diff --git a/docs/analysis/attack-pass-2026-10.md b/docs/analysis/attack-pass-2026-10.md
index 00d205f9f..6bd39c8ee 100644
--- a/docs/analysis/attack-pass-2026-10.md
+++ b/docs/analysis/attack-pass-2026-10.md
@@ -472,5 +472,26 @@ evidence is the first run's 331,672 seeds with one exhaustion.
 Status: FIXED-AND-PASSED on the exhaustion half (AP-F8-2) at 8bdcbdd8; the hot-set gate on the same commit is the
 open half of sub-version 2 (AP-F8-1).
 
+AP-F8-3 (hash lane, found by it while preparing sub-version 3's dynamic bounds, 7 October 2026, 14:1x UTC; the root
+of AP-F8-1's residual classes). `accept.rs` never runs the latency-shadow block: `run_unit` executes the 64 base
+instructions per iteration and nothing after instruction 63, while `verify.rs` and every kernel run the shadow 27
+times at the end of each iteration. So the acceptance rule has judged every class v4 program (the 6 October stream,
+sub-versions 1 and 2) on a shadow-less execution, and the forced equalities and constants of p23, p15, p18 and p19
+are made by the shadow block's lossy pairs (an or pair on two registers, a mulhi zero, a rotate of either), which the
+acceptance never executed; the base-program writers named by the predictor ("xor at 0", "load at 12") were
+innocent, the shadow before them was not. Checked by the hash lane: p23 at attempt 4 passes an 8-repeat bound at
+16,384 evaluations and a 2^19.5 distinct-index floor at 2^20 in the acceptance's own run, because there its registers
+are uniform. Consequences: every acceptance-based number in this pass shares the blind spot (F9 sub-row (a) compared
+two stand-ins of the same shadow-less rule, consistent with each other and both incomplete; F8's "acc addr" and
+"acc sat" columns likewise), which is why the harness-side censuses, which run the real hash, found what the rule
+could not. Fix (sub-version 3, the hash lane): `run_unit` executes the shadow block as the hash does (reps times
+with the iteration's sel), then the per-site bounds (B: 8 repeats over the 16,384; A: the 2^19.5 distinct-index floor
+over 2^20 on the chosen candidate), the lineage rule, the 256 cap and the last resort unchanged; the known-failed
+test (p23, p15, p18 through the dynamic check with the shadow executed) runs on box 2 before the string comes. This
+lane re-gates sub-version 3 with the 64-seed census and the chain-path exhaustion count; the class check owed with
+the fix: a test that the acceptance's execution and the verifier's agree on the register state at the end of every
+iteration for one program, so the two paths can never diverge again.
+Status: FINDING-OPEN; closes with sub-version 3's re-gate.
+
 Any further finding is logged here and in `docs/fud-ledger.md` with its owning lane (hash and algorithm: fixed in
 `igneum-pow` behind a test and re-gated; node: the node lane, relay agent) before the row is marked FIXED-AND-PASSED.