learn(exp-004): Req 6.4 pooled BLOCK gate with a family floor; manifest before units

Maintainer decision 2026-10-03: BLOCK precision is shown pooled across
families. Req 6.4 (amended; 7.4 annotated): one t_block on a per-family
Platt-normalised score (calibrate.fit_map, training folds, inner
cross-fit), Wilson 95% LB >= 0.95 over all families' training flags; a
family blocks only with >= F = 10 held-out flags and zero errors, else
advisory; t_quiet per family keeps >= 90% of training paired positives.

evaluate: normalised, pooled_threshold, quiet_threshold,
pooled_block_gate, family_floor, apply_floor. experiments: rule
pooled_block_gate (normalise: platt, floor, recall), decide_pooled on the
pre-floor pooled held-out LB, floor table, quiet_recall and after-floor
totals in analyse_rule; exp-003's analysis is unchanged.

plane/experiments/exp-004-pooled-block-gate.yaml: procedure and decision
rule registered now; units frozen at run time from the post-harvest
memory (units_status pending until then).

Tests (tests/test_learn_rules.py): pooled threshold where no family alone
reaches the bound, no held-out label moves a threshold or the map, a
family with one held-out error cannot block, t_quiet and no blind
silencing, the pooled verdict, the exp-004 manifest.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo

.kiro/specs/learned-decision-engine/requirements.md

@@ -119,6 +119,19 @@ rule hits.
 2. Without a trained model available the tools fall back to today's behaviour and say so in the report.
 3. Hand-written rules remain as signal producers; changing their enabled/shadow status no longer changes what is
    reported unless the engine's decision changes.
+4. *(Amended 2026-10-03, maintainer decision; refines the BLOCK rule of 7.3 and 7.4. Tested by exp-004.)* BLOCK
+   precision is demonstrated pooled across families, not per family:
+   - **Normalisation**, declared before any run: each family's raw score is put on one scale by a per-family Platt
+     map (`learn.calibrate.fit_map` with its floors) fitted on training folds only. A family the training folds
+     cannot map does not block on that fold.
+   - **t_block**: one threshold on the normalised score for all families, chosen on training folds as the lowest
+     one whose Wilson 95% lower bound on precision, over the training flags of all families together, is >= 0.95.
+     It is reported by the Wilson 95% lower bound over all held-out flags of all families, before the floor.
+   - **Family floor**: a family may block only if its own held-out flags have zero errors and number at least F
+     (F = 10, reasoned in the exp-004 registration); otherwise that family stays ADVISORY.
+   - **t_quiet**: per family, the highest threshold keeping >= 90% of the paired positives at or above it (chosen
+     on training folds, its held-out recall reported); findings below it are silent.
+   - No held-out label moves a threshold or a normalisation map; candidates whose verdict is `unsure` never block.
 
 ### Requirement 7: Built for unseen repositories -- a repo safety net
 
@@ -136,7 +149,8 @@ rule hits.
    ADVISORY, trading precision for recall; both reported with out-of-repository intervals. The decision unit is a
    candidate in changed code (`pre-push` delta), with whole-repository scans as a secondary mode.
 4. Calibration is checked on held-out repositories; where it does not hold, only ADVISORY is offered for that
-   family.
+   family. *(Amended 2026-10-03: once exp-004 adopts the pooled gate of 6.4, BLOCK precision is shown by its held-out
+   bound and family floor, and calibration is reported, not required.)*
 5. Any adaptation to a user's own repository (facts reuse, their dismissals) is opt-in, separate from the
    generalisation model, and never contributes to its reported numbers.
 

