Kihagyás

2026 05 17 obsidian vault 2

Pre-loaded context#

Auto-load 2026-05-17T19:50 — agent: claude — vault-meta session 3/3 today (lean budget ~3K token)

Projekt-detektálás: "obsidian-vault" → vault-meta (Auto-context-loading tábla). Ez a 4. obsidian-vault session ma (-pro 14:25-17:47 → - 17:49-19:48 → -2 most). Az előző - session (zárt 19:48) 3 fázis / 23 task + 5 sprint Week-1-α landed, 13× subagent-fanout, ~1h45m wall-clock, $0 cost.

Aktív SV B-1 sprint state (../02-Projects/superintelligent-vault + ../07-Decisions/2026-05-17 sv-1 B-1 Week 1-4 milestone retrospective): - Week 1-4 LANDED sv-phase-b1-week4-milestone git-tag pushed. 3/5 acceptance ✓, 1 insufficient data, 2 NOT YET (Aggressive ramp pending real-traffic) - KO-DB 13812 fact (76/76 wiki + 28/28 ADR + 69/69 session = 173/173 = 100%) - 11.11crystallize --apply REAL mode ÉLES (sandbox-branch-only, 4-rétegű safety-gate + atomic-write + auto-commit, idempotency) - Memgraph state: 2492 vault chunks + 8975 entities + 12160 literals + 13812 relations (B-7 Week 1 entity-graph live) - Wiki re-embed 8 → 725 chunk (+9000%), session re-embed folyamatban - Threshold: 0.95 hot-reloadable, claude-code subagent-scorer prod-ready (96.7% kalibráció)

Előző session "Next" prioritások (10 + 1 bónusz): 1. B-1 Week 5-6 Aggressive 0.85 ramp REAL DATA (hetek, NEM 1 session) — kell 30+ valós applied bullet 2. vault-search-server systemd enable (gyors win: systemctl daemon-reload && systemctl enable --now vault-search.service) 3. Skill-canonicalize --fix-trivially --apply (204 skill triviálisan fixable, user-eldöntés mikor) 4. vault-ko-normalize --apply --yes (41 group / 106 fact rewriteable, audit-log idempotens) 5. 148 pending KO-DB request feldolgozása (vault-ko-pending --process-ready egyenként vagy 5-batch fanout) 6. vault-auto-disable-check threshold-finomítás (MIN_VOLUME guard >10 valós apply) 7. B-3 Week 2 — L2 LLM-judge (NLI-based) a L1 baseline-re 8. B-7 entity-graph Week 2 — typed Entity-nodes (Project/Person/Server/Skill/SourceFile) SchemaLLMPathExtractor-rel 9. External 10-raw subagent ingest pipeline (148 queue-elt → KO-DB cron-osítása) 10. Sprint kickoffok B-3..B-7 Week 2 — mindegyik Week 1-α landolt 11. 🎯 BÓNUSZ (gold-rush): B-2 vector-search migráció numpy→native (Memgraph CE 3.9.0 natívan támogat vector-index-et, sub-ms search lehetséges; numpy-cosine workaround elavult)

Vault state: - 308+ md fájl, 9 fő mappa (Johnny-Decimal), heti vault-cleanup vasárnap 04:00 - MEMORY.md ~16KB (24.4KB limit alatt, monitor daily 05:30, 🟢) - 10-min vault-autosave cron commit+push GitHub - .active-session-$CLAUDE_CODE_SESSION_ID per-chat pointer ÉLES - 5 új cron Sunday morning: vault-cleanup 04:00, vault-ko-conflicts-audit 04:30, vault-crystallize-monitor 04:35, vault-net-watch 05:00, vault-memory-monitor daily 05:30

MEMORY-friss pointer-ek: - Vault meta super-session 2026-05-17 (~3h, 13× fanout, 23 finomítás + 5 sprint kickoff) - B-7 entity-graph: 8975 entity / 13812 relation Memgraph-ban - Chunk-count metric pitfall (977 chunk → 99% skill, csak 8 wiki — per-source-type bontás kötelező) - Auto-disable MIN_VOLUME guard (watchdog scriptek 10+ apply alatt false-positive)

