protocol: freeze oral-B-v2 cold-start recovery

11 files changed, 1606 insertions, 2 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.
diff --git a/experiments/analyze_bci_v2_recovery_confirmation.py b/experiments/analyze_bci_v2_recovery_confirmation.py
new file mode 100644
index 0000000..94143be
--- /dev/null
+++ b/experiments/analyze_bci_v2_recovery_confirmation.py
@@ -0,0 +1,498 @@
+#!/usr/bin/env python3
+"""Audit untouched oral-B-v2 cold-start recovery confirmation."""
+import argparse
+import glob
+import hashlib
+import json
+import math
+import os
+import statistics
+
+from experiments.bci_v2_recovery_confirmation import (
+    MODEL_SEEDS,
+    TASK_SEEDS,
+)
+from experiments.bci_v2_recovery_run import (
+    CHALLENGE_EPISODES,
+    FIXED_CONFIG,
+    TARGETS,
+    build_recovery_config,
+    source_paths,
+)
+from experiments.bci_v2_run import CONDITIONS
+
+
+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+RUNNER_PATH = os.path.join(
+    ROOT, "experiments", "bci_v2_recovery_confirmation.py"
+)
+T_CRITICAL_ONE_SIDED_95_DF5 = 2.015048373
+
+
+def require(condition, message):
+    if not condition:
+        raise ValueError(message)
+
+
+def sha256(path):
+    digest = hashlib.sha256()
+    with open(path, "rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def finite_tree(value):
+    if isinstance(value, dict):
+        return all(finite_tree(item) for item in value.values())
+    if isinstance(value, (list, tuple)):
+        return all(finite_tree(item) for item in value)
+    if isinstance(value, (int, float)):
+        return math.isfinite(value)
+    return True
+
+
+def lower_bound(values):
+    return (
+        statistics.mean(values)
+        - T_CRITICAL_ONE_SIDED_95_DF5
+        * statistics.stdev(values)
+        / math.sqrt(len(values))
+    )
+
+
+def upper_bound(values):
+    return (
+        statistics.mean(values)
+        + T_CRITICAL_ONE_SIDED_95_DF5
+        * statistics.stdev(values)
+        / math.sqrt(len(values))
+    )
+
+
+def summarize(values):
+    return {
+        "by_task_seed": values,
+        "mean": statistics.mean(values),
+        "one_sided_95pct_lower": lower_bound(values),
+        "one_sided_95pct_upper": upper_bound(values),
+    }
+
+
+def task_clusters(records, getter):
+    return [
+        statistics.mean(
+            getter(records[(task_seed, model_seed)])
+            for model_seed in MODEL_SEEDS
+        )
+        for task_seed in TASK_SEEDS
+    ]
+
+
+def confirmation_paths(r1_gate):
+    paths = source_paths()
+    paths["development_runner"] = paths["runner"]
+    paths["runner"] = RUNNER_PATH
+    paths["r1_gate"] = os.path.abspath(r1_gate)
+    return paths
+
+
+def validate(row, path, digests):
+    require(row.get("schema_version") == 3, f"{path}: schema")
+    args = row.get("args", {})
+    task_seed = args.get("task_seed")
+    model_seed = args.get("model_seed")
+    require(
+        task_seed in TASK_SEEDS and model_seed in MODEL_SEEDS,
+        f"{path}: seeds",
+    )
+    protocol = row.get("protocol", {})
+    require(
+        protocol.get("name")
+        == "oral_b_v2_cold_start_recovery_confirmation_v1"
+        and protocol.get("split") == "untouched_confirmation"
+        and protocol.get("training_task_seed") == task_seed
+        and protocol.get("model_seed") == model_seed
+        and protocol.get("fixed_config") == FIXED_CONFIG
+        and protocol.get("no_further_selection") is True
+        and protocol.get("confirmation_grid_size") == 30
+        and protocol.get("protocol_sha256") == digests["protocol"]
+        and protocol.get("r1_gate_sha256") == digests["r1_gate"],
+        f"{path}: protocol",
+    )
+    provenance = row.get("provenance", {})
+    require(
+        provenance.get("git_tracked_dirty") is False
+        and provenance.get("tracked_inputs")
+        and all(provenance["tracked_inputs"].values())
+        and set(provenance["tracked_inputs"]) == set(digests)
+        and provenance.get("input_sha256") == digests,
+        f"{path}: provenance",
+    )
+    require(
+        row.get("config") == vars(build_recovery_config()),
+        f"{path}: config",
+    )
+    require(
+        row.get("split") == "untouched_confirmation"
+        and row.get("finite") is True
+        and finite_tree(row),
+        f"{path}: finite/split",
+    )
+    require(
+        set(row.get("conditions", {})) == set(CONDITIONS),
+        f"{path}: conditions",
+    )
+    for name in CONDITIONS:
+        warmup = row["warmup"][name]
+        condition = row["conditions"][name]
+        cost = condition["cost"]
+        require(
+            warmup["batches"] == 100
+            and warmup["examples"] == 6400
+            and warmup["instruction_present"] is False
+            and warmup["role_cursor_scalar_observations"] == 12800
+            and warmup["predictor_max_abs_error"] <= 1e-5
+            and len(condition["daily_success"]) == 14
+            and cost["maximum_state_episode_steps"] == 25088
+            and cost["cursor_scalar_observations"]
+            == 2 * cost["role_probe_examples"]
+            and cost["task_loss_queries"] == 0
+            and cost["reverse_mode_calls"] == 0,
+            f"{path}: condition {name}",
+        )
+    challenge = row["assays"]["challenge"]
+    require(
+        challenge["targets"] == list(TARGETS)
+        and challenge["episodes_per_target"] == CHALLENGE_EPISODES
+        and challenge["selection_over_targets"] is False
+        and challenge["maximum_steps_per_episode"] == 28
+        and row["signatures"]["challenge_episodes"]
+        == len(TARGETS) * CHALLENGE_EPISODES,
+        f"{path}: target ladder",
+    )
+
+
+def checks(metrics, clustered, all_values):
+    return {
+        "all_records_finite_paired_and_cost_audited": True,
+        "learning_and_plasticity": {
+            "mean_intact_final_at_least_0p70":
+                metrics["intact_final"]["mean"] >= 0.70,
+            "every_task_mean_final_at_least_0p60":
+                min(clustered["intact_final"]) >= 0.60,
+            "intact_final_lower_bound_at_least_0p60":
+                metrics["intact_final"][
+                    "one_sided_95pct_lower"
+                ] >= 0.60,
+            "mean_learning_gain_at_least_0p10":