plane/experiments/exp-004-pooled-block-gate.yaml

@@ -0,0 +1,90 @@
+# Experiment exp-004 (learned-decision-engine Req 6.4, amended 2026-10-03 by maintainer decision): the pooled BLOCK
+# gate with a family floor. Registered before it runs; `ousast learn experiment register` records this file's sha256,
+# and a changed file is a new experiment. The procedure and the decision rule are fixed here; the units are NOT: they
+# are frozen at run time (`ousast learn experiment units`) from the memory after the negatives harvest (merged in
+# 5b2bd9b, 1133 static examples), committed, and registered again before the run. Until then the registration row
+# says `units_status: pending` and the run refuses.
+#
+# Scoring: kind rule, zero cost by construction -- one scoring pass of the six slice-2 incumbent programs, replayed from
+# the response cache behind a zero-ceiling meter; a unit with any uncached response is excluded and counted. The
+# harvest changed the memory the programs retrieve from, so most prompts are new: a paid cache fill (separately
+# approved, its own ceiling, as for exp-003's supplementary analysis) must precede the run, or the exclusions will be
+# most of the units.
+#
+# Normalisation (declared before the run): per-family Platt scaling, learn.calibrate.fit_map with its existing floors
+# (own slope and intercept per family at >= 20 positive and negative repository groups; a pooled slope with a
+# per-family intercept at >= 10; below that no map, so the family cannot block on that fold). Fitted on the training
+# folds' paired units only. Why Platt and not rank-to-quantile: a quantile map gives every family the same flag rate
+# whatever its precision, so one pooled threshold would flag the top slice of a weak family as readily as of a strong
+# one. Platt puts each family on the scale of its own training precision, which is what a pooled precision bound needs.
+# The map is a normaliser only: exp-003 showed calibration (ECE <= 0.05) does not hold per family, and precision is
+# demonstrated here by the bound on held-out flags, not by calibration.
+#
+# Thresholds, per outer fold j, from the other four folds only (fold j's labels never enter a threshold or a map;
+# tests/test_learn_rules.py):
+#   - the training units are normalised by an inner cross-fit (each training fold k by a map fitted on the training
+#     folds other than k), so t_block is chosen on out-of-sample scores;
+#   - t_block: the lowest normalised score at which the training paired units of ALL families together with p >= t
+#     and a verdict other than `unsure` have a Wilson 95% lower bound of precision >= 0.95; none -> no BLOCK on fold j;
+#   - t_quiet, per family: the highest normalised score keeping >= 90% of that family's training paired positives at
+#     or above it;
+#   - fold j's units, normalised by the map fitted on all its training units: BLOCK when not `unsure` and p >= t_block;
+#     ADVISORY when p >= t_quiet (or the family has no map or no t_quiet: nothing is silenced blind); silent below.
+#
+# Family floor F = 10: a family may block only if its own held-out paired flags (all five folds) number >= 10 with zero
+# errors; otherwise its flags are reported as ADVISORY. Why 10: zero errors in 10 flags is unlikely for a family that is
+# much worse than the bar (a family at 0.75 precision produces 10 clean flags with probability 0.056, at 0.80 with
+# 0.107); a larger F is out of reach of the smaller families (16-19 paired positives in exp-003's units, all folds); and
+# 10 is exp-003 arm C's training minimum. The floor uses held-out labels to decide which families block; it moves no
+# threshold. The primary metric is computed BEFORE the floor, over every family's held-out flags, so dropping a failing
+# family cannot raise it.
+#
+# Power note: the pooled bound needs >= 73 flagged units with no error (110 with one, 142 with two). exp-003's
