1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
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.
|