+                metrics["intact_gain"]["mean"] >= 0.10,
+            "learning_gain_lower_bound_at_least_0p05":
+                metrics["intact_gain"][
+                    "one_sided_95pct_lower"
+                ] >= 0.05,
+            "mean_fixed_role_gap_at_least_0p20":
+                metrics["fixed_role_gap"]["mean"] >= 0.20,
+            "fixed_role_gap_lower_bound_at_least_0p10":
+                metrics["fixed_role_gap"][
+                    "one_sided_95pct_lower"
+                ] >= 0.10,
+            "mean_oracle_deficit_at_most_0p10":
+                metrics["oracle_deficit"]["mean"] <= 0.10,
+            "oracle_deficit_upper_bound_at_most_0p20":
+                metrics["oracle_deficit"][
+                    "one_sided_95pct_upper"
+                ] <= 0.20,
+            "plasticity_half_margin_lower_bound_nonnegative":
+                metrics["plasticity_half_margin"][
+                    "one_sided_95pct_lower"
+                ] >= 0.0,
+            "mean_role_cosine_at_least_0p80":
+                metrics["role_cosine"]["mean"] >= 0.80,
+            "every_role_cosine_at_least_0p70":
+                min(all_values["role_cosine"]) >= 0.70,
+        },
+        "innovation_and_network_prediction": {
+            "mean_residual_soma_corr_at_most_0p10":
+                metrics["residual_soma_corr"]["mean"] <= 0.10,
+            "residual_soma_corr_upper_bound_at_most_0p12":
+                metrics["residual_soma_corr"][
+                    "one_sided_95pct_upper"
+                ] <= 0.12,
+            "mean_raw_residual_corr_gap_at_least_0p20":
+                metrics["raw_residual_corr_gap"]["mean"] >= 0.20,
+            "raw_residual_corr_gap_lower_bound_at_least_0p15":
+                metrics["raw_residual_corr_gap"][
+                    "one_sided_95pct_lower"
+                ] >= 0.15,
+            "mean_surrounding_accuracy_at_least_0p52":
+                metrics["surrounding_accuracy"]["mean"] >= 0.52,
+            "surrounding_accuracy_lower_bound_at_least_0p50":
+                metrics["surrounding_accuracy"][
+                    "one_sided_95pct_lower"
+                ] >= 0.50,
+            "mean_decoder_corr_at_least_0p05":
+                metrics["decoder_corr"]["mean"] >= 0.05,
+            "decoder_corr_lower_bound_nonnegative":
+                metrics["decoder_corr"][
+                    "one_sided_95pct_lower"
+                ] >= 0.0,
+            "positive_sign_in_at_least_25_of_30":
+                sum(
+                    value > 0 for value in all_values["sign_inversion"]
+                ) >= 25,
+            "positive_sign_in_every_task_cluster":
+                min(clustered["sign_inversion"]) > 0.0,
+            "mean_velocity_advantage_at_least_0p05":
+                metrics["velocity_advantage"]["mean"] >= 0.05,
+            "velocity_advantage_lower_bound_nonnegative":
+                metrics["velocity_advantage"][
+                    "one_sided_95pct_lower"
+                ] >= 0.0,
+        },
+        "target_ladder_outcome_surprise": {
+            "every_challenge_fraction_between_0p10_and_0p90":
+                min(all_values["challenge_fraction"]) >= 0.10
+                and max(all_values["challenge_fraction"]) <= 0.90,
+            "mean_terminal_accuracy_at_least_0p80":
+                metrics["terminal_accuracy"]["mean"] >= 0.80,
+            "terminal_accuracy_lower_bound_at_least_0p75":
+                metrics["terminal_accuracy"][
+                    "one_sided_95pct_lower"
+                ] >= 0.75,
+            "mean_terminal_separation_at_least_0p20":
+                metrics["terminal_separation"]["mean"] >= 0.20,
+            "terminal_separation_lower_bound_at_least_0p15":
+                metrics["terminal_separation"][
+                    "one_sided_95pct_lower"
+                ] >= 0.15,
+            "mean_outcome_lesion_drop_at_least_0p20":
+                metrics["outcome_lesion_drop"]["mean"] >= 0.20,
+            "outcome_lesion_drop_lower_bound_at_least_0p15":
+                metrics["outcome_lesion_drop"][
+                    "one_sided_95pct_lower"
+                ] >= 0.15,
+            "mean_critic_expectedness_at_least_0p05":
+                metrics["critic_expectedness"]["mean"] >= 0.05,
+            "critic_expectedness_lower_bound_at_least_0p02":
+                metrics["critic_expectedness"][
+                    "one_sided_95pct_lower"
+                ] >= 0.02,
+            "every_critic_value_corr_at_least_0p95":
+                min(all_values["critic_value_corr"]) >= 0.95,
+        },
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--results",
+        default="results/bci_v2_recovery_confirmation",
+    )
+    parser.add_argument(
+        "--r1-gate",
+        default="results/bci_v2_recovery_dev_gate.json",
+    )
+    parser.add_argument(
+        "--out",
+        default="results/bci_v2_recovery_confirmation_gate.json",
+    )
+    args = parser.parse_args()
+    with open(args.r1_gate) as handle:
+        r1 = json.load(handle)
+    require(
+        r1.get("protocol")
+        == "oral_b_v2_cold_start_recovery_development_v1"
+        and r1.get("status") == "passed"
+        and r1.get("complete_grid") is True
+        and r1.get("fixed_config") == FIXED_CONFIG
+        and r1.get("confirmation_seeds_touched") is False
+        and r1.get("recovery_confirmation_opened") is True,
+        "recovery R1 gate",
+    )
+    development_digests = {
+        name: sha256(path)
+        for name, path in source_paths().items()
+    }
+    require(
+        r1.get("input_sha256") == development_digests,
+        "R1 source binding",
+    )
+    digests = {
+        name: sha256(path)
+        for name, path in confirmation_paths(args.r1_gate).items()
+    }
+    expected = {
+        f"bci_v2_recovery_confirm_t{task_seed}_m{model_seed}.json"
+        for task_seed in TASK_SEEDS
+        for model_seed in MODEL_SEEDS
+    }
+    observed = {
+        os.path.basename(path)
+        for path in glob.glob(os.path.join(args.results, "*.json"))
+    }
+    require(
+        observed == expected,
+        (
+            f"confirmation drift: missing={sorted(expected-observed)}, "
+            f"extra={sorted(observed-expected)}"