+# instrument check found few positives above the first negative per family on the raw score (paired: 10, 1, 3, 0, 1);
+# per-family normalisation keeps each family's order, so the pooled top set can reach 73 only if the harvest's pairs
+# separate better. An inconclusive verdict is a plausible outcome and will be reported as such.
+id: exp-004-pooled-block-gate
+hypothesis: >
+  One BLOCK threshold pooled across families on a per-family Platt-normalised score, chosen on training folds only as
+  the lowest one whose Wilson 95% lower bound on flagged precision is >= 0.95, reaches that bound on held-out
+  repositories' paired units, where per-family thresholds (exp-003) could not for lack of flags; a per-family floor
+  (>= 10 held-out flags, none wrong) keeps a family that the pooling masks from blocking.
+kind: rule
+programs:                   # the slice-2 incumbents, as exp-003
+  access_control: 745cb1d76bbb09dc1574d333d2118489be62e39886517894d3699fa565b0e8c2
+  deserialization: 3c0c2831be7d78198a62351661c37867363c501cb40e3ed7b0119c155af09654
+  injection: b207a0710e0abf053b3b9397b25f9ee1cab4679144865e3a82c1b9be2c7b4982
+  output_encoding: 5d768f16724562c381c09949ff3df9cbb11d58cbb7a9c0515d0dc3f858d3818e
+  path: aea396680a4874dcb6de70e240ca0480c83424debdd3710bc3f1ea6a612661f3
+  untrusted_destination: 1242b3973fc89056aa08e00843043d68235df217d8e3ef8a0ea0e0b81c142753
+arms:
+  A:
+    label: today's rule -- Platt map cross-fitted per family, BLOCK only when the calibration check holds, BLOCK point nested on the other folds' calibrated p
+    rule: {name: calibrated_block}
+  B:
+    label: pooled gate -- one t_block over all families on a per-family Platt score (training folds, inner cross-fit), t_quiet per family at 90% recall, family floor 10 clean held-out flags
+    rule: {name: pooled_block_gate, precision: 0.95, recall: 0.9, floor: 10, normalise: platt, select_on: paired}
+unit: candidate
+pairing: paired             # every unit scored once; every arm decides every unit
+cluster: repository
+order: {randomise: none, seed: 4}
+# primary population: paired units (both sides of the candidate present); pooled (every unit) is reported beside it
+populations: {primary: paired, secondary: [pooled]}
+metric:
+  primary:
+    - {name: flagged_precision_lb, target: 0.95}   # held-out Wilson 95% LB of flagged precision, all families pooled, arm B, paired, before the floor
+  secondary: [flagged_precision, coverage, flagged, quiet_recall]  # quiet_recall: held-out positives at or above t_quiet / positives
+# Decision rule (decide_pooled), on arm B's held-out paired flags of all families together, before the floor:
+#   REJECT B if their point precision < 0.95;
+#   else ADOPT B if their Wilson 95% LB >= 0.95 and at least one family passes the floor (>= 10 held-out paired
+#   flags, zero errors) -- those families block, every other family stays advisory;
+#   else INCONCLUSIVE (the incumbent rule stays).
+# Reported, not deciding: per-family held-out precision and LB, the floor table, quiet_recall per family against 0.90,
+# the after-floor totals (what users would get), the repository-cluster bootstrap interval.
+decision: {arm: B, target_precision: 0.95, pooled: true}
+test: {resamples: 2000}     # repository-cluster bootstrap of the held-out flagged precision, beside the Wilson bound
+mde: 0.0                    # not a paired difference test; see the power note above
+budget_usd: {total: 0}      # the run itself is zero cost; the cache fill before it is a separate, approved spend
+families: [access_control, deserialization, injection, output_encoding, path, untrusted_destination]
+folds: {kind: outer, k: 5, seed: 0, compile_fraction: 0.25}   # the incumbent evaluation's folds (learn/compile.toml seed)
+units_file: plane/experiments/exp-004-pooled-block-gate.units.jsonl   # frozen at run time, after the negatives harvest