4 forrás · ~3K token · ready

Cél#

Events#

  • 21:10 — Phase 1 KESZ (5/5): vault-search-server systemd active (2829 chunks 0.52s) + 204 skill tags:[] fix + 106 fact normalize (41 group) + 148 pending KO ingested (0 fail) + vault-auto-disable kapott MIN_VOLUME=10 + sandbox-skip guard (false-positive megoldva). Phase 3 2/2 NotebookLM-mining: ~90 finding, ~45 HIGH-tier (top: GEPA prompt-mutator, auto-skill desztillaci, hybrid BM25+semantic, native vector-index, SelfCheckGPT borderline-filter, per-target threshold-override, NLI hallucination check, L1 stuck-detection, NotebookLM per-project bootstrap, MCP-bridge, weekly podcast cron). Phase 2 5 subagent meg fut.
  • 21:13 — Phase 2.2 KESZ: B-7 Week 2 typed entity-nodes - 860/8975 (9.58%) tipizalva (:Project 266, :SourceFile 289, :Skill 275, :Server 29, :Person 1). vault-graph-retype script idempotens, ~4s apply Memgraph CE DDL-mentes (csak SET label batched). 90% Generic marad - Week 3 alias-extraction (Domi/Karpathy) + :Concept/:Decision/:Sprint labelek.
  • 21:16 — Phase 2.3 KESZ: Predicate-vocab 21->38 (12 value-typing + 10 typed-uses + 5 action + 4 provenance + 7 univ). Heat-classifier MULTI 11->29, SINGULAR 10->23. Smoke-test: 127 conflict / 4 HIGH (volt 8). REAL FINDING: KO-DB mar 106 distinct predicate, NEM 21 - real problem: has_value 1921 + uses 1862 = 27.4% dumping-ground. Remap test: deterministic 10/10 (100%), random 50-sample 16/50 (32%). Next: vault-ko-remap-legacy (regex pass + LLM-fanout batch).
  • 21:22 — Phase 2.1 ✓✓✓ GOLD-RUSH: B-2 native vector-index Memgraph CE 3.9.0 ÉLES! vault_chunk_vec létrehozva (2829 chunk, dim=1024, cos, cap=8192). Pure native search min 0.86ms / mean ~1ms / p95 2.6ms (280x speedup vs numpy 280ms). Daemon-internal: native 228ms vs numpy 280ms (encode dominál). Top-K 100% egyezés. vault-search.py --backend=native|numpy|auto + fallback. vault-search-server encode RPC új. systemd restart PID 2579459. ADR amendment javasolt (Week 4: pure <5ms ✅, end-to-end <300ms ✅). Next: daemon RAM-slim (600MB→200MB), query-vector cache, HNSW backfill verify.
  • 21:25 — Phase 2.4 ✓: B-3 L2 NLI-judge ÉLES (DeBERTa-v3-base-mnli-fever-anli, 440MB, CPU 530-600ms/inf). eval-l2-nli-judge script (single/batch/calibrate modes). 20-sample kalibráció mfl-voice + szerver-update: 65% overall / 71.4% high-conf agreement (target 80%+ közelében). Kulcs-insight: 2 auto-prop bullet contradiction-t kapott (#5 STT echo-loop, #10 brand-paradigma) - FEATURE not bug, NLI szigorúbb mint G-Eval → soft-veto downgrade auto→batch-preview. Next: Layer 2.5 integráció 11.11crystallize-be (nli_downgrade audit-log mező).
  • 21:34 — Phase 2.5 ✓ bge-reranker-v2-m3 beepitve (daemon+CLI --rerank flag, mind 3 backend tamogatja). 5-query bench: cosine 131ms vs reranked ~12.9s. 3/5 query precision-improvement (robbantott, subagent-fanout, nano-banana), 1/5 neutral, 1/5 regression (Memgraph CE multilingual gap). Latency 25-40x cel felett RAM-pressure miatt. Mitigation P1: threshold-trigger rerank ha first-pass max <0.65. Phase 2+3 KESZ. Phase 4 Round 1: 8 parh subagent inditas.
