3.5 KiB
LoCoMo headline integrity — 87.66% is not reproducible; canonical = 86.49% (settled 2026-07-01)
TL;DR. The published LoCoMo SOTA headline 87.66% (1350/1540) does not reproduce on a fresh
judge pass on any substrate. Fresh 7-lane W4 + fresh gpt-4.1-mini judge = 85.19% on its own
archived 2026-06-11 substrate and 86.49% on the current substrate. The inflation is at the
judgment layer (the harness's documented stale-verdict-replay bug), not substrate drift. Founder
adopted 86.49% as the canonical, reproducible, still-SOTA number (+4.54pp over Memori 81.95,
z=4.64, p<10⁻⁵). Verify: benchmarks/results/locomo-sota-2026-06/recount.mjs.
The investigation chain (each step verified)
-
Evidence not in the repos. The 87.66 report + raw answers/judgments lived only in the throwaway
hive-mind-testrepo + git-ignored local disk (.gitignore **/benchmarks/results/*); OSS showed the old 73.1%. (Seelocomo-sota-evidence-drift-2026-06-30.md.) -
Committed judgments recount low.
memori-gpt41mini-ours-judgments.jsonl@05f2146(the report's cited input) recounts to 1313/1540 = 85.26%, not 1350/87.66. Method validated: it reproduces the Mem0 (73.96%), Config-C (76.62%) and theirs-arm (82.14%) numbers exactly. -
The committed answers were the WRONG config (founder's "7-lane W4?" catch). That answers file has
raw_detail=0, importance=0, ~3098 tok— a reduced 2-lane run (distilled+semantic), not 7-lane W4. Re-judged fresh = 85.39%. The 87.66 needed the full 7-lane stack (raw-detail≈16, ~3700 tok). -
Regenerated the true 7-lane W4 (PROFILES+DATEWIN+EPISODIC+RAWDETAIL, uncapped) on the current substrate → 86.49% (1332/1540). Better than 2-lane (+1.1pp), validates the Pareto — but still 1.17pp below 87.66.
-
Substrate confound ruled out. The 87.66 (2026-06-11) ran on minds archived as
minds-pre-wave3c(2026-06-11 00:38); the current minds were rebuilt larger on 2026-06-29. Fresh 7-lane W4 on the archived original substrate = 85.19% — lower than current. So the gap is NOT substrate drift (newer substrate scores higher); 87.66 doesn't reproduce even on its own substrate. -
Cause = stale-verdict replay. The harness note (2026-06-15,
RESULT-backlog-closeout): "judge resumes by question_id and replayed stale verdicts." The original 1350-correct judgment pass included replayed/inflated verdicts; it is lost and no fresh judge (85.19 / 86.49) reproduces it. The41-judgeresume-by-linecount is the mechanism — reusing an OUT_FILE skips fresh judging.
Definitive numbers (all fresh gpt-4.1-mini judge, Memori verbatim prompt, N=1540)
| Configuration | Overall | vs Memori |
|---|---|---|
| Published claim (2026-06-11) | 87.66% (1350) | +5.71pp — unreproducible |
| Archived 2026-06-11 substrate, 7-lane W4 | 85.19% (1312) | +3.24pp |
| Current substrate, 7-lane W4 (CANONICAL) | 86.49% (1332) | +4.54pp, z=4.64 |
| 2-lane committed file, fresh judge | 85.39% | +3.44pp |
Resolution (2026-07-01)
- Adopt 86.49% as canonical. Pinned in
benchmarks/results/locomo-sota-2026-06/(report + answers + judgments +recount.mjs), git-tracked. - Correct 87.66→86.49 across all surfaces (docs, public
apps/www,BenchmarkApp, arXiv draft, OSS, memory index); recompute stats (+5.71→+4.54pp, z=4.42→4.64). - Prevent recurrence: the harness must use a fresh
OUT_TAGper run; benchmark SOTA evidence must pin substrate+answers+judgments together in-repo.hive-mind-testis throwaway.