1 files changed, 128 insertions, 0 deletions
diff --git a/ORAL_B_V2_RECOVERY.md b/ORAL_B_V2_RECOVERY.md
new file mode 100644
index 0000000..d929755
--- /dev/null
+++ b/ORAL_B_V2_RECOVERY.md
@@ -0,0 +1,128 @@
+# Oral-B-v2 cold-start recovery: dense innovation plus target psychometrics
+
+This branch begins only after the complete, frozen
+`oral_b_v2_development_v1` grid failed. That failure is retained at
+`results/bci_v2_dev_gate.json`: none of 24 records was eligible, no candidate
+exceeded 0.78% full-horizon evaluation success, and confirmation task seeds
+`30--35` were not touched. The recovery does not edit that protocol, result,
+or gate.
+
+## Development-only diagnosis
+
+The failure localized cleanly. Learned-role cosine was about `0.995`, the
+neutral residual remained decorrelated from soma, and the surrounding-state
+and velocity-over-error diagnostics passed, but even the exact-role oracle
+could not learn. Before the first success, the critic can observe only the
+dense performance-change term. Frozen v2 scaled that term by `0.25`, so the
+largest actor rate was effectively one quarter of the already validated
+temporal-difference rule and the terminal reward was unreachable.
+
+On already used development task seed 20, a causal postmortem changed only
+the dense velocity scale from `0.25` to `1.0`: final success changed from
+about 1% to 100%. With critic enabled, success rose sharply on days 7--8;
+without critic it did not rise until days 13--14. At scale 1, the diagnostic
+also recovered learned-role cosine `0.983`, residual--soma correlation
+`0.058`, surrounding decoder `54.4%`, velocity advantage `0.643`, and an
+acute-outcome-lesion loss of `0.386` in role-aligned terminal separation.
+These are exploratory postmortem values and are not R1 or R2 evidence.
+
+That postmortem exposed a second structural issue: the learned fixed-target
+policy reaches target `0.8` within four steps on every trial, so no
+horizon-only assay can produce an unrewarded class. A development-only
+uncensored probe placed the maximum cursor transition between targets 1.55
+and 1.80. The recovery therefore replaces the failed horizon ladder with a
+fixed target psychometric ladder. This is disclosed development tuning, not
+claimed as preregistration against task seed 20.
+
+## Frozen recovery
+
+The mechanism, local updates, six paired conditions, neutral warmup, cost
+accounting, full-horizon performance assay, and all other parameters remain
+as in `ORAL_B_V2.md`, except:
+
+```text
+velocity_reward_scale = 1.0
+forward_eta            = 0.1
+gamma                  = 0.8
+critic_eta              = 0.03
+```
+
+There is no recovery hyperparameter grid. Endpoint-free tests establish that
+the nonterminal TD and actor update are exactly four times the failed v2
+cold-start drive when critic and terminal reward are absent. They also bind
+the ordered target ladder.
+
+The signature assay uses six targets
+`{1.55, 1.60, 1.65, 1.70, 1.75, 1.80}`, 128 independent full-28-step episodes
+per target, and no target selection or reweighting. All targets are pooled.
+The same trained intact model is replayed under intact, acute critic lesion,
+and acute terminal-outcome lesion. This is a psychometric generalization
+assay: ordinary performance remains measured at the trained target `0.8`.
+The target ladder is intended to create rewarded and censored timeout trials
+without defining outcome directly from a balanced random label.
+
+## R1 independent development validation
+
+R1 uses previously untouched development-validation task seeds `23--25` and
+model seed 0. Full performance seed is `460000 + task_seed`; target-ladder
+seed for zero-based target index `k` is
+`470000 + 1000 task_seed + k`. There is one fixed candidate and exactly three
+records.
+
+Every seed must pass every check:
+
+- intact final success at least 70%, early-to-late gain at least 10 points,
+  fixed-role gap at least 20 points, oracle deficit at most 10 points,
+  plasticity lesion at most half the intact gain, and role cosine at least
+  `0.80`;
+- pooled target-ladder success fraction in `[0.15, 0.85]`;
+- terminal residual outcome balanced accuracy at least 75%, role-aligned
+  outcome separation at least `0.20`, acute outcome lesion separation drop at
+  least `0.20`, critic expectedness contribution at least `0.05`, and paired
+  critic contribution/value correlation at least `0.95`;
+- residual--soma correlation at most `0.10`, raw-minus-residual correlation
+  gap at least `0.20`, surrounding-event accuracy at least 52%, and
+  decoder-distance correlation at least `0.02`; and
+- sign inversion at least `0.01` and velocity-over-error advantage at least
+  `0.05`.
+
+R1 cannot change the score. Any failure closes this recovery and leaves
+confirmation untouched.
+
+## R2 untouched confirmation
+
+Only a complete R1 pass permits the fixed candidate on task seeds `30--35`
+and model seeds `0--4`, exactly 30 records with no replacement or further
+selection. Full performance seed is `520000 + task_seed`; target-ladder seed
+is `530000 + 1000 task_seed + k`. The task seed is the uncertainty unit: the
+five models are averaged within each task, then one-sided 95% Student-t
+bounds use five degrees of freedom.
+
+Learning/plasticity requirements are unchanged from the prior frozen R2:
+mean final at least 70%, every task mean and lower bound at least 60%; mean
+gain at least 10 points and lower bound 5; mean fixed-role gap at least 20
+points and lower bound 10; mean oracle deficit at most 10 points and upper
+bound 20; nonnegative lower bound for
+`0.5 intact_gain - lesion_gain`; mean role cosine at least `0.80` and every
+record at least `0.70`.
+
+Innovation requirements are: mean residual--soma correlation at most `0.10`
+with upper bound `0.12`; raw-minus-residual gap at least `0.20` with lower
+bound `0.15`; surrounding accuracy at least 52% with lower bound 50%;
+decoder-distance correlation at least `0.05` with nonnegative lower bound;
+positive sign in at least 25/30 records and every task mean; velocity
+advantage at least `0.05` with nonnegative lower bound.
+
+Outcome-surprise requirements are: every pooled success fraction in
+`[0.10, 0.90]`; terminal accuracy mean at least 80% and lower bound 75%;
+role-aligned separation mean at least `0.20` and lower bound `0.15`; acute
+outcome-lesion drop mean at least `0.20` and lower bound `0.15`; critic
+expectedness mean at least `0.05` and lower bound `0.02`; every paired
+critic/value correlation at least `0.95`.
+
+All records require a clean tracked tree, exact filenames, and SHA-256 binding
+to both preserved failure gates, the D4 gate, protocol, dynamics, metrics,
+runners, analyzers, and R1 gate where applicable. A complete R2 pass raises
+the formal repository milestone from 7 to 8 and permits a separately frozen
+oral-A-v2 protocol. It does not erase either failure, open the old oral-A
+gate, or establish online apical control.