test: compaction storm -- an irreducible protected message re-escalates to level 8 on every request (#fwut)

docs/lanes/fwut-compaction-storm/evidence/own-captures-compaction-summary.tsv

@@ -0,0 +1,58 @@
+# session_uuid	compaction_events	over_target	at_level_8	both(storm_criterion)
+0000000000000000-040a4cf2780940a7_an	30	0	0	0
+0000000000000000-1f4bec9b6f67470e_an	1	0	0	0
+0000000000000000-27fd2a47a61e4e47_an	1	0	0	0
+0000000000000000-2ddd506f3a494b18_an	2	1	2	1
+0000000000000000-3d698bd370ac4128_ap	2	0	0	0
+0000000000000000-52c44298074d4e7b_an	4	3	3	3
+0000000000000000-5600b60010c54cb1_an	0	0	0	0
+0000000000000000-5966d52bf1834ea7_an	4	0	0	0
+0000000000000000-5e32001dea4e4c76_an	11	0	0	0
+0000000000000000-5fccece65e2c4e5c_ap	2	0	0	0
+0000000000000000-6ffd3b727ae847bc_an	2	0	0	0
+0000000000000000-760e7ff5fa164a83_ap	1	0	0	0
+0000000000000000-8236a4a70c8b444c_an	1	0	0	0
+0000000000000000-8a54f833396e4d5f_fo	1	0	0	0
+0000000000000000-a86be4664d6d4aec_an	1	0	0	0
+0000000000000000-a957109ce9fb4306_an	3	0	0	0
+0000000000000000-bf61825c1fbb45fc_an	1	0	0	0
+0000000000000000-c2180ed2049f4592_an	1	0	0	0
+0000000000000000-cbdb7bf135584005_an	1	0	0	0
+0000000000000000-d0bd9db3d2444ff6_an	1	0	0	0
+0000000000000000-ed402cdcbf02406f_an	1	0	0	0
+?	11	0	0	0
+?	18	1	2	1
+184c7503-6b2a-41aa-aadb-d43511fbc153	0	0	0	0
+1ee4a784-6ff2-4d71-94f9-916cdc56fbf1	7	2	0	0
+?	2	0	0	0
+?	20	3	5	3
+?	21	3	5	3
+?	22	1	4	1
+22df78a2-2b04-420b-8db4-a7896e3db0d7	21	11	10	6
+2c930a2b-6962-493d-8fb3-b4a5d519aac4	16	7	11	6
+?	31	2	13	2
+?	4	0	0	0
+45bb3743-13a5-43e3-bd9b-1cee398ab13f	1	0	0	0
+4d3bff85-0180-4516-96de-e9df66d5f9b2	14	4	0	0
+4e294b23-76e9-4ba1-bc31-1252b04cc76a	0	0	0	0
+?	5	0	0	0
+521af98d-1005-4bb7-90e1-5c0a7306e061	1	0	0	0
+5971952c-cd00-452d-a391-a5a35eb441fa	19	19	19	19
+5f24140d-fa97-4541-81f5-1c34a1d71cc6	11	0	0	0
+?	6	0	1	0
+?	7	0	0	0
+71d12e65-040d-417f-a8cb-c7407f48af6a	2	0	0	0
+?	8	1	3	1
+82579f6f-d888-483d-90d7-872fb9bf8b98	22	22	22	22
+829b5867-4eaf-485e-93b0-d09ef22b5094	0	0	0	0
+83eeda13-ddf3-46cc-814d-48490955ff40	2	0	0	0
+a037d79f-e957-425d-90e5-a372dd5d2af6	2	0	0	0
+a14c76b9-9b9f-4b76-9e44-e4c54c131c9d	25	25	25	25
+b790851b-5a22-40ff-8d9f-c02364a52f40	15	2	7	2
+f42aa377-bd4c-4b1a-a994-583542b4ee66	8	0	0	0
+f7ce58e8-d004-40f6-adc4-96b1c49a5a4b	17	3	12	3
+-root-val-dw-tdg-oai-03/sessions/000	1	0	0	0
+-root-val-dw-tstrip-oai-01/sessions/	1	0	0	0
+-root-val-dw-tstrip-oai-02/sessions/	2	0	0	0
+-root-val-rr2-s1-oai-02/sessions/000	1	0	0	0
+-root-val-s4-oai-terra-high-01/sessi	1	0	0	0

