probe(P-D4): refit the base rate inside each bootstrap draw; the post-hoc gate does not pass

.claude/board/entries/2026-10-22-probes-pd4-pe2-pd3.md

@@ -1,6 +1,8 @@
 # 2026-10-22: probes P-D4, P-E2, P-D3 — measured
 
-**Status:** MEASURED (lab probes; no production code changed). These are the
+**Status:** MEASURED (lab probes; no production code changed). **Corrected
+the same day** (codex review of #1724): P-D4's post-hoc pass does not hold, so
+P-D4 fails as well (see point 2 below). These are the
 top three probes pre-registered in `2026-10-22-coresearch-pal256-evidence-probes.md`.
 
 ## P-E2: count each source once (`crates/causal-edge/tests/p_e2_stamp_gate_probe.rs`)
@@ -57,7 +59,7 @@ Results:
    - the top candidate has ≥ 2 sources;
    - the support flag is clear.
 
-   Brier(always-answer) − Brier(gate) = **0.0242 [0.0079, 0.0409]** at 44% coverage. This clears the 0.02 bar, but it is post-hoc, so it is weaker evidence than a pre-registered pass.
+   Brier(always-answer) − Brier(gate) = 0.0242 at 44% coverage. ~~[0.0079, 0.0409]~~: that interval held the base rate π fixed, even though π is estimated from the same blocks being resampled. **With π refitted in every draw, the interval is [−0.0062, 0.0552]. It includes 0, so the kill fires.** `post_hoc_gate_gain_is_not_established` pins this. Fixing π on a separate calibration set would be the alternative.
 3. **Label-blind control.** The pre-registered shuffled-label Brier control cannot stay silent: replacing p ≠ π with π lowers the Brier score on any labels. It is replaced by within-block selective accuracy:
 
    | labels | answered − abstained accuracy |
@@ -83,4 +85,6 @@ Results:
 - The calibration claim needs real held-out labelled outcomes, not a synthetic world.
 - CI runs `bgz-tensor` with `--lib` only, so P-D3 runs locally, not in CI.
 
-Closeout: `STATUS: measured | OUTCOME: P-E2 and P-D3 pass as registered; P-D4's registered margin is vacuous and its post-hoc gate passes on the margin alone | OPEN: separating world for P-D4, G5, real labels`
+The ablation gains in point 4 used the same fixed-π interval, so they also overstate certainty.
+
+Closeout: `STATUS: measured | OUTCOME: P-E2 and P-D3 pass as registered; P-D4 fails: the registered margin is vacuous, and the post-hoc gate's gain is not established once π's error is included | OPEN: a separating world with a separate calibration set for π, G5, real labels`

.claude/board/entries/README.md

@@ -25,7 +25,7 @@ index row, (3) no duplicate entry id. Checks 1 and 2 are deliberately
 opposite directions; the stranding this convention prevents shows up in
 exactly one of them, never both.
 
-440 entries, 2026-08-06 .. 2026-10-22.
+441 entries, 2026-08-06 .. 2026-10-22.
 
 | date | entry id | finding | file |
 |---|---|---|---|

crates/causal-edge/tests/p_d4_decision_gate_probe.rs

@@ -31,6 +31,11 @@
 //! cannot clear with 3–9 observations. The rule is killed by vacuity, and
 //! `preregistered_interval_margin_is_vacuous` pins that.
 //!
+//! CORRECTION (codex review of #1724): π is estimated from the same blocks
+//! being resampled, so it is refitted inside every bootstrap draw. With that,
+//! the post-hoc gain is 0.0242 [−0.0062, 0.0552]: the interval includes 0 and
+//! the kill fires. The earlier [0.0079, 0.0409] held π fixed.
+//!
 //! POST-HOC arm (b′), declared after that run and before any other: the
 //! margin is on evidence counted once — the top candidate needs at least 2
 //! more distinct sources than the runner-up. The other two conditions are
@@ -224,6 +229,19 @@ fn base_rate(os: &[Outcome], parity: usize) -> f64 {
     other.iter().map(|o| o.y_a).sum::<f64>() / other.len() as f64
 }
 
+/// The same base rate refitted on the drawn blocks (with multiplicity), so a
+/// bootstrap interval carries the base rate's own estimation error.
+fn base_rate_in(os: &[Outcome], blocks: &[usize], parity: usize) -> f64 {
+    let (mut s, mut n) = (0.0_f64, 0.0_f64);
+    for &blk in blocks.iter().filter(|&&b| b % 2 != parity) {
+        for o in os.iter().filter(|o| o.block == blk) {
+            s += o.y_a;
+            n += 1.0;
+        }
+    }
+    s / n.max(1.0)
+}
+
 fn brier_diff(os: &[Outcome], blocks: &[usize], pi: [f64; 2]) -> f64 {
     let (mut a, mut b, mut n) = (0.0, 0.0, 0.0);
     for &blk in blocks {
@@ -312,21 +330,31 @@ fn preregistered_interval_margin_is_vacuous() {
 }
 
 #[test]
-fn gate_beats_always_answer_on_brier_with_a_block_interval() {
+/// Codex review of #1724: with π refitted inside every bootstrap draw (π is
+/// estimated from the same blocks), the interval includes 0. The post-hoc
+/// gate therefore does NOT clear the kill condition; this pins that. The
+/// earlier fixed-π interval [0.0079, 0.0409] omitted π's estimation error.
+fn post_hoc_gate_gain_is_not_established() {
     let os = evaluate(&world(), Margin::Sources);
     let coverage = os.iter().filter(|o| o.answered).count() as f64 / os.len() as f64;
     let acc_a = os.iter().map(|o| o.y_a).sum::<f64>() / os.len() as f64;
     let pi = [base_rate(&os, 0), base_rate(&os, 1)];
-    let (d, lo, hi) = block_ci(|b| brier_diff(&os, b, pi));
+    let (_, flo, fhi) = block_ci(|b| brier_diff(&os, b, pi));
+    let (d, lo, hi) =
+        block_ci(|b| brier_diff(&os, b, [base_rate_in(&os, b, 0), base_rate_in(&os, b, 1)]));
     println!(
         "coverage {coverage:.3}, always-answer accuracy {acc_a:.3}, pi {pi:.3?}, \
-         Brier(a)-Brier(b) {d:.4} [{lo:.4}, {hi:.4}]"
+         Brier(a)-Brier(b) {d:.4}, refitted-pi interval [{lo:.4}, {hi:.4}] \
+         (fixed-pi interval [{flo:.4}, {fhi:.4}], which omits pi's error)"
     );
     assert!(
         (0.2..0.95).contains(&coverage),
         "gate is vacuous: coverage {coverage}"
     );
-    assert!(d >= 0.02 && lo > 0.0, "kill: gain {d} [{lo}, {hi}]");
+    assert!(
+        lo <= 0.0,
+        "the gain is now established: {d} [{lo}, {hi}] — re-pin P-D4"
+    );
 }
 
 #[test]