Skip to content

v6.2.2: Forensic-8 zero-vector recall (QA2 85% All@5, LoCoMo 47.3% R@1, LME-S 90% R@1, BEAM 73.2%)

Latest

Choose a tag to compare

@samvallad33 samvallad33 released this 12 Nov 06:54
842c80b

v6.2.2 is the Forensic-8 zero-vector recall engine plus the Strata, MCP, Autopilot, and dashboard fixes that ship with it.

Zero embeddings. Zero vectors. Zero cloud model calls. Exact handles stay exact. Causal edges stay backward-only.

Measured on the merged kernel

Track Result
RULER QA1 (197K) Answer-doc RecallAny@5 98.00%, RecallAny@10 99.00%
RULER QA2 (421K, 2-hop) RecallAny@5 100%, RecallAll@5 85%, RecallAll@10 94%
LoCoMo-10 CORE Session R@5 92.45%, Turn Exact R@1 47.33%, Turn Exact R@5 65.89%
LoCoMo-10 ALL Session R@5 93.24%, Turn Exact R@1 47.93%, Turn Exact R@5 68.16%
LongMemEval-S official 470 RecallAny@1 90.00%, RecallAny@10 99.36%, NDCG@10 0.9229
BEAM 100K (355 scored) Stratified packet session Any@5 73.24%, session Any@5 70.99%, session Any@10 79.44%
BEAM categories contradiction 100%, temporal 95%, knowledge update 92.5% session Any@5

Artifacts: benchmarks/results/v6.2.2-forensic8-clean/.

What changed

  • Direct query-to-header canonical title match and top-3 hop anchors.
  • Interrogative-only temporal when, plus a scope-matched standing-rule winner.
  • User-milestone routing for order, timeline, and summary questions.
  • Deterministic version, tech-major, orthography, and abbreviation stemming.
  • Final-clause question-echo detection, named-partner answer swap, and unique speaker-prefix binding.
  • Topical preference scoring on the whole session, not only the best window.
  • Strata recall, memory packet, search, contradictions, cross-reference, curate, session context, smart ingest, autopilot, and dashboard importance now follow the same recorded-edge contracts.

Product binaries are built by the release workflow from this tag.