+        ),
+    )
+    records = {}
+    commits = set()
+    source_sha256 = {}
+    for task_seed in TASK_SEEDS:
+        for model_seed in MODEL_SEEDS:
+            path = os.path.join(
+                args.results,
+                (
+                    f"bci_v2_recovery_confirm_t{task_seed}"
+                    f"_m{model_seed}.json"
+                ),
+            )
+            with open(path) as handle:
+                row = json.load(handle)
+            validate(row, path, digests)
+            records[(task_seed, model_seed)] = row
+            commits.add(row["provenance"]["git_commit"])
+            source_sha256[path] = sha256(path)
+    require(len(commits) == 1, "confirmation source revision drift")
+
+    getters = {
+        "intact_final": lambda row:
+            row["conditions"]["intact"]["final_success"],
+        "intact_gain": lambda row:
+            row["conditions"]["intact"]["learning_gain"],
+        "fixed_role_gap": lambda row: (
+            row["conditions"]["intact"]["final_success"]
+            - row["conditions"]["fixed_role"]["final_success"]
+        ),
+        "oracle_deficit": lambda row: (
+            row["conditions"]["oracle_role"]["final_success"]
+            - row["conditions"]["intact"]["final_success"]
+        ),
+        "plasticity_half_margin": lambda row: (
+            0.5 * row["conditions"]["intact"]["learning_gain"]
+            - row["conditions"]["plasticity_lesion"]["learning_gain"]
+        ),
+        "role_cosine": lambda row:
+            row["conditions"]["intact"]["role_cosine_after_training"],
+        "critic_training_gap": lambda row: (
+            row["conditions"]["intact"]["final_success"]
+            - row["conditions"]["critic_training_lesion"][
+                "final_success"
+            ]
+        ),
+        "outcome_training_gap": lambda row: (
+            row["conditions"]["intact"]["final_success"]
+            - row["conditions"]["outcome_training_lesion"][
+                "final_success"
+            ]
+        ),
+        "residual_soma_corr": lambda row:
+            row["signatures"]["mean_abs_residual_soma_corr"],
+        "raw_residual_corr_gap": lambda row:
+            row["signatures"][
+                "raw_minus_residual_abs_soma_corr"
+            ],
+        "surrounding_accuracy": lambda row:
+            row["signatures"][
+                "surrounding_event_decoder_balanced_acc"
+            ],
+        "decoder_corr": lambda row:
+            row["signatures"]["decoder_distance_residual_corr"],
+        "sign_inversion": lambda row:
+            row["signatures"][
+                "causal_role_sign_inversion_index"
+            ],
+        "velocity_advantage": lambda row:
+            row["signatures"][
+                "velocity_minus_error_abs_cv_corr"
+            ],
+        "challenge_fraction": lambda row:
+            row["signatures"]["challenge_success_fraction"],
+        "terminal_accuracy": lambda row:
+            row["signatures"][
+                "terminal_residual_outcome_balanced_acc"
+            ],
+        "terminal_separation": lambda row:
+            row["signatures"][
+                "terminal_role_aligned_outcome_separation"
+            ],
+        "outcome_lesion_drop": lambda row:
+            row["signatures"][
+                "terminal_outcome_separation_drop_under_acute_lesion"
+            ],
+        "critic_expectedness": lambda row:
+            row["signatures"][
+                "mean_critic_expectedness_contribution"
+            ],
+        "critic_value_corr": lambda row:
+            row["signatures"][
+                "critic_contribution_value_prediction_corr"
+            ],
+    }
+    clustered = {
+        name: task_clusters(records, getter)
+        for name, getter in getters.items()
+    }
+    metrics = {
+        name: summarize(values)
+        for name, values in clustered.items()
+    }
+    all_values = {
+        name: [getter(row) for row in records.values()]
+        for name, getter in getters.items()
+    }
+    gate_checks = checks(metrics, clustered, all_values)
+    passed = all(
+        value
+        for category in gate_checks.values()
+        for value in (
+            [category]
+            if isinstance(category, bool)
+            else category.values()
+        )
+    )
+    output = {
+        "protocol":
+            "oral_b_v2_cold_start_recovery_confirmation_v1",
+        "status": "passed" if passed else "failed",
+        "complete_grid": True,
+        "fixed_config": FIXED_CONFIG,
+        "checks": gate_checks,
+        "metrics": metrics,
+        "positive_sign_count": sum(
+            value > 0 for value in all_values["sign_inversion"]
+        ),
+        "source_commit": next(iter(commits)),
+        "source_sha256": source_sha256,
+        "input_sha256": digests,
+        "oral_b_v2_outcome_surprise_established": passed,
+        "old_r2_status_preserved": "failed",
+        "failed_v2_development_status_preserved": "failed",
+        "old_oral_a_gate_remains_closed": True,
+        "new_oral_a_v2_protocol_may_be_frozen": passed,
+        "review_score_before": 7,
+        "review_score_after": 8 if passed else 7,
+        "score_change_rule": (
+            "only a complete untouched recovery R2 pass establishes "
+            "role-vectorized TD outcome surprise; both prior failures "
+            "and the old oral-A gate remain unchanged"
+        ),
+    }
+    os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True)
+    if os.path.exists(args.out):
+        with open(args.out) as handle:
+            existing = json.load(handle)
+        require(existing == output, "existing recovery R2 gate differs")
+    else:
+        with open(args.out, "w") as handle:
+            json.dump(output, handle, indent=2, sort_keys=True)
+            handle.write("\n")
+    print(json.dumps(output, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/analyze_bci_v2_recovery_development.py b/experiments/analyze_bci_v2_recovery_development.py
new file mode 100644
index 0000000..a38582b
--- /dev/null
+++ b/experiments/analyze_bci_v2_recovery_development.py
@@ -0,0 +1,284 @@
+#!/usr/bin/env python3
+"""Audit the frozen oral-B-v2 cold-start recovery development validation."""