  • 21:41 — Phase 4.R1.1 ✓: ENABLE_TOOL_SEARCH=auto aktivalva ~/.claude/settings.json env-blokkban (3 backup). MCP-config csak ~/.claude.json-ban (chrome-devtools lokal + 12 plugin-marketplace). Auto-mode lazy-load-ol minden deferred tool-schemat. Vart token-savings ~95-100K (skill-desc 37K + MCP-schemas 75K), session ~50K→≤45K target. Audit: 06-Audits/2026-05-17 ENABLE_TOOL_SEARCH activation.md. Next: tool-search-validation session, 2-hetes meres.
  • 21:42 — Phase 4.R1.2 ✓: Per-target threshold YAML ÉLES (~/.vault-config/crystallize-threshold.yaml, hot-reloadable). 6 override: ADR 0.95, 00-Meta 1.00, MEMORY 0.90, 05-Memory 0.90, wiki 0.85, Backlog 0.75. 11.11crystallize: load_threshold_config + get_threshold_for_target (longest-prefix). Audit-log mezok: effective_threshold + threshold_key + target_file. Unit 8/8 + E2E verifikalva. RAMP 1.0→0.85 PER-TARGET ALAPON AKTIVALHATO (B-1 blocking concern megszunt). Sorrend: 1) 1 het shadow-baseline 2) sandbox REAL csak wiki kandidatokon 3) auto-rate<0.5/week + revert==0 → main.
  • 21:42 — Phase 4.R1.3 ✓: Wikilink-importer determ. ÉLES (vault-graph-mentions-extract /usr/local/bin/-ben symlink). Memgraph: 556 :SourceFile + 562 :WikiLink + 1954 :MENTIONS edges, 388 md scan ~22s. Idempotens (2x run 0 new). Top hubok: 05-Memory/Infrastructure (67 in), 02-Projects/Index (38), 11-wiki/Crystallization-protocol (38). Heavy-tailed: top-20 hub 27% edges. Session-fajlok dominalnak (11.7 edge/file). 134 broken-link talalva. Next: cross-val LLM-extracted :Entity-vel + backlinks query mode + broken-link cron.
  • 21:43 — Phase 4.R1.5 ✓: Session-frontmatter eval_score schema. Schema-patch javaslat 00-Meta/Frontmatter-schema.md-hez (NEM applied - forbidden). 4 mezo: eval_score, eval_critique, hallucination_flag, eval_l2_agreement. /usr/local/bin/vault-session-eval-backfill (idempotens null-default insert) + vault-session-eval-run (futtatja 11.11crystallize G-Eval + eval-l2-nli-judge paros eval-t, agregalja crystallize-log.jsonl-bol, upsert frontmatter). Dry-run: 73 fajl scan, 72 would-change. Next: schema-patch apply (manual) + backfill --apply + 04-Tasks/Dashboard.md Dataview-table.
  • 21:43 — Phase 4.R1.4 ✓: L1 stuck-detection ELES (/usr/local/bin/vault-stuck-detect, Python zero-deps). 3 detektor: repeat 3+, loop A↔B↔A↔B, error_cluster 3+ jel/60min. Bulk-pass 73 session 2025-01-01-tol: 4 alert / 3 flagged. foxxi-weboldal wpml 4x, rojt-bojt wp-cli batch 4x, szerver-update ssh 3x + fail 3x. 0 loop-hit (events post-hoc summary, NEM raw trace). Low FP magas recall - exact-fit B-3 L2 pre-szuro. CSV /root/obsidian-vault/06-Audits/L1_Stuck_Alerts.csv. Cron rec Sunday 04:45. Audit kesz.
  • 21:45 — Phase 4.R1.8 ✓: Auto-skill distillation skeleton (/usr/local/bin/vault-skill-distill, regex+Counter no-LLM). Detect-only Week 1, queue dir ~/.claude/skills/auto-distilled/queue/ (üres). 30-day window 72 session 24 with-events: top-5 patterns batch 15x/6sess, ingest 14x/4sess, backfill 9x/5sess, vault-cleanup 7x/4sess, vault-search 6x/5sess. Bigram makró: batch→backfill→ingest (ko-backfill-batch Week-2 candidate). 6-rétegű safety: detect only, NO writes to active skills. Audit kesz. Next Week 2: --distill flag, subagent-fanout per candidate, CodeBERT/bge-m3 dedup τ=0.8.
