Files
waggle-os/docs/analysis/locomo-sota-evidence-drift-2026-06-30.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
moving
2026-09-02 10:10:29 +02:00

5.4 KiB
Raw Permalink Blame History

LoCoMo SOTA evidence drift — root cause + consolidation (2026-06-30)

UPDATE 2026-07-01: consolidating the evidence surfaced a deeper defect — the 87.66% number itself does not reproduce (fresh 7-lane W4 = 86.49% current / 85.19% archived substrate; stale- verdict-replay inflation). Canonical headline is now 86.49%. This doc's gitignore/side-repo root cause still stands; the number correction is in locomo-87.66-vs-85.26-integrity-2026-06-30.md.

Symptom (founder): "old benchmark results show, and the SOTA claim is not within waggle-os or hive-mind." The 87.66% LoCoMo memory-SOTA is real and verified, but looking at the canonical repos you see old numbers — the reproducible evidence isn't there.

Verdict: Confirmed. This is a recurrence of the §7.5 substrate-drift pattern, but in the evidence/results dimension, not the code dimension — with an aggravating .gitignore rule that was silently swallowing the result report.


How it was diagnosed (4-repo evidence matrix)

Repo / location LoCoMo number visible 87.66 evidence present?
waggle-os committed benchmarks/results/ April-2026 GEPA + v4v6 manifests + agentic-locomo-2026-04-25 No
waggle-os docs/ (methodology §0, arXiv .tex/.docx, WAGGLE-CORNERSTONE) 87.66 prose Claim only — no reproducible data
waggle-os on-disk benchmarks/results/memori-phase22-RESULT.md 82.21% (Phase-2.2 precursor) No — and git-ignored, never committed
OSS marolinik/hive-mind benchmarks/locomo/RESULTS.md + README badge 73.1% (N=320, Opus self-judge) No
hive-mind-test @ 05f2146 (private side repo) 87.66% (N=1540, same-judge) Yes — committed

Substrate code (the engine) is NOT the problem this time. All four SOTA-critical elements — inprocess-reranker.ts, search.ts reranker wiring, resolve-relative-date.ts / parse-date-window.ts (write-time temporal dating), raw-detail-lane.ts — are present in waggle-os main (packages/hive-mind-core/src/mind/) and in the OSS mirror. The reranker was reverse-ported from OSS to monorepo in f47ee8f (2026-06-11). (Minor: the OSS reranker sits at an older commit 974ad7b but is content-equivalent.)

Root cause (three compounding failures)

  1. .gitignore swallow. .gitignore line **/benchmarks/results/* ignores everything directly under benchmarks/results/. The June W3.3 result report was generated there and silently never committed. Older results (gepa-faza1/…, agentic-locomo-2026-04-25, manifest-v4/v5) survive only because they were force-added / committed before the rule — so the directory shows a stale snapshot.
  2. Evidence produced in a throwaway side repo. The actual 87.66 run (report + 1,540×2 answers + judgments) was produced and committed in hive-mind-test, which is a private benchmark working repo — not the product monorepo and not the public OSS repo. It was never forward-ported. This is the §7.5 "benchmark work in a side checkout is throwaway unless reverse-ported" failure mode.
  3. OSS public repo never refreshed. marolinik/hive-mind still advertises the earlier 73.1% (N=320) result in RESULTS.md + README badge; the 87.66 number was never published there even though the winning-stack code was (PR #14).

Net effect: the SOTA claim (prose) shipped to the monorepo docs, but its reproducible evidence lived only on local disk (git-ignored) + a side repo. A fresh clone of either canonical repo shows old numbers — exactly the founder's report.

Red herring: the founder pointed at D:/Projects/waggle-os-w4 (branch feature/w4-port). That worktree is 226 commits behind main and only 3 doc/lint commits ahead, all of whose content was already re-ported to main (paper via 4193e68a). It holds nothing main lacks.

The fix (this change)

A. Commit the canonical evidence into the product monorepo (done here). benchmarks/results/locomo-sota-2026-06/ now holds the two canonical reports (verbatim, with provenance headers), an INDEX.md (the previously-untracked SOTA single-source-of-truth), and a README.md reproduction recipe. The raw ~3.8 MB answers/judgments are intentionally not duplicated into the lean product repo — they're pointered to hive-mind-test + OSS.

B. Stop the silent swallow (structural drift-closure). A .gitignore negation exception re-includes benchmarks/results/locomo-sota-2026-06/** so this evidence stays committed and future canonical SOTA evidence has a non-ignored home.

C. OSS public update — prepared, founder-gated. Updating marolinik/hive-mind benchmarks/locomo/RESULTS.md + README badge from 73.1% → 87.66% is outward-facing (it pre-announces the SOTA ahead of arXiv submission, whose citation pass + endorsement are still open). Left as a go/no-go for the founder rather than pushed unilaterally.

  • Add a one-line contract to CLAUDE.md §7.5 / packages/hive-mind-core/CONTRIBUTING.md: "Benchmark SOTA evidence MUST land in benchmarks/results/locomo-sota-*/ (gitignore-excepted) in the monorepo. hive-mind-test is throwaway — forward-port the report the same arc."
  • Optionally extend scripts/oss-drift-check.sh to assert RESULTS.md headline parity between the monorepo evidence dir and the OSS benchmarks/locomo/.