+import argparse
+import glob
+import hashlib
+import json
+import os
+
+from experiments.bci_v2_run import CONDITIONS
+from experiments.bci_v2_recovery_run import (
+    CHALLENGE_EPISODES,
+    FIXED_CONFIG,
+    MODEL_SEEDS,
+    TARGETS,
+    TASK_SEEDS,
+    build_recovery_config,
+    source_paths,
+)
+
+
+def require(condition, message):
+    if not condition:
+        raise ValueError(message)
+
+
+def sha256(path):
+    digest = hashlib.sha256()
+    with open(path, "rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def checks(row):
+    conditions = row["conditions"]
+    signatures = row["signatures"]
+    intact = conditions["intact"]
+    return {
+        "final_success_at_least_0p70":
+            intact["final_success"] >= 0.70,
+        "learning_gain_at_least_0p10":
+            intact["learning_gain"] >= 0.10,
+        "fixed_role_gap_at_least_0p20": (
+            intact["final_success"]
+            - conditions["fixed_role"]["final_success"]
+        ) >= 0.20,
+        "oracle_deficit_at_most_0p10": (
+            conditions["oracle_role"]["final_success"]
+            - intact["final_success"]
+        ) <= 0.10,
+        "plasticity_lesion_retains_at_most_half_gain":
+            conditions["plasticity_lesion"]["learning_gain"]
+            <= 0.5 * intact["learning_gain"],
+        "role_cosine_at_least_0p80":
+            intact["role_cosine_after_training"] >= 0.80,
+        "challenge_fraction_between_0p15_and_0p85":
+            0.15 <= signatures["challenge_success_fraction"] <= 0.85,
+        "terminal_outcome_accuracy_at_least_0p75":
+            signatures[
+                "terminal_residual_outcome_balanced_acc"
+            ] >= 0.75,
+        "terminal_role_separation_at_least_0p20":
+            signatures[
+                "terminal_role_aligned_outcome_separation"
+            ] >= 0.20,
+        "outcome_lesion_separation_drop_at_least_0p20":
+            signatures[
+                "terminal_outcome_separation_drop_under_acute_lesion"
+            ] >= 0.20,
+        "critic_expectedness_at_least_0p05":
+            signatures[
+                "mean_critic_expectedness_contribution"
+            ] >= 0.05,
+        "critic_contribution_value_corr_at_least_0p95":
+            signatures[
+                "critic_contribution_value_prediction_corr"
+            ] >= 0.95,
+        "residual_soma_corr_at_most_0p10":
+            signatures["mean_abs_residual_soma_corr"] <= 0.10,
+        "raw_residual_corr_gap_at_least_0p20":
+            signatures[
+                "raw_minus_residual_abs_soma_corr"
+            ] >= 0.20,
+        "surrounding_event_accuracy_at_least_0p52":
+            signatures[
+                "surrounding_event_decoder_balanced_acc"
+            ] >= 0.52,
+        "decoder_distance_corr_at_least_0p02":
+            signatures[
+                "decoder_distance_residual_corr"
+            ] >= 0.02,
+        "sign_inversion_at_least_0p01":
+            signatures[
+                "causal_role_sign_inversion_index"
+            ] >= 0.01,
+        "velocity_advantage_at_least_0p05":
+            signatures[
+                "velocity_minus_error_abs_cv_corr"
+            ] >= 0.05,
+    }
+
+
+def validate(row, path, task_seed, digests):
+    require(row.get("schema_version") == 3, f"{path}: schema")
+    require(
+        row.get("args", {}).get("task_seed") == task_seed
+        and row["args"].get("model_seed") in MODEL_SEEDS,
+        f"{path}: seeds",
+    )
+    protocol = row.get("protocol", {})
+    require(
+        protocol.get("name")
+        == "oral_b_v2_cold_start_recovery_development_v1"
+        and protocol.get("selection_split") == "development_validation"
+        and protocol.get("hyperparameter_selection") is False
+        and protocol.get("confirmation_seeds_touched") is False
+        and protocol.get("fixed_config") == FIXED_CONFIG,
+        f"{path}: protocol",
+    )
+    require(
+        protocol.get("protocol_sha256") == digests["protocol"]
+        and protocol.get("d4_gate_sha256") == digests["d4_gate"]
+        and protocol.get("old_r2_gate_sha256")
+        == digests["old_r2_gate"]
+        and protocol.get("failed_v2_gate_sha256")
+        == digests["failed_v2_gate"],
+        f"{path}: parent digest",
+    )
+    provenance = row.get("provenance", {})
+    require(
+        provenance.get("git_tracked_dirty") is False
+        and provenance.get("tracked_inputs")
+        and all(provenance["tracked_inputs"].values())
+        and set(provenance["tracked_inputs"]) == set(digests)
+        and provenance.get("input_sha256") == digests,
+        f"{path}: provenance",
+    )
+    require(
+        row.get("config") == vars(build_recovery_config()),
+        f"{path}: config",
+    )
+    require(
+        row.get("split") == "development"
+        and row.get("finite") is True,
+        f"{path}: finite/split",
+    )
+    require(
+        set(row.get("conditions", {})) == set(CONDITIONS),
+        f"{path}: conditions",
+    )
+    for name in CONDITIONS:
+        warmup = row["warmup"][name]
+        require(
+            warmup["batches"] == 100
+            and warmup["examples"] == 6400
+            and warmup["instruction_present"] is False
+            and warmup["role_cursor_scalar_observations"] == 12800
+            and warmup["predictor_max_abs_error"] <= 1e-5,
+            f"{path}: warmup {name}",
+        )
+        cost = row["conditions"][name]["cost"]
+        require(
+            len(row["conditions"][name]["daily_success"]) == 14
+            and cost["maximum_state_episode_steps"] == 25088
+            and cost["cursor_scalar_observations"]
+            == 2 * cost["role_probe_examples"]
+            and cost["task_loss_queries"] == 0
+            and cost["reverse_mode_calls"] == 0,
+            f"{path}: condition {name}",
+        )
+    challenge = row["assays"]["challenge"]
+    require(
+        challenge["targets"] == list(TARGETS)
+        and challenge["episodes_per_target"] == CHALLENGE_EPISODES
+        and challenge["selection_over_targets"] is False
+        and challenge["maximum_steps_per_episode"] == 28,
+        f"{path}: target ladder",
+    )
+    require(
+        row["signatures"]["challenge_episodes"]
+        == len(TARGETS) * CHALLENGE_EPISODES,
+        f"{path}: challenge count",
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--results", default="results/bci_v2_recovery_dev"
+    )
+    parser.add_argument(
+        "--out", default="results/bci_v2_recovery_dev_gate.json"
+    )
+    args = parser.parse_args()
+    digests = {
+        name: sha256(path)
+        for name, path in source_paths().items()
+    }
+    expected = {
+        f"bci_v2_recovery_t{task_seed}_m0.json"
+        for task_seed in TASK_SEEDS
+    }
+    observed = {
+        os.path.basename(path)
+        for path in glob.glob(os.path.join(args.results, "*.json"))
+    }
+    require(
+        observed == expected,
+        (
+            f"recovery grid drift: missing={sorted(expected-observed)}, "
+            f"extra={sorted(observed-expected)}"
+        ),
+    )
+    records = {}
+    commits = set()
+    source_sha256 = {}
+    for task_seed in TASK_SEEDS:
+        path = os.path.join(
+            args.results,
+            f"bci_v2_recovery_t{task_seed}_m0.json",
+        )
+        with open(path) as handle:
+            row = json.load(handle)
+        validate(row, path, task_seed, digests)
+        records[task_seed] = row
+        commits.add(row["provenance"]["git_commit"])
+        source_sha256[path] = sha256(path)