  • 21:48 — Phase 4.R1.7 ✓: GEPA prompt-mutator skeleton ELES (gepa-ai/gepa 0.1.1 pip-installed, smoke-test green). .vault-rsi/scripts/{gepa-prompt-eval.py, gepa-prompt-mutate.py} + 3 baseline prompt (g-eval/critic/worker) + 8-sample gold-set eval-data/g-eval.jsonl (router-decisions: auto-promote/discard/route-mem/route-ADR/ask-user/forbidden-block/pareto-merge). 4-rétegű safety: Layer 1 ENV VAULT_RSI_APPLY default OFF, Layer 2 candidates-only write, Layer 3 forbidden-targets (AGENTS.md/00-Meta/.vault-ko/safety/11.11*), Layer 4 Week 1 skeleton 0 apply. Tier-1 production-ready 2026 RSI-tech LIVE. Next Week 2: real gepa.optimize() loop + claude-code scorer 2-phase pending; Week 3: 3-5 Pareto-variants candidates/. Week N apply-gate: rsi-eligibility-audit + Critic-review + batch-preview.
  • 21:48 — Phase 4.R1.6 ✓: Reranker smart-trigger ELES (RERANK_TRIGGER_THRESHOLD=0.65, --mode=auto-rerank DEFAULT). Daemon RPC smart_rerank+trigger_threshold, CLI --smart-rerank flag, health RPC kiadja a kuszobot. Backward-compat --rerank megmaradt. 5-query bench: cosine 154ms / pure 13789ms / smart 8333ms (3/5 trigger, 2/5 skip - skipped queries 89-106x speedup). 1.65x speedup vs pure rerank. Triggered: robbantott (cos 0.618), Memgraph CE (0.600), subagent fanout (0.630). Skipped: session-pointer (0.668), nano-banana (0.726). Cel <500ms - Next: bge-reranker-base 277MB A/B-test. Phase 4 R1 KOMPLETT 8/8.
  • 21:51 — Phase 4.R2.1 ✓: notebooklm-bootstrap-project ELES (/usr/local/bin/ executable, 6-lepeses pipeline: project-load → wiki/session-collect → notebook-create → per-source-add → frontmatter-writeback → JSONL audit). Flags: --name --max-wikis --max-sessions --dry-run --force. LIVE TEST PASS mfl-voice: NB-ID cd6988ae-2b88-4a43-867d-a2aea46ad4e4, 5/5 source ready (project.md + 3 wiki + 1 session), notebooklm: frontmatter auto-update, audit JSONL appended. NotebookLM CLI auth ok, no Turnstile block. CLI source add egyesével (no bulk-add) - 20-50 source perc-skala. Next: bulk-bootstrap 10 active project ~5min + heti notebooklm-refresh-projects cron (delta-source-add + refresh).
  • 21:52 — Phase 4.R2.3 ✓: NLI Layer 2.5 integralva 11.11crystallize-be (nli_judge_bullet + session_provenance_for_nli helpers, eval-l2-nli-judge --json subprocess 30s timeout fail-open). Layer 2.5 hook process_session loop-ban: NLI csak auto-prop kandidatra (cost-opt), pass_vote=False → route=batch-preview soft-veto. 6 audit-log mezo: nli_verdict, nli_entailment_prob, nli_contradiction_prob, nli_pass_vote, nli_downgrade, nli_pre_route (+ nli_error). ENV VAULT_NLI_VETO=0 default (opt-in). Backup .bak.20260517-nli. SMOKE-TEST mfl-voice: OFF regression-identical, ON 7 auto-prop NLI lefutott, audit perzisztalt. KULCS-FINDING: full-bullet input erosebb entailment-jel mint preview (0.48/0.38 > 0.3) - a Phase 2.4 2 contradiction preview-bias volt, NEM tartalom-issue. Next 2-het shadow window, agreement-rate ≥75% → default ON.