src/openultrasast/learn/evaluate.py

@@ -25,7 +25,7 @@ from collections.abc import Callable, Iterable, Mapping, Sequence
 from dataclasses import dataclass, replace
 from typing import Any
 
-from .calibrate import Row, check, cross_fit, offered
+from .calibrate import Row, check, cross_fit, fit_map, offered
 from .examples import Example
 from .folds import Fold
 from .labels import PER_FAMILY_FLOOR, POOLED_FLOOR
@@ -210,6 +210,98 @@ def calibrated_block(preds: Sequence[Prediction], *, seed: int = 0, **floors: in
     return out, info
 
 
+# --- the pooled BLOCK gate (Req 6.4, amended 2026-10-03; exp-004) ---------------------------------------------------------
+
+FAMILY_FLOOR = 10  # a family blocks only with at least this many held-out flags and none of them wrong
+
+
+def normalised(train: Sequence[Prediction], rows: Sequence[Prediction], **floors: int) -> list[float | None]:
+    """``rows``' raw scores on one scale across families: the per-family Platt map of :func:`.calibrate.fit_map`
+    (pooled slope with a per-family intercept between the floors) fitted on ``train`` only. ``None`` for a family the
+    training rows cannot map (below the pooled floor): it cannot block on that fold."""
+    cmap = fit_map([Row(r.fold, r.family, r.group, r.label, r.s) for r in train], **floors)
+    return [cmap.apply(r.family, r.s) for r in rows]
+
+
+def pooled_threshold(rows: Sequence[Prediction], *, precision: float = TARGET_PRECISION) -> float | None:
+    """The lowest normalised score ``t`` at which the blockable rows of every family together with ``p >= t`` have a
+    Wilson 95% lower bound of precision >= ``precision``; candidate thresholds are the rows' own ``p``."""
+    blockable = [r for r in rows if r.blockable and r.p is not None]
+    for t in sorted({float(r.p) for r in blockable if r.p is not None}):
+        chosen = [r for r in blockable if (r.p or 0.0) >= t]
+        if wilson_lower(sum(r.label for r in chosen), len(chosen)) >= precision:
+            return t
+    return None
+
+
+def quiet_threshold(rows: Sequence[Prediction], *, recall: float = TARGET_RECALL) -> float | None:
+    """The highest normalised score ``t`` that keeps at least ``recall`` of the rows' positives at or above it."""
+    positives = sorted((float(r.p) for r in rows if r.label == 1 and r.p is not None), reverse=True)
+    if not positives:
+        return None
+    return positives[math.ceil(recall * len(positives) - 1e-9) - 1]
+
+
+def pooled_block_gate(
+    preds: Sequence[Prediction],
+    *,
+    eligible: Callable[[Prediction], bool] | None = None,
+    precision: float = TARGET_PRECISION,
+    recall: float = TARGET_RECALL,
+    **floors: int,
+) -> tuple[list[Prediction], dict[str, dict[str, Any]]]:
+    """One BLOCK threshold for every family, on a normalised score, chosen per outer fold j on the other folds only.
+
+    The training rows (``eligible`` rows of the other folds) are normalised by an inner cross-fit: each training fold
+    k by a map fitted on the training folds other than k, so ``t_block`` is chosen on out-of-sample scores. ``t_block``
+    is the lowest normalised score whose pooled Wilson 95% lower bound of precision is >= ``precision``; ``t_quiet``
+    per family is the highest one keeping >= ``recall`` of that family's training positives. Fold j's rows are then
+    normalised by the map fitted on all its training rows and decided: ``BLOCK`` when blockable and ``p >= t_block``;
+    ``ADVISORY`` when ``p >= t_quiet`` (or the family has no map, or no ``t_quiet``: nothing is silenced blind);
+    ``DROP`` (silent) below ``t_quiet``. Fold j's labels never enter its thresholds or its map."""
+    chosen = [r for r in preds if eligible is None or eligible(r)]
+    families = sorted({r.family for r in preds})
+    decided: dict[int, Prediction] = {}
+    gates: dict[str, dict[str, Any]] = {}
+    for fold in sorted({r.fold for r in preds}):
+        train = [r for r in chosen if r.fold != fold]
+        inner: list[Prediction] = []
+        for k in sorted({r.fold for r in train}):
+            mine = [r for r in train if r.fold == k]
+            inner += [replace(r, p=p) for r, p in zip(mine, normalised([r for r in train if r.fold != k], mine, **floors), strict=True)
+                      if p is not None]  # fmt: skip
+        t_block = pooled_threshold(inner, precision=precision)
+        t_quiet = {f: quiet_threshold([r for r in inner if r.family == f], recall=recall) for f in families}
+        held = [(i, r) for i, r in enumerate(preds) if r.fold == fold]
+        for (i, r), p in zip(held, normalised(train, [r for _, r in held], **floors), strict=True):
+            quiet = t_quiet[r.family]
+            if p is not None and t_block is not None and r.blockable and p >= t_block:
+                point = "BLOCK"
+            elif p is None or quiet is None or p >= quiet:
+                point = "ADVISORY"
+            else:
+                point = "DROP"
+            decided[i] = replace(r, p=p, point=point)
+        gates[fold] = {"t_block": t_block, "t_quiet": t_quiet, "training_rows": len(train), "normalised_training_rows": len(inner)}
+    return [decided[i] for i in range(len(preds))], gates
+
+
+def family_floor(rows: Sequence[Prediction], *, floor: int = FAMILY_FLOOR) -> dict[str, dict[str, Any]]:
+    """Per family, its held-out ``BLOCK`` flags among ``rows`` (the primary population, every fold): it may block only
+    when none is wrong and there are at least ``floor``; otherwise its flags are reported as ADVISORY."""
+    out: dict[str, dict[str, Any]] = {}
+    for family in sorted({r.family for r in rows}):
+        flagged = [r for r in rows if r.family == family and r.point == "BLOCK"]
+        errors = sum(1 for r in flagged if r.label == 0)
+        out[family] = {"flagged": len(flagged), "errors": errors, "floor": floor, "may_block": errors == 0 and len(flagged) >= floor}
+    return out
+
+
+def apply_floor(rows: Sequence[Prediction], floors: Mapping[str, Mapping[str, Any]]) -> list[Prediction]:
+    """``BLOCK`` becomes ``ADVISORY`` in every family that does not pass :func:`family_floor`."""
+    return [replace(r, point="ADVISORY") if r.point == "BLOCK" and not floors.get(r.family, {}).get("may_block") else r for r in rows]
+
+
 def _at(rows: Sequence[Prediction], level: str) -> dict[str, float | None]:
     flagged = [r for r in rows if r.point in (("BLOCK",) if level == "block" else ("BLOCK", "ADVISORY"))]
     positives = sum(r.label for r in rows)
@@ -408,7 +500,8 @@ def candidate_from_example(
 
