feat(jme): Phase 3 — temporal filter + vector ceiling + jme-stats - #30
Merged
Conversation
…hboard Pieza 1 — Temporal dedup in queryMemory(): - New deduplicateFacts() groups recall candidates by cosine similarity (threshold=0.85) and keeps only the most-recent fact per cluster. - Prevents contradictory v1/v3 versions of the same fact from landing in the same context window. - factEmbeddings map collected during vector search, passed to dedup. - Only activates when vector embeddings are present (FTS5-only path passes through unchanged). Pieza 2 — Vector ceiling surveillance in consolidateAll(): - Counts jme_facts with embeddings at each nightly run. - Emits console.warn when count >= VECTOR_CEILING_WARN (400). - LIMIT 500 in queryMemory() starts degrading at ~400 rows. - No pruning yet (Phase 4); this gives visibility before it becomes a bottleneck. Pieza 3 — mc-ctl jme-stats command: - New cmd_jme_stats() function in mc-ctl. - Shows: fact store counts, vector ceiling status, turn buffer depth, recall utility (was_used breakdown), daily breakdown (last 7d), baseline comparison across all sources, and verdict runway to 2026-07-27. - Registered in dispatch (jme-stats) and usage() help text. Tests: 6 new cases for deduplicateFacts() in jme.test.ts. All 33 tests pass. tsc --noEmit: zero errors.
…e newest
The anchor loop broke out on the FIRST newer candidate, leaving the rest
of the cluster un-absorbed — a v1/v2/v3 trio of the same fact surfaced
both v2 AND v3 (repro: identical embeddings, ts 100/200/300 → [v2, v3]),
defeating the feature's whole purpose ("solo v3 llega al contexto").
Now each anchor absorbs every ≥threshold member first and keeps the
single max-ts survivor at the anchor's rank position. 3-member regression
test added — 2-member cases cannot express this failure.
Relands the fix from task 595947aa, which ran in the nanoclaw sandbox and
never reached the host.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0186BhSFY9sFFHCPeFnpbvki
Owner
Author
|
Operator-session review + fix relanded ( Blocking bug (now fixed): Non-blocking notes for a future pass: the new function's docstring still references a 🤖 Generated with Claude Code |
kosm1x
pushed a commit
that referenced
this pull request
Jul 17, 2026
…losure record Records the three arcs (consolidator fix proven in prod, gh_create_pr cwd/upstream root cause + gate PASS 66.87, PR #30 forensics incl. the nanoclaw sandbox-evaporation finding) and queues the systemic follow-ups (nanoclaw work-landing check, coding-playbook repo map). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0186BhSFY9sFFHCPeFnpbvki
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
JME Phase 3 — PR #30
Three independent monitoring/quality improvements. No schema changes, no wiring changes, no new infra.
Pieza 1 — Temporal filter in queryMemory()
New deduplicateFacts() groups recall candidates by cosine similarity (TEMPORAL_DEDUP_THRESHOLD=0.85) and keeps only the most-recent fact per cluster. Prevents contradictory v1/v3 facts from appearing in the same context window.
Pieza 2 — Vector ceiling surveillance
consolidateAll() now counts embedded facts at each nightly run. Warns at >=400 (VECTOR_CEILING_WARN). LIMIT 500 in queryMemory() starts degrading at ~400 rows.
Pieza 3 — mc-ctl jme-stats dashboard
New cmd_jme_stats() command showing: fact store counts, vector ceiling status, recall utility (was_used breakdown), daily breakdown (7d), baseline comparison, and verdict runway to 2026-07-27.
Tests
6 new cases for deduplicateFacts(). All 33 tests pass. tsc: zero errors.