+    require(len(commits) == 1, "recovery source revision drift")
+    checks_by_seed = {
+        str(seed): checks(records[seed]) for seed in TASK_SEEDS
+    }
+    passed = all(
+        value for seed_checks in checks_by_seed.values()
+        for value in seed_checks.values()
+    )
+    output = {
+        "protocol":
+            "oral_b_v2_cold_start_recovery_development_v1",
+        "status": "passed" if passed else "failed",
+        "complete_grid": True,
+        "fixed_config": FIXED_CONFIG,
+        "checks_by_task_seed": checks_by_seed,
+        "intact_final_by_task_seed": [
+            records[seed]["conditions"]["intact"]["final_success"]
+            for seed in TASK_SEEDS
+        ],
+        "challenge_success_fraction_by_task_seed": [
+            records[seed]["signatures"]["challenge_success_fraction"]
+            for seed in TASK_SEEDS
+        ],
+        "terminal_outcome_accuracy_by_task_seed": [
+            records[seed]["signatures"][
+                "terminal_residual_outcome_balanced_acc"
+            ]
+            for seed in TASK_SEEDS
+        ],
+        "confirmation_seeds_touched": False,
+        "recovery_confirmation_opened": passed,
+        "source_commit": next(iter(commits)),
+        "source_sha256": source_sha256,
+        "input_sha256": digests,
+        "review_score_before": 7,
+        "review_score_after": 7,
+        "score_change_rule": (
+            "development validation never changes the formal score"
+        ),
+    }
+    os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True)
+    if os.path.exists(args.out):
+        with open(args.out) as handle:
+            existing = json.load(handle)
+        require(
+            existing == output,
+            "existing recovery development gate differs",
+        )
+    else:
+        with open(args.out, "w") as handle:
+            json.dump(output, handle, indent=2, sort_keys=True)
+            handle.write("\n")
+    print(json.dumps(output, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/bci_v2_confirmation.sh b/experiments/bci_v2_confirmation.sh
index da7328c..a071526 100644
--- a/experiments/bci_v2_confirmation.sh
+++ b/experiments/bci_v2_confirmation.sh
@@ -13,4 +13,4 @@ for task_seed in 30 31 32 33 34 35; do
   done
 done
 
-python experiments/analyze_bci_v2_confirmation.py
+PYTHONPATH=. python experiments/analyze_bci_v2_confirmation.py
diff --git a/experiments/bci_v2_development_screen.sh b/experiments/bci_v2_development_screen.sh
index 8612cae..b2c5141 100644
--- a/experiments/bci_v2_development_screen.sh
+++ b/experiments/bci_v2_development_screen.sh
@@ -19,4 +19,4 @@ for forward_eta in 0.03 0.1; do
   done
 done
 
-python experiments/analyze_bci_v2_development.py
+PYTHONPATH=. python experiments/analyze_bci_v2_development.py
diff --git a/experiments/bci_v2_recovery_confirmation.py b/experiments/bci_v2_recovery_confirmation.py
new file mode 100644
index 0000000..9d45626
--- /dev/null
+++ b/experiments/bci_v2_recovery_confirmation.py
@@ -0,0 +1,141 @@
+#!/usr/bin/env python3
+"""Run one untouched oral-B-v2 cold-start recovery confirmation cell."""
+import argparse
+import json
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+from experiments.bci_v2_recovery_run import (
+    FIXED_CONFIG,
+    PROTOCOL_PATH,
+    finite_tree,
+    provenance,
+    require_parent_gates,
+    run_cell,
+    sha256,
+    source_paths,
+)
+
+
+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+R1_GATE_PATH = os.path.join(
+    ROOT, "results", "bci_v2_recovery_dev_gate.json"
+)
+TASK_SEEDS = tuple(range(30, 36))
+MODEL_SEEDS = tuple(range(5))
+
+
+def require_r1_gate(r1):
+    digests = {
+        name: sha256(path)
+        for name, path in source_paths().items()
+    }
+    if not (
+        r1.get("protocol")
+        == "oral_b_v2_cold_start_recovery_development_v1"
+        and r1.get("status") == "passed"
+        and r1.get("complete_grid") is True
+        and r1.get("fixed_config") == FIXED_CONFIG
+        and r1.get("confirmation_seeds_touched") is False
+        and r1.get("recovery_confirmation_opened") is True
+        and r1.get("review_score_after") == 7
+        and r1.get("input_sha256") == digests
+    ):
+        raise ValueError(
+            "recovery R2 requires the complete eligible R1 gate"
+        )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--task-seed", type=int, choices=TASK_SEEDS, required=True
+    )
+    parser.add_argument(
+        "--model-seed", type=int, choices=MODEL_SEEDS, required=True
+    )
+    parser.add_argument(
+        "--r1-gate",
+        default="results/bci_v2_recovery_dev_gate.json",
+    )
+    parser.add_argument(
+        "--outdir", default="results/bci_v2_recovery_confirmation"
+    )
+    args = parser.parse_args()
+    require_parent_gates()
+    with open(args.r1_gate) as handle:
+        r1 = json.load(handle)
+    require_r1_gate(r1)
+    source = provenance({
+        "development_runner": os.path.join(
+            ROOT, "experiments", "bci_v2_recovery_run.py"
+        ),
+        "runner": os.path.abspath(__file__),
+        "r1_gate": os.path.abspath(args.r1_gate),
+    })
+    if (
+        source["git_tracked_dirty"]
+        or not all(source["tracked_inputs"].values())
+    ):
+        raise RuntimeError(
+            "recovery R2 requires clean, tracked, frozen inputs"
+        )
+    cell = run_cell(
+        args.task_seed,
+        args.model_seed,
+        split="untouched_confirmation",
+        performance_seed_offset=520_000,
+        challenge_seed_offset=530_000,
+    )
+    result = {
+        "schema_version": 3,
+        "protocol": {
+            "name":
+                "oral_b_v2_cold_start_recovery_confirmation_v1",
+            "split": "untouched_confirmation",
+            "training_task_seed": args.task_seed,
+            "model_seed": args.model_seed,
+            "fixed_config": FIXED_CONFIG,
+            "no_further_selection": True,
+            "confirmation_grid_size": (
+                len(TASK_SEEDS) * len(MODEL_SEEDS)
+            ),
+            "protocol_sha256": sha256(PROTOCOL_PATH),
+            "r1_gate_sha256": sha256(args.r1_gate),
+        },
+        "args": vars(args),
+        "provenance": source,
+        **cell,
+    }
+    result["finite"] = finite_tree(result)
+    if not result["finite"]:
+        raise RuntimeError("non-finite recovery confirmation record")
+    os.makedirs(args.outdir, exist_ok=True)
+    path = os.path.join(
+        args.outdir,
+        (
+            f"bci_v2_recovery_confirm_t{args.task_seed}"
+            f"_m{args.model_seed}.json"
+        ),
+    )
+    if os.path.exists(path):
+        raise FileExistsError(f"refusing to overwrite {path}")
+    with open(path, "w") as handle:
+        json.dump(result, handle, indent=2, sort_keys=True)
+        handle.write("\n")
+    print(json.dumps({
+        "path": path,
+        "intact_final": result["conditions"]["intact"]["final_success"],
+        "challenge_success_fraction": result["signatures"][
+            "challenge_success_fraction"
+        ],
+        "terminal_outcome_accuracy": result["signatures"][
+            "terminal_residual_outcome_balanced_acc"
+        ],
+        "finite": result["finite"],
+    }, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/bci_v2_recovery_confirmation.sh b/experiments/bci_v2_recovery_confirmation.sh
new file mode 100644
index 0000000..16534b1
--- /dev/null
+++ b/experiments/bci_v2_recovery_confirmation.sh
@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+# Frozen v2 recovery R2: 6 untouched task seeds x 5 model seeds.
