Skip to content

v5.50.0 - the knowledge layer stops guessing

Choose a tag to compare

@JosephOIbrahim JosephOIbrahim released this 15 Aug 17:43
· 507 commits to master since this release

v5.50.0 — the knowledge layer stops guessing

Wave 4 (H22 retrieval repair): one session from recon report to shipped wave —
5 Opus 4.8 legs in worktrees, adversarial crucible, merge train, 80/80 integration green.

What's in it

  • Retrieval repair (knowledge.py +459): scout now sees the node corpus
    (previously zero rows reached it); node index keyed (context, type) with a
    disambiguation list for ambiguous bare types; a type-name intent test replaces
    the old 2-token bail-out, so sentence-shaped questions route to the node path
    or an honest not-found instead of stale H21 prose — 0/25 confident-wrong
    on fresh adversarial probes; datasheets return all internal parm names +
    channels (12-label cap removed); similarity floor on the dense path; corpus
    build-stamp verified loudly at load.
  • Tool schema: knowledge_lookup gains optional context and k — additive,
    back-compat held (171/161 census controls).
  • Freshness release gate + ingest ledger (checks.py, ingest_ledger.py):
    served-corpus build stamp must match the ratified build, release-blocking;
    the per-context ingest ledger is single-writer by design.
  • helpdoc parameterized: callers choose the help-archive build (.368/.400),
    loud SystemExit on a missing archive; no ingest path hardcodes a build.
  • Gate P fork packet (harness/notes/h22/crux-ruling.md): the bookish-AST
    source ruling adversarially re-executed — regen reproduced to the integer
    (5,481 pages / 0 errors), one number corrected, four claims sharpened, and a
    staged-adoption advisory for the parser fork.
  • Harness hardening (CRX0): _template.md now mandates commit-before-receipt
    after a systemic receipt-sentinel race was caught, recorded, and repaired.

Honest state — read before relying

  • Freshness gate is RED on purpose: shipped corpus is .368, ratified build
    is .400. It stays red until the ING-DELTA re-ingest. That is the gate working.
  • P@1 0.755 vs the 0.98 campaign bar on the hybrid scout — capped by node
    entries being absent from the dense semantic index (out-of-scope surface);
    the proven fix is held as spawn S1 for wave 5.
  • suite_baseline is RED on master (R31 flat-baseline) — pre-existing,
    unrelated to this wave, queued for its own leg.

Receipts

Every claim above traces to harness/notes/receipts/W4-*.json; the crucible
verdict W4-CRUX.json is the wave's truth document — all builder numbers were
independently re-run, none inherited. Blueprint: harness/notes/h22/BLUEPRINT.md.