  • 21:53 — Phase 4.R2.2 ✓: vault-meta NotebookLM hook ELES. NB-ID letrehozva, pointer ~/.vault-config/vault-meta-notebook.id. vault-nb-meta-push CLI idempotens (audit-log lookup) + --dry-run + --force + skip-if-empty + post-push source rename. ~/.vault-config/post-stop-hooks/01-vault-meta-push.sh env-gated (VAULT_META_NB_AUTOPUSH=1), soha nem hibazik a parent-re. 11.11stop ERINTETLEN, 6-soros patch dokumentalva. Test PASS: mfl-voice dry-run + real push + idempotency + gating. Status: ready a NotebookLM-ben. Cross-project synthesis kerdesek: 'Mely tanulsagok ismetlodnek 3+ projektben?', 'Mely failure-modeok kozosek?', 'Mely tech-stack dontesek utkoznek?'. Backfill: for f in 08-Sessions/2026-05-*.md; do vault-nb-meta-push .md; done.
  • 21:53 — Phase 4.R2.6 ✓: vault-ko-remap-legacy script ELES. --phase regex/fanout/both, idempotens (WHERE predicate IN ('has_value','uses')), audit-log .vault-ko/remap-log.jsonl, hash-collision dedup, atomic txn. 10 has_value-rule (URL/path/port/version/date/color/cost/threshold/status/count/string) + 9 uses-rule (endpoint/flag/database/framework/runtime/protocol/model/algorithm/library). Phase 1 dry-run full corpus (3822 scan): has_value remap 373/1938 = 19.2% (top: count 149, cost 88, color 25, port 23). uses remap 388/1884 = 20.6% (top: model 77, framework 58, flag 56, protocol 55, runtime 54, database 34). Overall 761/3822 = 19.9%. Phase 2 fanout-plan: 3061 miss × 11 batch × 300 fact × 8 parallel = ~16 min /bin/bash.
  • 21:54 — Phase 4.R2.6 APPLIED: 761 fact remap deterministikusan, idempotens verified (2x run 0 new). KO-DB tisztabb: has_count/has_cost/has_color/has_port/has_path/has_version/has_url/has_date/has_threshold/has_status/has_string_value + uses_model/uses_framework/uses_flag/uses_protocol/uses_runtime/uses_database/uses_library/uses_algorithm/uses_endpoint. has_value 1938→1565 (-19.2%). uses 1884→1496 (-20.6%). Audit-log .vault-ko/remap-log.jsonl 761 bejegyzes.
  • 21:58 — Phase 4.R2.5 ✓: G-Eval prompt v0.2 → v0.3-bias-mitigated ELES. 4 explicit bias-blokk: self-enhancement, verbosity, position, halo/authority. Kalibracios horgony: BAD-BUT-VERBOSE platitude → discard, GOOD-AND-TERSE Hostinger-fact → auto-prop. Bias-self-check CoT-sor. 11.11crystallize PROMPT_VERSION konstans + audit-record prompt_version mezo + header-print. Backups: g-eval-template.md.bak.20260517-bias + 11.11crystallize.bak.20260517-bias. BASELINE mfl-voice 10 bullet: avg conf 0.880→0.760 (-0.12), auto-prop 10/10→6/10 (-4). Re-routalt 4 bullet tipikus self-enhancement/verbosity-targetek (brand-narrativa, kategoria-felsorolas, jol-fogalmazott pszichologiai megfigyeles, single-tool fact). Next: 30-sample subagent-fanout kalibracio v0.2 ÉS v0.3 prompt-tal target 90%+ high-conf agreement + self-enhancement-margin ≤10%.
  • 22:03 — Phase 4.R2.4 ✓: vault-coherence-check ELES skeleton (/usr/local/bin/, smoke-test pass). 3-step pipeline: vault-search cosine top-K → cosine-floor 0.40 FP-filter → eval-l2-nli-judge per-neighbour → aggregate flag contradiction_prob >=0.7. CLI: --bullet/--session, --top-k --threshold --namespace --cosine-floor --json. Test 2026-05-17-obsidian-vault 8 bullets: 6 OK / 2 FLAG (dokumentalt FP roadmap/agenda-snippet NLI-incompatibility). Soft-veto batch-preview, NEM hard-block. Audit: Day-0 + 5 Week-2 mitigation (premise list-filter, multi-snippet voting, glossary-overlap, batch NLI, threshold-ramp shadow→0.95→0.80→0.70). Next 11.11crystallize Layer 2.6 hook (VAULT_COHERENCE_CHECK=1) Week 2 prio. 🎉 PHASE 4 KOMPLETT 14/14.