docs/lanes/fwut-compaction-storm/evidence/pr33_recompute.py

@@ -0,0 +1,65 @@
+"""Recompute the PR#33 primary criterion after excluding compaction-STORM sessions.
+
+Inputs are the PUBLISHED figures from _dashboard/pr33-repro-hunt.md (Part 2 global
+cells + the robustness section). Nothing is re-derived from raw turns here; this is a
+transparent sensitivity analysis on the hunt's own numbers, so it inherits their
+rounding (rates are quoted to 4 dp).
+
+Exclusion criterion, fixed BEFORE looking at the effect on the ratio and applied
+identically to both arms: a session whose `context:compaction` events are >=90%
+at strategy_level 8 (max) with after_tokens above target is a compaction STORM --
+a different defect from the standing-stale-notice PR #33 addresses. Measured
+against the graph, both arms' top outliers are 100% level-8 storms.
+"""
+
+# --- published global cells (later-notice turns) ---
+SON_TURNS, SON_RATE = 11_810, 0.2451
+OPU_TURNS, OPU_RATE = 22_267, 0.1226
+
+# --- published outliers (both verified as 100%-level-8 storms) ---
+SON_STORM_TURNS, SON_STORM_DUPS = 972, 2_188      # 445ac89c... anchors-amp-dev-architect
+OPU_STORM_TURNS, OPU_STORM_DUPS = 2_163, 2_248    # 877774a5... foundation-explorer
+
+son_dups = SON_TURNS * SON_RATE
+opu_dups = OPU_TURNS * OPU_RATE
+
+
+def rate(d, t):
+    return d / t if t else 0.0
+
+
+rows = [
+    ("as published (no exclusion)",
+     rate(son_dups, SON_TURNS), rate(opu_dups, OPU_TURNS)),
+    ("exclude sonnet storm only",
+     rate(son_dups - SON_STORM_DUPS, SON_TURNS - SON_STORM_TURNS),
+     rate(opu_dups, OPU_TURNS)),
+    ("exclude opus storm only",
+     rate(son_dups, SON_TURNS),
+     rate(opu_dups - OPU_STORM_DUPS, OPU_TURNS - OPU_STORM_TURNS)),
+    ("exclude BOTH storms (the principled cut)",
+     rate(son_dups - SON_STORM_DUPS, SON_TURNS - SON_STORM_TURNS),
+     rate(opu_dups - OPU_STORM_DUPS, OPU_TURNS - OPU_STORM_TURNS)),
+]
+
+print(f"sonnet-5 dups implied by published cell: {son_dups:,.0f}")
+print(f"opus-5   dups implied by published cell: {opu_dups:,.0f}")
+print(f"sonnet storm share of sonnet dups: {SON_STORM_DUPS / son_dups:.1%} "
+      f"(hunt says 76%)")
+print(f"opus   storm share of opus dups:   {OPU_STORM_DUPS / opu_dups:.1%} "
+      f"(hunt says 83%)")
+print()
+print(f"{'cut':<42} {'sonnet':>8} {'opus':>8} {'ratio':>8}   verdict")
+print("-" * 82)
+for label, s, o in rows:
+    r = s / o if o else float("inf")
+    if r >= 2.0:
+        v = "REPRODUCED"
+    elif r < 1.3:
+        v = "NOT-REPRODUCED"
+    else:
+        v = "INCONCLUSIVE"
+    print(f"{label:<42} {s:>8.4f} {o:>8.4f} {r:>7.2f}x   {v}")
+print()
+print("pre-registered rule: >=2.0x REPRODUCED, <1.3x NOT-REPRODUCED, else INCONCLUSIVE")
+print("hunt's own leave-top-3-sessions-out variant: 0.0488 / 0.0187 = 2.61x")

docs/lanes/fwut-compaction-storm/evidence/scan_own_captures.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+# Safe capture scan (v2, resumable). grep only -- no file parsed whole.
+ROOT=/home/bkrabach/dev/openai-evals-team-ci/.amplifier/evaluation
+OUT=/tmp/fwut/captures_scan2.tsv
+DONEF=/tmp/fwut/scanned_files.txt
+touch "$OUT" "$DONEF"
+find "$ROOT" \( -name 'events.jsonl' -o -name 'root_events.jsonl' \) -print0 2>/dev/null |
+while IFS= read -r -d '' f; do
+  grep -qxF "$f" "$DONEF" && continue
+  c=$(grep -c '"context:compaction"' "$f" 2>/dev/null || true)
+  c=${c:-0}
+  echo "$f" >> "$DONEF"
+  [ "$c" -eq 0 ] 2>/dev/null && continue
+  stats=$(grep -o '"message_count": *[0-9]\+' "$f" 2>/dev/null | grep -o '[0-9]\+$' |
+    sort -n | awk '{a[n++]=$1} END{if(n==0){print "0\t0\t0"; exit} print a[0]"\t"a[int(n/2)]"\t"n}')
+  printf '%s\t%s\t%s\n' "$c" "$stats" "$f" >> "$OUT"
+done
+echo "DONE files=$(wc -l < "$DONEF") with_compactions=$(wc -l < "$OUT")"