+set -euo pipefail
+
+export OMP_NUM_THREADS=1
+export MKL_NUM_THREADS=1
+
+for task_seed in 30 31 32 33 34 35; do
+  for model_seed in 0 1 2 3 4; do
+    python experiments/bci_v2_recovery_confirmation.py \
+      --task-seed "$task_seed" \
+      --model-seed "$model_seed"
+  done
+done
+
+PYTHONPATH=. python experiments/analyze_bci_v2_recovery_confirmation.py
diff --git a/experiments/bci_v2_recovery_development.sh b/experiments/bci_v2_recovery_development.sh
new file mode 100644
index 0000000..68f22f5
--- /dev/null
+++ b/experiments/bci_v2_recovery_development.sh
@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+# Frozen v2 recovery R1: one fixed mechanism x 3 new development seeds.
+set -euo pipefail
+
+export OMP_NUM_THREADS=1
+export MKL_NUM_THREADS=1
+
+for task_seed in 23 24 25; do
+  python experiments/bci_v2_recovery_run.py --task-seed "$task_seed"
+done
+
+PYTHONPATH=. python experiments/analyze_bci_v2_recovery_development.py
diff --git a/experiments/bci_v2_recovery_run.py b/experiments/bci_v2_recovery_run.py
new file mode 100644
index 0000000..29507b3
--- /dev/null
+++ b/experiments/bci_v2_recovery_run.py
@@ -0,0 +1,377 @@
+#!/usr/bin/env python3
+"""Run one frozen oral-B-v2 cold-start recovery development cell."""
+import argparse
+from dataclasses import replace
+import hashlib
+import json
+import os
+import platform
+import resource
+import subprocess
+import sys
+import time
+
+import torch
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+from experiments.bci_v2_run import (
+    ACUTE_MODES,
+    CONDITIONS,
+    build_config,
+    evaluate_performance,
+    finite_tree,
+    neutral_warmup,
+    train,
+)
+from sdil.bci import generate_trajectories
+from sdil.bci_v2 import BCIV2, run_day_v2
+from sdil.bci_v2_recovery_metrics import (
+    annotate_target_events,
+    recovery_signature_metrics,
+)
+
+
+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+PROTOCOL_PATH = os.path.join(ROOT, "ORAL_B_V2_RECOVERY.md")
+D4_GATE_PATH = os.path.join(
+    ROOT, "results", "kp_dynamic_projection_confirmation_gate.json"
+)
+OLD_R2_GATE_PATH = os.path.join(
+    ROOT, "results", "bci_td_confirmation_gate.json"
+)
+FAILED_V2_GATE_PATH = os.path.join(
+    ROOT, "results", "bci_v2_dev_gate.json"
+)
+TASK_SEEDS = (23, 24, 25)
+MODEL_SEEDS = (0,)
+TARGETS = (1.55, 1.60, 1.65, 1.70, 1.75, 1.80)
+CHALLENGE_EPISODES = 128
+FIXED_CONFIG = {
+    "forward_eta": 0.1,
+    "gamma": 0.8,
+    "critic_eta": 0.03,
+    "velocity_reward_scale": 1.0,
+}
+
+
+def sha256(path):
+    digest = hashlib.sha256()
+    with open(path, "rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def source_paths():
+    return {
+        "runner": os.path.abspath(__file__),
+        "development_analyzer": os.path.join(
+            ROOT,
+            "experiments",
+            "analyze_bci_v2_recovery_development.py",
+        ),
+        "confirmation_runner": os.path.join(
+            ROOT, "experiments", "bci_v2_recovery_confirmation.py"
+        ),
+        "confirmation_analyzer": os.path.join(
+            ROOT,
+            "experiments",
+            "analyze_bci_v2_recovery_confirmation.py",
+        ),
+        "base_dynamics": os.path.join(ROOT, "sdil", "bci.py"),
+        "v2_dynamics": os.path.join(ROOT, "sdil", "bci_v2.py"),
+        "v2_metrics": os.path.join(ROOT, "sdil", "bci_v2_metrics.py"),
+        "recovery_metrics": os.path.join(
+            ROOT, "sdil", "bci_v2_recovery_metrics.py"
+        ),
+        "protocol": PROTOCOL_PATH,
+        "d4_gate": D4_GATE_PATH,
+        "old_r2_gate": OLD_R2_GATE_PATH,
+        "failed_v2_gate": FAILED_V2_GATE_PATH,
+    }
+
+
+def provenance(extra_paths=None):
+    def run(command):
+        return subprocess.run(
+            command,
+            cwd=ROOT,
+            check=True,
+            capture_output=True,
+            text=True,
+        ).stdout.strip()
+
+    paths = source_paths()
+    if extra_paths:
+        paths.update(extra_paths)
+    relative = {
+        name: os.path.relpath(path, ROOT)
+        for name, path in paths.items()
+    }
+    tracked = {
+        name: subprocess.run(
+            ["git", "ls-files", "--error-unmatch", path],
+            cwd=ROOT,
+            capture_output=True,
+        ).returncode == 0
+        for name, path in relative.items()
+    }
+    return {
+        "git_commit": run(["git", "rev-parse", "HEAD"]),
+        "git_tracked_dirty": bool(run([
+            "git", "status", "--porcelain", "--untracked-files=no"
+        ])),
+        "tracked_inputs": tracked,
+        "input_sha256": {
+            name: sha256(path) for name, path in paths.items()
+        },
+    }
+
+
+def require_parent_gates():
+    with open(D4_GATE_PATH) as handle:
+        d4 = json.load(handle)
+    with open(OLD_R2_GATE_PATH) as handle:
+        old_r2 = json.load(handle)
+    with open(FAILED_V2_GATE_PATH) as handle:
+        failed_v2 = json.load(handle)
+    if not (
+        d4.get("status") == "passed"
+        and d4.get("review_score_after") == 7
+        and old_r2.get("status") == "failed"
+        and old_r2.get("review_score_after") == 7
+        and failed_v2.get("protocol") == "oral_b_v2_development_v1"
+        and failed_v2.get("status") == "failed"
+        and failed_v2.get("complete_grid") is True
+        and failed_v2.get("oral_b_v2_confirmation_opened") is False
+        and failed_v2.get("review_score_after") == 7
+    ):
+        raise ValueError(
+            "cold-start recovery requires D4 and both preserved failures"
+        )
+
+
+def build_recovery_config():
+    return replace(
+        build_config(
+            FIXED_CONFIG["forward_eta"],
+            FIXED_CONFIG["gamma"],
+            FIXED_CONFIG["critic_eta"],
+        ),
+        velocity_reward_scale=FIXED_CONFIG[
+            "velocity_reward_scale"
+        ],
+    )
+
+
+def evaluate_target_ladder(model, task_seed, seed_offset):
+    mode_events = {name: [] for name in ACUTE_MODES}
+    costs = {name: 0 for name in ACUTE_MODES}
+    seeds = {}
+    for target_index, target in enumerate(TARGETS, start=1):
+        challenge_cfg = replace(model.cfg, target=target)
+        trajectory_seed = (
+            seed_offset + 1_000 * task_seed + target_index - 1
+        )
+        seeds[str(target)] = trajectory_seed
+        trajectories = generate_trajectories(
+            challenge_cfg,
+            trajectory_seed,
+            days=1,
+            episodes=CHALLENGE_EPISODES,
+        )
+        for mode in ACUTE_MODES:
+            settings = {
+                "intact": {},
+                "acute_critic_lesion": {"critic_enabled": False},