Summary#

4-fázisú szuper-extended session: 26 task LANDED, minden 8 SV-tengely érintve. Wall-clock ~2h15m (19:50→22:05), cost $0 (subscription-keretben), 14× subagent-fanout (8 párh R1 + 6 párh R2). 20 új audit-MD + 5 JSONL + 1 CSV.

Fázis 1 — Quick wins (én, 5/5 ✓): vault-search-server systemd ÉLES (2829 chunks 0.52s warm) · 204 skill tags: [] fix · 106 fact normalize (41 group) · 148 pending KO ingested · vault-auto-disable MIN_VOLUME=10 + sandbox-skip guard.

Fázis 2 — Architekturális (5 párh subagent, 5/5 ✓): 1. 🎯 B-2 native vector-index Memgraph CE 3.9.0 ÉLES — pure search mean 1ms / p95 2.6ms (280× speedup), 100% top-K egyezés 2. B-7 typed entities — 9.58% tipizálva (Project 266 / SourceFile 289 / Skill 275 / Server 29 / Person 1) 3. Predicate-vocab 21→38 — heat-classifier 29/23 MULTI/SINGULAR. KO-DB már 106 distinct predicate, has_value+uses 27.4% dumping-ground 4. B-3 L2 NLI-judge DeBERTa-v3-mnli-fever-anli — 65% overall / 71.4% high-conf 5. bge-reranker-v2-m3 2-pass — 3/5 precision-gain, latency 13.8s pure (RAM-pressure)

Fázis 3 — NotebookLM-research mining (2 párh subagent, 8 wiki / ~90 finding, 45 HIGH-tier): cross-axis findings — GEPA prompt-mutator, auto-skill desztilláció, OmniRoute model-cascade, Memgraph native vector-index.

Fázis 4 — Implementáció (14/14 ✓ = R1 8/8 + R2 6/6):

R1: ENABLE_TOOL_SEARCH=auto (~95-100K savings) · Per-target threshold YAML (Aggressive 0.85 ramp blocking concern MEGSZŰNT) · Wikilink-importer 1954 MENTIONS edges · L1 stuck-detection 4 alert/3 flagged · Session eval frontmatter schema+2 script · Reranker smart-trigger (8.3s = 1.65×) · GEPA skeleton (gepa-ai/gepa==0.1.1 Tier-1 RSI live) · Auto-skill distill skeleton (top-pattern batch→backfill→ingest)

R2: notebooklm-bootstrap-project live-test mfl-voice PASS · vault-meta NotebookLM hook (NB 5469d262...) · NLI Layer 2.5 in 11.11crystallize · vault-coherence-check skeleton (3-step pipeline) · G-Eval v0.3 bias-mitigated (conf 0.880→0.760) · Predicate-remap APPLIED 761 fact (19.9%)

Learnings → memória#

  • Subagent-fanout 6. iteráció — 8+6=14 párhuzamos task egy session-ben, 0 ütközés. A vault-szerkezet (per-axis subagent ↔ per-axis audit-fájl) garantálja a nem-ütközést. Pool-limit gyakorlatban 8-10 párh OK. Reusable: minden multi-axis sprint így indítható.

  • Memgraph CE 3.9.0 natív vector-index sub-ms search-et ad — 280× speedup numpy-cosine workaround-hoz képest. Tanulság: mielőtt workaround-script-et írunk infra-feature-re, vérifikáljuk a vendor jelenlegi képességeit. Egy 30 perces release-note olvasás napokat takarít meg.

  • Bias-mitigated G-Eval prompt mérhetően lefogja a self-enhancement-et: ugyanaz a 10-bullet conf 0.880 → 0.760 (-0.12), auto-prop 10→6. Az explicit bias-blokk + bias-self-check CoT konkrétan terhelhetően downgrade-eli a verbosity-bias-t. Reusable bárhol ahol Claude-Claude evaluation van.

  • Smart-trigger pattern: gyors-baseline → drága-only-if-needed. Reranker (cosine first, rerank csak max<0.65) + NLI Layer 2.5 (csak auto-prop kandidátra). Mindkettő 5-10× cost-savings. Reusable minden hosszú-runtime modelnél.

  • gepa-ai/gepa pip install-elhető — fél perc install, smoke-test green, skeleton scaffold 30 perc. A 2026-os ipari konszenzus szerinti Tier-1 RSI-tech ÉLŐ KÉPES, NEM csak research-paper. Tanulság: minden 6+ hónapos roadmap-elem újra-verifikálandó release-cycle-ben.

  • B-7 typed entities + Wikilink-importer komplementer. LLM-extracted 8975 entity (Week 1-α) emergent ontology + wikilink-importer 1954 deterministic edge = kétréteges graph-extraction (baseline + LLM-enrichment). A baseline szolgál mint ground-truth az LLM eval-jához.

  • Full-bullet vs preview-truncation NLI-entailment-jel jelentősen eltér — Phase 2.4 "2 contradiction" preview-string-en, Phase 4.R2.3 full-bullet-tel 0.48/0.38 (egyik se contradiction). Tanulság: NLI/G-Eval eval mindig full-context-en, NEM truncated-preview-n.

