral-A gate, or establish online apical control. -- cgit v1.2.3