 
 __all__ = [
-    "BOUNDS", "PARSE_FAILURE_LIMIT", "TARGET_PRECISION", "TARGET_RECALL", "Points", "Prediction", "calibrated_block",
+    "BOUNDS", "FAMILY_FLOOR", "PARSE_FAILURE_LIMIT", "TARGET_PRECISION", "TARGET_RECALL", "Points", "Prediction", "apply_floor",
+    "calibrated_block", "family_floor", "normalised", "pooled_block_gate", "pooled_threshold", "quiet_threshold",
     "candidate_from_example", "choose_points", "cluster_bootstrap", "evaluate", "explain", "floor_status", "incompatible",
     "learning_curve", "nested_points", "nested_precision_block", "precision_threshold", "predict", "strata", "summary", "targets",
     "unevaluable_families",

src/openultrasast/learn/experiments.py

@@ -44,7 +44,20 @@ import yaml
 from ..plane.budget import BudgetExhausted, MeteredClient
 from ..plane.memory import MemoryStore, row_id
 from .compile import load_program, spec_of
-from .evaluate import BOUNDS, TARGET_PRECISION, Prediction, calibrated_block, candidate_from_example, nested_precision_block, wilson_lower
+from .evaluate import (
+    BOUNDS,
+    FAMILY_FLOOR,
+    TARGET_PRECISION,
+    TARGET_RECALL,
+    Prediction,
+    apply_floor,
+    calibrated_block,
+    candidate_from_example,
+    family_floor,
+    nested_precision_block,
+    pooled_block_gate,
+    wilson_lower,
+)
 from .examples import EXAMPLE_KIND, Example
 from .folds import Fold, compile_split, folds_digest, outer_folds
 from .program import Caller, Decision, Program, ProgramSpec
@@ -60,12 +73,17 @@ RUN_TASK = "run"
 ARM_KINDS = ("program", "plane", "rule")
 # decision-rule experiments (kind: rule): one replay-only scoring pass of the incumbent programs, arms differ only in
 # the BLOCK rule applied to the same scores
-RULES = ("calibrated_block", "precision_bound_block")
+RULES = ("calibrated_block", "precision_bound_block", "pooled_block_gate")
+NORMALISATIONS = ("platt",)  # pooled_block_gate: how the scores of different families are put on one scale (Req 6.4)
+RULE_PARAMETERS: Mapping[str, set[str]] = {
+    "precision_bound_block": {"precision", "bound", "min_flagged", "select_on"},
+    "pooled_block_gate": {"precision", "recall", "floor", "normalise", "select_on"},
+}
 POPULATIONS = ("paired", "pooled")  # paired: units whose candidate has both sides present; pooled: every unit
 SCORE_ARM = "score"  # the arm name of a rule experiment's outcome rows (the shared scores)
 MISS_VERDICT = "replay_miss"  # an outcome row of a unit excluded because a response was not cached
 # rule metrics, per family and population of held-out (outer-fold) units: direction only (no paired test)
-RULE_METRICS: Mapping[str, int] = {"flagged_precision_lb": 1, "flagged_precision": 1, "coverage": 1, "flagged": 1}
+RULE_METRICS: Mapping[str, int] = {"flagged_precision_lb": 1, "flagged_precision": 1, "coverage": 1, "flagged": 1, "quiet_recall": 1}
 UNITS = ("candidate", "file", "repository")
 PAIRINGS = ("paired", "assigned")
 # O'Brien-Fleming boundaries for two equally spaced looks at alpha 0.05, two-sided (design section 6).
@@ -277,17 +295,24 @@ def _check_rule_manifest(path: Path, raw: Mapping[str, Any], arms: Mapping[str,
         name = arm.rule.get("name")
         if name not in RULES:
             raise ExperimentError(f"{path}: arm {arm.name} rule.name {name!r} (one of {', '.join(RULES)})")
-        allowed = {"name"} | ({"precision", "bound", "min_flagged", "select_on"} if name == "precision_bound_block" else set())
+        allowed = {"name"} | RULE_PARAMETERS.get(str(name), set())
         if set(arm.rule) - allowed:
             raise ExperimentError(f"{path}: arm {arm.name} rule parameters {sorted(set(arm.rule) - allowed)} are not understood")
         if arm.rule.get("bound", "wilson") not in BOUNDS or arm.rule.get("select_on", "paired") not in POPULATIONS:
             raise ExperimentError(f"{path}: arm {arm.name} bound in {BOUNDS} and select_on in {POPULATIONS}")
+        if name == "pooled_block_gate":
+            if arm.rule.get("normalise") not in NORMALISATIONS:
+                raise ExperimentError(f"{path}: arm {arm.name} normalise in {NORMALISATIONS} (declared before the run)")
+            if int(arm.rule.get("floor", FAMILY_FLOOR)) < 1 or not 0 < float(arm.rule.get("recall", TARGET_RECALL)) <= 1:
+                raise ExperimentError(f"{path}: arm {arm.name} floor >= 1 and recall in (0, 1]")
     populations = dict(raw.get("populations") or {})
     if populations.get("primary") not in POPULATIONS or any(p not in POPULATIONS for p in populations.get("secondary", [])):
         raise ExperimentError(f"{path}: populations.primary (and secondary) in {POPULATIONS}")
     decision = dict(raw.get("decision") or {})
     if not 0 < float(decision.get("target_precision", 0)) <= 1 or decision.get("arm") not in arms:
         raise ExperimentError(f"{path}: decision.target_precision in (0, 1] and decision.arm naming an arm")
+    if decision.get("pooled") and arms[str(decision["arm"])].rule.get("name") != "pooled_block_gate":
+        raise ExperimentError(f"{path}: decision.pooled judges a pooled_block_gate arm only")
 
 
 # --- register ------------------------------------------------------------------------------------------------------------
@@ -1087,6 +1112,17 @@ def _apply_rule(arm: Arm, preds: Sequence[Prediction], seed: int) -> tuple[list[
     if arm.rule["name"] == "calibrated_block":
         return calibrated_block(preds, seed=seed)
     select_on = str(arm.rule.get("select_on", "paired"))
+    if arm.rule["name"] == "pooled_block_gate":  # the family floor is applied by analyse_rule, on held-out flags
+        rows, gates = pooled_block_gate(
+            preds, eligible=(lambda r: bool(r.pair)) if select_on == "paired" else None,
+            precision=float(arm.rule.get("precision", TARGET_PRECISION)), recall=float(arm.rule.get("recall", TARGET_RECALL)),
+        )  # fmt: skip
+        info = {
+            f: {"t_block_by_fold": {j: g["t_block"] for j, g in gates.items()},
+                "t_quiet_by_fold": {j: g["t_quiet"].get(f) for j, g in gates.items()}}
+            for f in sorted({r.family for r in preds})
+        }  # fmt: skip
+        return rows, info
     decided, thresholds = nested_precision_block(
         preds, eligible=(lambda r: bool(r.pair)) if select_on == "paired" else None,
         precision=float(arm.rule.get("precision", TARGET_PRECISION)), bound=str(arm.rule.get("bound", "wilson")),
@@ -1095,11 +1131,38 @@ def _apply_rule(arm: Arm, preds: Sequence[Prediction], seed: int) -> tuple[list[
     return decided, {f: {"thresholds_by_fold": t} for f, t in thresholds.items()}
 