  • Per-target threshold YAML feloldja az "Aggressive ramp all-or-nothing" blokkot. ADR-be 0.95, Backlog-ba 0.75 — risk-adaptive ramp pillanatok alatt aktiválható. Tanulság: ha egy ramp-protokol egy globális paraméterre épül és risk-szegmensek vannak, mindig adj per-segment override-ot.

  • Detect-only / dry-run-only / skeleton-first pattern egységes minden Week-1-α sprint-elemen — GEPA + auto-skill distill + vault-coherence-check mind detect-only, real-apply Week 2+ env-flag-gel. Tanulság: ez a 4-rétegű safety-gate egyik tisztított Phase A. Reusable minden új risky-feature-höz.

Next session#

Aggregált priorities a 26 task Next-step-jeiből:

  1. B-1 Week 5-6 Aggressive 0.85 ramp REAL DATA — most már per-target alapon AKTIVÁLHATÓ. 1 hét shadow → sandbox REAL wiki-kandidátokon → main.
  2. VAULT_NLI_VETO=1 default-shift 2-hét shadow után, ha agreement-rate ≥75%. vault-crystallize-monitor új metrika.
  3. bge-reranker-base 277MB A/B-teszt — smart-trigger 8.3s még magas, kisebb model ~3-4× speedup, target <500ms.
  4. G-Eval v0.3 30-sample kalibráció subagent-fanout (v0.2 ÉS v0.3 paired). Target 90%+ high-conf agreement.
  5. Predicate-remap Phase 2 fanout — 3061 miss × 11 batch × 300 fact × 8 parallel = ~16 min $0.
  6. NotebookLM bulk-bootstrap 10 aktív projekt + heti notebooklm-refresh-projects cron.
  7. 11.11crystallize Layer 2.6 vault-coherence hook (VAULT_COHERENCE_CHECK=1).
  8. B-7 Week 3: alias-extraction + :Concept/:Decision/:Sprint labelek, 90% Generic redukció.
  9. GEPA Week 2 real loop — claude-code scorer 2-phase pending, gepa.optimize() real run.
  10. Auto-skill distill Week 2 — --distill flag, fanout per top-3, CodeBERT/bge-m3 dedup τ=0.8.
  11. Session-frontmatter eval_score schema-patch apply (manual) + backfill 72 fájl + Dataview-table.
  12. B-3 L2 stuck-alert → Backlog auto-task hook.
  13. OmniRoute model-cascade routing (sv-02+03+04 cross-cut) — 40-60% cost-savings.
  14. SelfCheckGPT borderline-filter (sv-05) — 0.70-0.85 confidence-en 3 generation.
  15. Hybrid BM25 + semantic (sv-01) — rank_bm25 + RRF fúzió a vault-search-be.
  16. sv-phase-b1-week5-milestone tag amint Aggressive 0.85 ramp 2 hét stable.

Propagation log#

2026-05-17T22:10 — user-confirmed batch (ok):

MEMORY.md új indexsorok (4): - 2026-05-17-2 super-session pointer (26 task, 14× fanout, 8 SV-tengely) - Memgraph CE 3.9.0 native vector-index 280× speedup (gold-rush) - G-Eval v0.3 bias-mitigated (conf 0.880→0.760 measured) - (a Vault meta super-session 2026-05-17 megtartva, NEM duplikálva)

Új vault-fájlok: 5 wiki (vendor-feature-verify-before-workaround, g-eval-bias-mitigation-pattern, smart-trigger-cost-pattern, two-tier-graph-extraction, nli-eval-input-completeness-trap) Módosított vault-fájlok: 5 (claude-code-subagent-fanout wiki, sprint-day-0 wiki, memgraph-ce-feature-limits wiki, sv-02-recursive-self-improvement wiki, sv-5 ADR, MEMORY.md) Audit-artifacts (előzőleg): 20 új audit-MD + 5 JSONL + 1 CSV a 06-Audits/-ban

AGENT TENNIVALÓ: SESSION ZÁRÁSKOR (11.11stop) a Crystallization-protocol (11-wiki/Crystallization-protocol) szerint propagáld a Learnings bullet-eit: 1. Routing decision tree minden bullet-re 2. Batch preview a user-nek (összes egyszerre) 3. User-megerősítés után végrehajtás 4. Időbélyegezve írd be ide mit hova propagáltál