+                "acute_outcome_lesion": {
+                    "terminal_outcome_enabled": False
+                },
+            }[mode]
+            assay_model = model.clone()
+            assay_model.cfg = challenge_cfg
+            report = run_day_v2(
+                assay_model,
+                trajectories,
+                0,
+                horizon=challenge_cfg.steps_per_episode,
+                plasticity_gain=0.0,
+                learn_role=False,
+                probe_role=False,
+                learn_predictor=False,
+                learn_critic=False,
+                collect=True,
+                **settings,
+            )
+            episode_offset = (
+                2_000_000
+                + (target_index - 1) * CHALLENGE_EPISODES
+            )
+            annotate_target_events(
+                report["events"],
+                0,
+                report["success"],
+                episode_offset,
+                target_index,
+            )
+            mode_events[mode].extend(report["events"])
+            costs[mode] += report["active_transitions"]
+    return mode_events, {
+        "targets": list(TARGETS),
+        "episodes_per_target": CHALLENGE_EPISODES,
+        "trajectory_seeds": seeds,
+        "active_state_episode_steps_by_mode": costs,
+        "selection_over_targets": False,
+        "maximum_steps_per_episode": model.cfg.steps_per_episode,
+    }
+
+
+def run_cell(
+    task_seed,
+    model_seed,
+    *,
+    split,
+    performance_seed_offset,
+    challenge_seed_offset,
+):
+    cfg = build_recovery_config()
+    trajectories = generate_trajectories(cfg, task_seed)
+    performance_seed = performance_seed_offset + task_seed
+    performance_trajectories = generate_trajectories(
+        cfg, performance_seed, days=1, episodes=256
+    )
+    initial = BCIV2(cfg, model_seed)
+    conditions = {}
+    trained = {}
+    warmups = {}
+    training_events = None
+    started = time.perf_counter()
+    for name in CONDITIONS:
+        model, warmup = neutral_warmup(
+            initial, task_seed, model_seed, name
+        )
+        events, report = train(
+            model, trajectories, name, collect=name == "intact"
+        )
+        warmups[name] = warmup
+        conditions[name] = report
+        trained[name] = model
+        if name == "intact":
+            training_events = events
+    for name in CONDITIONS:
+        report = evaluate_performance(
+            trained[name], performance_trajectories
+        )
+        conditions[name]["final_success"] = report["success_rate"]
+        conditions[name]["evaluation_active_state_episode_steps"] = (
+            report["active_transitions"]
+        )
+    challenge_events, challenge_cost = evaluate_target_ladder(
+        trained["intact"], task_seed, challenge_seed_offset
+    )
+    signatures = recovery_signature_metrics(
+        training_events,
+        challenge_events,
+        cfg,
+        trained["intact"].role,
+        TARGETS,
+    )
+    return {
+        "config": vars(cfg),
+        "warmup": warmups,
+        "conditions": conditions,
+        "signatures": signatures,
+        "assays": {
+            "performance_evaluation_seed": performance_seed,
+            "performance_evaluation_episodes": 256,
+            "challenge": challenge_cost,
+        },
+        "wall_s": time.perf_counter() - started,
+        "peak_rss_mib": (
+            resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024
+        ),
+        "hardware": {
+            "device": "cpu",
+            "platform": platform.platform(),
+            "torch_version": torch.__version__,
+            "threads": torch.get_num_threads(),
+        },
+        "split": split,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--task-seed", type=int, choices=TASK_SEEDS, required=True
+    )
+    parser.add_argument(
+        "--model-seed", type=int, choices=MODEL_SEEDS, default=0
+    )
+    parser.add_argument(
+        "--outdir", default="results/bci_v2_recovery_dev"
+    )
+    args = parser.parse_args()
+    require_parent_gates()
+    source = provenance()
+    if (
+        source["git_tracked_dirty"]
+        or not all(source["tracked_inputs"].values())
+    ):
+        raise RuntimeError(
+            "v2 recovery R1 requires clean, tracked, frozen inputs"
+        )
+    cell = run_cell(
+        args.task_seed,
+        args.model_seed,
+        split="development",
+        performance_seed_offset=460_000,
+        challenge_seed_offset=470_000,
+    )
+    result = {
+        "schema_version": 3,
+        "protocol": {
+            "name": "oral_b_v2_cold_start_recovery_development_v1",
+            "selection_split": "development_validation",
+            "hyperparameter_selection": False,
+            "confirmation_seeds_touched": False,
+            "fixed_config": FIXED_CONFIG,
+            "protocol_sha256": source["input_sha256"]["protocol"],
+            "d4_gate_sha256": source["input_sha256"]["d4_gate"],
+            "old_r2_gate_sha256": source["input_sha256"]["old_r2_gate"],
+            "failed_v2_gate_sha256": source["input_sha256"][
+                "failed_v2_gate"
+            ],
+        },
+        "args": vars(args),
+        "provenance": source,
+        **cell,
+    }
+    result["finite"] = finite_tree(result)
+    if not result["finite"]:
+        raise RuntimeError("non-finite v2 recovery development record")
+    os.makedirs(args.outdir, exist_ok=True)
+    path = os.path.join(
+        args.outdir,
+        f"bci_v2_recovery_t{args.task_seed}_m0.json",
+    )
+    if os.path.exists(path):
+        raise FileExistsError(f"refusing to overwrite {path}")
+    with open(path, "w") as handle:
+        json.dump(result, handle, indent=2, sort_keys=True)
+        handle.write("\n")
+    print(json.dumps({
+        "path": path,
+        "intact_final": result["conditions"]["intact"]["final_success"],
+        "challenge_success_fraction": result["signatures"][
+            "challenge_success_fraction"
+        ],
+        "terminal_outcome_accuracy": result["signatures"][
+            "terminal_residual_outcome_balanced_acc"
+        ],
+        "finite": result["finite"],
+    }, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/experiments/bci_v2_recovery_smoke.py b/experiments/bci_v2_recovery_smoke.py
new file mode 100644
index 0000000..39055c1
--- /dev/null
+++ b/experiments/bci_v2_recovery_smoke.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+"""Endpoint-free checks for the v2 cold-start recovery."""