 
-def decide_rule(manifest: Manifest, families: Mapping[str, Mapping[str, Any]]) -> tuple[str, str]:
+def decide_pooled(manifest: Manifest, families: Mapping[str, Mapping[str, Any]], overall: Mapping[str, Any]) -> tuple[str, str]:
+    """The pre-registered rule of a pooled gate (Req 6.4, exp-004), on the primary population's held-out flags of
+    every family together, before the family floor: **reject** when they flag at a point precision below the target;
+    **adopt** when their Wilson 95% lower bound reaches the target and at least one family passes the floor (no
+    held-out error, at least ``floor`` flags) -- those families block, the others stay advisory; otherwise
+    **inconclusive** (the incumbent rule stays)."""
+    decision = dict(manifest.raw["decision"])
+    arm, target = str(decision["arm"]), float(decision["target_precision"])
+    primary = str(manifest.raw["populations"]["primary"])
+    cell = overall[arm][primary]
+    if cell["flagged"] and not cell["meets_target_point"]:
+        return "reject", f"{arm} flags held-out {primary} units of all families together below precision {target}"
+    passing = sorted(f for f, v in families.items() if v[arm]["floor"]["may_block"])
+    if cell["meets_target_lb"] and passing:
+        return "adopt", f"{arm}'s pooled held-out Wilson lower bound reaches {target}; {', '.join(passing)} pass the family floor and block"
+    if not cell["flagged"]:
+        return "inconclusive", f"{arm} offers no BLOCK on held-out {primary} units; the incumbent stays"
+    if not cell["meets_target_lb"]:
+        return "inconclusive", f"{arm}'s pooled held-out Wilson lower bound is below {target}; the incumbent stays"
+    return "inconclusive", f"{arm} reaches {target} pooled but no family passes the floor; every family stays advisory"
+
+
+def decide_rule(manifest: Manifest, families: Mapping[str, Mapping[str, Any]], overall: Mapping[str, Any] | None = None) -> tuple[str, str]:
     """The pre-registered rule: **reject** the decision arm when it flags held-out units of the primary population in
     some family at a point precision below the target; **adopt** it when it does not and some family's held-out
-    Wilson lower bound reaches the target; otherwise **inconclusive** (the incumbent rule stays)."""
+    Wilson lower bound reaches the target; otherwise **inconclusive** (the incumbent rule stays). A manifest whose
+    ``decision.pooled`` is set is judged by :func:`decide_pooled` instead."""
     decision = dict(manifest.raw["decision"])
+    if decision.get("pooled"):
+        if overall is None:
+            raise ExperimentError(f"{manifest.id}: a pooled decision needs the pooled-over-families totals")
+        return decide_pooled(manifest, families, overall)
     arm, target = str(decision["arm"]), float(decision["target_precision"])
     primary = str(manifest.raw["populations"]["primary"])
     cells = {f: v[arm][primary] for f, v in families.items()}
@@ -1142,6 +1205,19 @@ def analyse_rule(
         return [r for r in rows if r.pair] if population == "paired" else list(rows)
 
     decided = {name: _apply_rule(arm, preds, manifest.seed) for name, arm in manifest.arms.items()}
+    gated = {name for name, arm in manifest.arms.items() if arm.rule.get("name") == "pooled_block_gate"}
+    floors = {  # held-out flags of the primary population decide which families may block; no threshold moves
+        name: {f: family_floor([r for r in select(decided[name][0], populations[0]) if r.family == f],
+                               floor=int(manifest.arms[name].rule.get("floor", FAMILY_FLOOR))).get(f)
+               or {"flagged": 0, "errors": 0, "floor": int(manifest.arms[name].rule.get("floor", FAMILY_FLOOR)), "may_block": False}
+               for f in manifest.families}
+        for name in gated
+    }  # fmt: skip
+
+    def quiet_recall(rows: Sequence[Prediction]) -> float | None:
+        positives = [r for r in rows if r.label == 1]
+        return round(sum(1 for r in positives if r.point in ("BLOCK", "ADVISORY")) / len(positives), 4) if positives else None
+
     families: dict[str, dict[str, Any]] = {}
     for family in manifest.families:
         cell: dict[str, Any] = {}
@@ -1149,6 +1225,9 @@ def analyse_rule(
             mine = [r for r in rows if r.family == family]
             cell[name] = {**info.get(family, {}), **{p: flag_summary(select(mine, p), target, resamples=n, seed=manifest.seed)
                                                       for p in populations}}  # fmt: skip
+            if name in gated:
+                cell[name]["floor"] = floors[name][family]
+                cell[name]["quiet_recall"] = {p: quiet_recall(select(mine, p)) for p in populations}
         mine = [r for r in preds if r.family == family]
         folds = sorted({r.fold for r in mine})
         cell["capacity"] = {
@@ -1160,7 +1239,12 @@ def analyse_rule(
         name: {p: flag_summary(select(rows, p), target, resamples=n, seed=manifest.seed) for p in populations}
         for name, (rows, _) in decided.items()
     }
-    verdict, reason = decide_rule(manifest, families)
+    after_floor = {  # what users would get: BLOCK only in the families that pass the floor (secondary, not deciding)
+        name: {p: flag_summary(select(apply_floor(decided[name][0], floors[name]), p), target, resamples=n, seed=manifest.seed)
+               for p in populations}
+        for name in sorted(gated)
+    }  # fmt: skip
+    verdict, reason = decide_rule(manifest, families, overall)
     spend = {SCORE_ARM: round(sum(o.usd or 0.0 for o in outcomes if o.arm == SCORE_ARM), 6)}
     return {
         "experiment": manifest.id, "manifest_digest": manifest.digest, "hypothesis": manifest.hypothesis,
@@ -1172,7 +1256,8 @@ def analyse_rule(
         "populations": {"primary": populations[0], "secondary": populations[1:]}, "target_precision": target,
         "min_flagged_for_bound": min_flagged_for_bound(target), "primary": [m.as_dict() for m in manifest.primary],
         "secondary": [m.as_dict() for m in manifest.secondary], "families": families, "overall": overall, "resamples": n,
-        "decision": verdict, "reason": reason, "spend_usd": spend, **dict(provenance or {}),
+        **({"overall_after_floor": after_floor} if after_floor else {}), "decision": verdict, "reason": reason, "spend_usd": spend,
+        **dict(provenance or {}),
     }  # fmt: skip
 
 

tests/test_learn_rules.py

@@ -4,13 +4,24 @@ today's calibrated BLOCK rule, scored once from the response cache at zero cost
 from __future__ import annotations
 
 import json
+from dataclasses import replace
 from pathlib import Path
 
 import pytest
 from test_learn_experiments import CODE, EXAMPLES, PRICES, _json_shape, artifact, committed, example_rows, excerpt_text, paid, put_program
 
 from openultrasast.learn import experiments as ex
-from openultrasast.learn.evaluate import Prediction, calibrated_block, nested_precision_block, precision_threshold, wilson_lower
+from openultrasast.learn.evaluate import (
+    Prediction,
+    apply_floor,
+    calibrated_block,
+    family_floor,
+    nested_precision_block,
+    pooled_block_gate,
+    precision_threshold,
+    quiet_threshold,
+    wilson_lower,
+)  # fmt: skip
 from openultrasast.learn.program import Candidate, Program
 from openultrasast.learn.retrieve import Target
 from openultrasast.plane.memory import FileStore
@@ -245,3 +256,123 @@ def test_the_registered_exp003_manifest_loads() -> None:
     assert manifest.kind == "rule" and manifest.budget_total_usd == 0 and manifest.raw["populations"]["primary"] == "paired"
     assert manifest.arms["A"].rule == {"name": "calibrated_block"} and manifest.arms["B"].rule["bound"] == "wilson"
     assert len(manifest.families) == 6 and set(manifest.programs) == set(manifest.families)
+
+
+# --- the pooled BLOCK gate with a family floor (Req 6.4 amended 2026-10-03; exp-004) -------------------------------------------
+
+LOW_FLOORS = {"per_family_floor": 2, "pooled_floor": 1}  # small synthetic families still get a Platt map
+
+
+def two_scales(per_fold: int = 12, *, error: tuple[int, str] | None = None) -> list[Prediction]:
+    """Two families on different raw scales: injection positives at 0.9, negatives at 0.1; path at 0.7 and 0.3. With
+    ``error = (fold, family)`` one negative of that fold and family scores like a positive and says vulnerable."""
+    rows = []
+    for f in range(5):
+        for family, hi, lo in (("injection", 0.9, 0.1), ("path", 0.7, 0.3)):
+            for i in range(per_fold):
+                key = f"{family}{f}-{i}"
+                bad = error == (f, family) and i == 0
+                rows.append(pred(f, 1, hi, family=family, group=f"org/{key}", verdict="vulnerable", pair=key))
+                verdict = "vulnerable" if bad else "not_vulnerable"
+                rows.append(pred(f, 0, hi if bad else lo, family=family, group=f"org/{key}", verdict=verdict, pair=key))
+    return rows
+
+
+def test_one_pooled_threshold_blocks_where_no_family_alone_reaches_the_bound() -> None:
+    rows = two_scales()  # 48 training positives per family: fewer than the 73 the bound needs, 96 together
+    _, alone = nested_precision_block(rows)
+    assert all(t is None for fam in alone.values() for t in fam.values())
+    decided, gates = pooled_block_gate(rows, **LOW_FLOORS)
+    assert all(g["t_block"] is not None for g in gates.values())
+    blocked = [r for r in decided if r.point == "BLOCK"]
+    assert {r.family for r in blocked} == {"injection", "path"} and all(r.label == 1 for r in blocked)
+    assert sum(r.label for r in blocked) == sum(r.label for r in rows)  # every positive, on both raw scales
+
+
+def test_no_held_out_label_moves_a_threshold_or_the_map() -> None:
+    """Fold outer-0's labels are flipped (and its scores kept): its t_block, t_quiet and normalised scores must not move,
+    while the other folds, which train on outer-0, do see the change."""
+    rows = two_scales()
+    flipped = [replace(r, label=1 - r.label) if r.fold == "outer-0" else r for r in rows]
+    decided, gates = pooled_block_gate(rows, **LOW_FLOORS)
+    again, regates = pooled_block_gate(flipped, **LOW_FLOORS)
+    assert regates["outer-0"] == gates["outer-0"]
+    assert [r.p for r in decided if r.fold == "outer-0"] == [r.p for r in again if r.fold == "outer-0"]
+    assert [r.point for r in decided if r.fold == "outer-0"] == [r.point for r in again if r.fold == "outer-0"]
+    assert any(regates[f"outer-{f}"] != gates[f"outer-{f}"] for f in range(1, 5))  # the test has teeth
+
+
+def test_a_family_with_one_held_out_error_cannot_block() -> None:
+    rows = two_scales(16, error=(2, "path"))
+    decided, _ = pooled_block_gate(rows, **LOW_FLOORS)
+    floors = family_floor(decided, floor=10)
+    assert floors["path"]["errors"] == 1 and floors["path"]["flagged"] >= 10 and floors["path"]["may_block"] is False