+from dataclasses import replace
+import os
+import sys
+
+import torch
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+from experiments.bci_v2_recovery_run import (
+    TARGETS,
+    build_recovery_config,
+)
+from sdil.bci_v2 import BCIV2
+
+
+def copy_state(source, target):
+    for name in ("W", "A", "coupling", "P", "P_bias", "critic"):
+        setattr(target, name, getattr(source, name).clone())
+
+
+def check_dense_velocity_cold_start():
+    recovered_cfg = replace(
+        build_recovery_config(),
+        days=2,
+        episodes_per_day=8,
+        steps_per_episode=2,
+        target=10.0,
+        inertia=0.0,
+        process_noise=0.0,
+        forward_eta=0.1,
+        perturb_every=99,
+    )
+    failed_cfg = replace(recovered_cfg, velocity_reward_scale=0.25)
+    recovered = BCIV2(recovered_cfg, model_seed=1)
+    recovered.P.copy_(recovered.coupling)
+    recovered.A.copy_(recovered.role)
+    recovered.critic.zero_()
+    failed = BCIV2(failed_cfg, model_seed=1)
+    copy_state(recovered, failed)
+    context = torch.zeros(8, recovered_cfg.context_dim)
+    context[:, 0] = 1.0
+    noise = torch.zeros(8, recovered_cfg.n_neurons)
+    noise[:, :recovered_cfg.n_plus] = torch.linspace(
+        0.1, 0.8, 8
+    ).unsqueeze(1)
+    perturbations = torch.ones_like(noise)
+    recovered_initial_w = recovered.W.clone()
+    failed_initial_w = failed.W.clone()
+    recovered_state = recovered.initial_episode_state(8)
+    failed_state = failed.initial_episode_state(8)
+    _, recovered_record = recovered.step(
+        context,
+        noise,
+        perturbations,
+        recovered_state,
+        global_step=1,
+        final_step=False,
+        learn_role=False,
+        probe_role=False,
+        learn_predictor=False,
+        learn_critic=False,
+        critic_enabled=False,
+    )
+    _, failed_record = failed.step(
+        context,
+        noise,
+        perturbations,
+        failed_state,
+        global_step=1,
+        final_step=False,
+        learn_role=False,
+        probe_role=False,
+        learn_predictor=False,
+        learn_critic=False,
+        critic_enabled=False,
+    )
+    assert torch.allclose(
+        recovered_record["td_innovation"],
+        4.0 * failed_record["td_innovation"],
+    )
+    assert torch.allclose(
+        recovered.W - recovered_initial_w,
+        4.0 * (failed.W - failed_initial_w),
+        atol=1e-7,
+    )
+    print("recovery dense cold-start drive is exactly 4x failed v2")
+
+
+def check_fixed_target_ladder():
+    assert TARGETS == (1.55, 1.60, 1.65, 1.70, 1.75, 1.80)
+    maxima = torch.tensor([1.58, 1.63, 1.68, 1.73, 1.78])
+    fractions = torch.tensor([
+        (maxima >= target).float().mean() for target in TARGETS
+    ])
+    assert torch.all(fractions[:-1] >= fractions[1:])
+    assert bool((fractions > 0).any()) and bool((fractions < 1).any())
+    print("recovery target ladder is fixed, ordered, and monotone")
+
+
+if __name__ == "__main__":
+    check_dense_velocity_cold_start()
+    check_fixed_target_ladder()
+    print("ALL V2 COLD-START RECOVERY MECHANICS CHECKS PASSED")
diff --git a/sdil/bci_v2_recovery_metrics.py b/sdil/bci_v2_recovery_metrics.py
new file mode 100644
index 0000000..78d3449
--- /dev/null
+++ b/sdil/bci_v2_recovery_metrics.py
@@ -0,0 +1,44 @@
+"""Target-ladder wrapper for the oral-B-v2 cold-start recovery."""
+
+from sdil.bci_v2_metrics import (
+    annotate_events,
+    signature_metrics_v2,
+)
+
+
+def annotate_target_events(
+    events,
+    day,
+    final_success,
+    episode_offset,
+    target_index,
+):
+    """Reuse paired grouping while encoding the fixed target-ladder index."""
+    return annotate_events(
+        events,
+        day,
+        final_success,
+        episode_offset,
+        target_index,
+    )
+
+
+def recovery_signature_metrics(
+    train_events,
+    challenge_mode_events,
+    cfg,
+    role,
+    targets,
+):
+    values = signature_metrics_v2(
+        train_events,
+        challenge_mode_events,
+        cfg,
+        role,
+    )
+    by_index = values.pop("horizon_success_fraction")
+    values["target_success_fraction"] = {
+        str(target): by_index[str(index)]
+        for index, target in enumerate(targets, start=1)
+    }
+    return values