+    assert floors["injection"] == {"flagged": 80, "errors": 0, "floor": 10, "may_block": True}
+    shipped = apply_floor(decided, floors)
+    assert not any(r.point == "BLOCK" for r in shipped if r.family == "path")
+    assert all(r.point == "ADVISORY" for r, d in zip(shipped, decided, strict=True) if d.point == "BLOCK" and r.family == "path")
+    few = [r for r in decided if r.family == "injection"][:18]  # 9 clean flags: below the floor
+    assert family_floor(few, floor=10)["injection"] == {"flagged": 9, "errors": 0, "floor": 10, "may_block": False}
+
+
+def test_quiet_threshold_keeps_ninety_percent_of_positives_and_nothing_is_silenced_blind() -> None:
+    rows = [replace(pred(0, 1, 0.5), p=v / 10) for v in range(1, 11)]
+    assert quiet_threshold(rows) == 0.2  # 9 of 10 positives at or above 0.2; 0.3 keeps 8
+    assert quiet_threshold([replace(pred(0, 0, 0.5), p=0.9)]) is None
+    lone = [pred(0, 1, 0.9, family="deserialization", group="org/lone")]  # a family with no training rows: no map, no t_quiet
+    decided, _ = pooled_block_gate([*two_scales(), *lone], **LOW_FLOORS)
+    assert decided[-1].p is None and decided[-1].point == "ADVISORY"
+    assert any(r.point == "DROP" for r in decided)  # low scores below t_quiet are silent
+
+
+POOLED_MANIFEST = RULE_MANIFEST.replace(
+    "  B: {{label: bound, rule: {{name: precision_bound_block, bound: wilson, precision: 0.95, select_on: paired}}}}\n"
+    "  C: {{label: point, rule: {{name: precision_bound_block, bound: point, precision: 0.95, min_flagged: 2, select_on: paired}}}}",
+    "  B: {{label: gate, rule: {{name: pooled_block_gate, precision: 0.95, recall: 0.9, floor: 10, normalise: platt, select_on: paired}}}}",
+).replace("decision: {{arm: B, target_precision: 0.95}}", "decision: {{arm: B, target_precision: 0.95, pooled: true}}")
+
+
+def pooled_outcomes(error: tuple[int, str] | None) -> tuple[list[ex.Outcome], list[ex.Unit]]:
+    units, outcomes = [], []
+    for r in two_scales(16, error=error):
+        u = ex.Unit(f"{r.pair}-{r.label}", r.group, r.family, r.label, r.fold, r.pair)
+        units.append(u)
+        outcomes.append(ex.Outcome(u.unit, u.group, u.family, u.label, u.pair, ex.SCORE_ARM, 0, r.s, r.verdict, 0.0))
+    return outcomes, units
+
+
+def test_the_pooled_verdict_follows_the_registered_rule(tmp_path: Path) -> None:
+    path = tmp_path / "exp-904-pooled.yaml"
+    path.write_text(POOLED_MANIFEST.format(id="exp-904-pooled", pid="0" * 64, units=tmp_path / "u.jsonl"))
+    manifest = ex.load_manifest(path)
+    assert manifest.arms["B"].rule["normalise"] == "platt"
+    outcomes, units = pooled_outcomes(None)
+    report = ex.analyse_rule(manifest, outcomes, units, resamples=20)
+    assert report["decision"] == "adopt" and report["overall"]["B"]["paired"]["meets_target_lb"]
+    assert all(report["families"][f]["B"]["floor"]["may_block"] for f in ("injection", "path"))
+    assert report["families"]["path"]["B"]["quiet_recall"]["paired"] == 1.0
+    outcomes, units = pooled_outcomes((2, "path"))
+    report = ex.analyse_rule(manifest, outcomes, units, resamples=20)
+    assert report["families"]["path"]["B"]["floor"]["may_block"] is False and report["decision"] == "adopt"
+    assert "injection" in report["reason"] and "path" not in report["reason"]
+    assert report["overall_after_floor"]["B"]["paired"]["flagged"] == report["families"]["injection"]["B"]["paired"]["flagged"]
+    text = path.read_text()
+    for bad, why in (
+        (text.replace("normalise: platt", "normalise: rank"), "normalise"),
+        (text.replace("floor: 10", "floor: 0"), "floor"),
+        (text.replace("name: pooled_block_gate", "name: precision_bound_block"), "not understood"),
+    ):
+        path.write_text(bad)
+        with pytest.raises(ex.ExperimentError, match=why):
+            ex.load_manifest(path)
+
+
+def test_the_exp004_manifest_registers_the_procedure_before_its_units() -> None:
+    path = ROOT / "plane/experiments/exp-004-pooled-block-gate.yaml"
+    if not path.exists():
+        pytest.skip("exp-004 is not written yet")
+    manifest = ex.load_manifest(path)
+    gate = manifest.arms["B"].rule
+    assert manifest.kind == "rule" and manifest.budget_total_usd == 0 and manifest.raw["decision"]["pooled"] is True
+    assert gate["name"] == "pooled_block_gate" and gate["normalise"] == "platt" and gate["floor"] == 10 and gate["precision"] == 0.95
+    assert manifest.raw["populations"]["primary"] == "paired" and set(manifest.programs) == set(manifest.families)