mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-27 22:15:33 +00:00
9a5606af6d
* feat(eval-contradictions): types + pure helpers for v0.33.0 probe Foundational module for the contradiction measurement probe (v0.33.0 plan). Pure, hermetic, no engine or LLM dependencies. Sets the wire contract for the rest of the implementation. - types.ts: schema_version + PROMPT_VERSION + TRUNCATION_POLICY constants, ProbeReport + ContradictionPair + JudgeVerdict + cache/run row shapes. - calibration.ts: Wilson 95% CI on the headline percentage with exact clamping at p=0 and p=1 (floating-point overshoot regression guard); small_sample_note when n<30. - judge-errors.ts: first-class typed error collector (Codex fix — bias guard for the silent-skip-on-throw decision); classifier maps to parse_fail/refusal/timeout/http_5xx/unknown. - severity-classify.ts: parseSeverity defaults to 'low' on garbage input; bucketBySeverity + buildHotPages (descending rank + tie-break by severity). - date-filter.ts: three-rule A1 pre-filter — same-paragraph-dual-date beats the separation rule (flip-flop case); missing dates falls through to the judge; only "both explicit AND >30d apart" actually skips. 51 hermetic tests across the four pure modules; typecheck clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval-contradictions): schema migrations + engine methods (v0.33.0) Adds the persistent surface the contradiction probe needs: two new tables plus five BrainEngine methods, mirrored cleanly across PGLite + Postgres. Migrations v51 + v52 (idempotent on both engines): - eval_contradictions_cache: composite PK on (chunk_a_hash, chunk_b_hash, model_id, prompt_version, truncation_policy) per Codex outside-voice fix; verdict JSONB; expires_at-driven TTL. - eval_contradictions_runs: one row per probe run; Wilson CI bounds, judge-error totals, source-tier breakdown, full report_json. Engine methods (interface + 2 impls each): - listActiveTakesForPages(pageIds, opts): P1 batched per-page fetch. Single WHERE page_id = ANY($1) AND active = true; replaces the O(K) loop the probe would otherwise pay per query. - writeContradictionsRun(row): M5 time-series insert; idempotent on run_id via ON CONFLICT DO NOTHING. - loadContradictionsTrend(days): M5 history read, newest first. - getContradictionCacheEntry(key): P2 cache lookup; 5-component key includes prompt_version + truncation_policy. - putContradictionCacheEntry(opts): cache upsert with TTL refresh. - sweepContradictionCache(): periodic expired-row purge. JSONB writes use sql.json() on Postgres (matches existing eval_takes_quality + raw_data patterns; not the literal-template-tag pattern banned by scripts/check-jsonb-pattern.sh). PGLite uses $N::jsonb positional binds. 17 hermetic tests on PGLite cover P1 (4 cases: empty, grouped, supersede- excludes, holder-allow-list), M5 (5 cases: write+read, idempotent run_id, newest-first, days-window, JSONB round-trip), P2 (6 cases: miss, put-get, prompt-version differs, truncation differs, upsert refreshes, sweep deletes expired). Existing 109 migrate + bootstrap tests still green. Schema mirror in pglite-schema.ts; source.sql regenerated to schema-embedded. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval-contradictions): cross-source + cost-tracker + cache wrappers Three pure-orchestration modules between the engine surface and the runner. Each is independently testable; the cache wrapper does hit the PGLite engine end-to-end since its job is to round-trip through P2. - cross-source.ts (M6): classifySlugTier maps a slug to curated/bulk/other using DEFAULT_SOURCE_BOOSTS (boost > 1.05 = curated, < 0.95 = bulk). buildSourceTierBreakdown produces the {curated_vs_curated, curated_vs_bulk, bulk_vs_bulk, other} counts; order-independent on the pair members. - cost-tracker.ts (A2 + P3): estimateUpperBoundCost for pre-flight refuse. CostTracker records judge calls (per-token-pricing per model) AND embedding calls (Codex P3 fix). Soft-ceiling semantics documented in the estimate_note string surfaced in the final report (Codex caveat: "hard ceiling" was overclaimed for token estimates). Anthropic + OpenAI pricing baked in; unknown models fall back to Haiku rates. - cache.ts (P2 wrapper): hashContent (sha256), buildCacheKey with lex-sorted (a, b) so verdicts are order-independent and key bakes in PROMPT_VERSION + TRUNCATION_POLICY (Codex outside-voice fix). JudgeCache class tracks hits/misses for the run report. Shape validation guards against corrupt rows: a cache row that doesn't parse as JudgeVerdict treats as a miss instead of crashing downstream. 40 hermetic tests across the three modules. Cache tests hit PGLite for real round-trip coverage of the new engine methods committed in C2. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval-contradictions): judge + auto-supersession + fixture-redact Three modules that together turn an LLM into a contradiction probe and its output into actionable resolutions. - judge.ts: judgeContradiction() is the single LLM call. Query-conditioned prompt (Codex outside-voice fix — the judge sees what the user asked). Holder context for take pairs (C3). UTF-8-safe truncation at maxPairChars (default 1500, --max-pair-chars overridable; C4 wire-up). C1 double-enforcement: orchestrator filters contradicts:true with confidence < 0.7 to false regardless of prompt rules. parseJudgeJSON is a 3-strategy generic parser (direct → fence-strip → trailing-comma + quote + first-{} extraction) — we don't reuse parseModelJSON because that's shape-locked to cross-modal-eval's scores payload. Refusal detection via stopReason AND text-pattern fallback. chatFn injection for hermetic tests. - auto-supersession.ts (M7): proposeResolution classifies each pair into takes_supersede / dream_synthesize / takes_mark_debate / manual_review and emits a paste-ready CLI command. Judge's hint wins on cross-slug pairs (it has semantic context); structural fallback prefers dream_synthesize when either side is a curated entity slug (companies/, people/, deals/, projects/). pairToFinding merges a pair + verdict into a ContradictionFinding. - fixture-redact.ts (T2): privacy-redacted pass for the gold fixture build. Layers PII scrubber (v0.25.0 eval-capture-scrub) + slug rewrites (people/<name> → people/alice-example, deterministic per session) + capitalized firstname-lastname detection + monetary obfuscation (multiply revenues by session salt to preserve magnitude shape). isCleanForCommit is the pre-commit safety net: blocks if any raw name or email shape survives. Audit trail records every redaction made. 60 hermetic tests. Judge tests use direct chatFn stub (cleaner than module-level transport seam for one-shot wrapper). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval-contradictions): trends + runner orchestrator (v0.33.0) The heart of the probe — runner.ts ties every prior module together, trends.ts writes one row per run to eval_contradictions_runs and produces the trend chart for the CLI `trend` sub-subcommand. runner.ts: - Pair generation: cross-slug across top-K results (same-slug skipped) + intra-page chunk-vs-take via P1 batched listActiveTakesForPages. - A1 date pre-filter wired: pairs separated by >30 days skip without judge calls; same-paragraph-dual-date overrides separation rule (flip-flop case sees the judge). - A3 deterministic sampling: combined_score DESC, slug-lex tiebreaker, stable across re-runs. - A2 soft budget ceiling: pre-flight estimate refuses without --yes; mid-run cumulative cost stops the run and emits a partial report. - P2 cache integration: lookup before judge call, store after; hit/miss counters drive the cache stats block in the report. - C2 first-class judge_errors: every throw counted via the typed collector, surfaced in report.judge_errors with the no-silent-skip `note` field. - Wilson CI on the headline percentage; small_sample_note when n<30. - source_tier_breakdown + hot_pages aggregated across all findings. - AbortSignal propagation for cancellation mid-run. - PreFlightBudgetError exported as a discriminable rejection class. - Hermetic via judgeFn + searchFn dependency injection — runner tests stub both without ever touching the real gateway or hybridSearch. trends.ts: - writeRunRow flattens a ProbeReport into the eval_contradictions_runs row shape, including Wilson CI bounds + duration_ms. - loadTrend reads back as typed TrendRow[]. - renderTrendChart produces a fixed-width ASCII bar chart; empty input prints a friendly message naming the command to populate runs. 41 new hermetic tests on PGLite (15 trends, 26 runner). Full eval-contradictions suite at 194/194 across 13 files. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval-contradictions): CLI + eval dispatch + mini fixture (v0.33.0) User-facing surface: `gbrain eval suspected-contradictions [run|trend|review]`. Engine-required sub-subcommand, dispatched via the existing eval.ts pattern (matches `replay`). Run mode: --queries-file FILE | --query "..." | --from-capture (mutually exclusive) --top-k N=5 --judge MODEL=claude-haiku-4-5 --limit N --budget-usd N (default $5 TTY / $1 non-TTY) --yes --output FILE --max-pair-chars N=1500 --sampling deterministic|score-first --no-cache --refresh-cache --json Trend mode: --days N=30 [--json] Review mode: --severity low|medium|high --since YYYY-MM-DD A4 wired: --from-capture detects empty eval_candidates and exits 2 with hint naming GBRAIN_CONTRIBUTOR_MODE=1 / eval.capture config key. Human summary on stderr always prints Wilson CI band, judge_errors counts broken out by class, cache hit-rate, source-tier breakdown, hot pages. Partial-report warning when mid-run budget cap fires. Run-row persistence (M5) writes to eval_contradictions_runs every successful run; subsequent `trend` and `review` invocations read from there. PreFlightBudgetError surfaces as exit 1 with the calculated estimate + cap in the message — operators see the exact number to pass to --budget-usd or override with --yes. TrendRow type extended with report_json so `review` can fetch the latest run's findings without a second query. test/fixtures/contradictions-mini.jsonl: 5 redacted queries for CLI smoke. Full eval-contradictions suite: 194 hermetic tests across 13 files. Real- brain CLI smoke covered by the E2E in commit 9. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval-contradictions): doctor + MCP + synthesize integrations (M1+M2+M3) Three thin wire-ups that turn the probe's output into action surfaces: M1 (doctor): src/commands/doctor.ts adds a `contradictions` check after the eval_capture check. Reads loadContradictionsTrend(7), surfaces the latest run's headline + severity breakdown + Wilson CI band + first 3 high-severity findings with paste-ready resolution commands. ok status when no runs exist or no findings; warn when high-severity > 0. Graceful skip when the table doesn't exist yet (pre-migration brain). M3 (MCP): src/core/operations.ts adds `find_contradictions` op (scope: read, NOT localOnly — agent-callable over HTTP MCP). Params: slug (substring match), severity (low|medium|high), limit. Reads loadContradictionsTrend(30), returns the latest run's findings filtered. NOT in the subagent allowlist by design — user-initiated only, not autonomous-action surface. New FIND_CONTRADICTIONS_DESCRIPTION constant in operations-descriptions.ts. M2 (synthesize): src/core/cycle/synthesize.ts pre-fetches the latest probe findings once at phase start (loadPriorContradictionsBlock helper) and threads up to 5 highest-severity items into buildSynthesisPrompt as an informational block. Subagent sees what to reconcile when writing compiled_truth to flagged slugs. Empty trend yields empty block (existing behavior unchanged on fresh installs). Try/catch around the engine call keeps synthesize robust even when the contradiction tables don't exist yet. 11 new hermetic tests for the MCP op (registry presence, scope, empty case, slug+severity+limit filters) and the M1/M2 data-shape contracts (end-to-end runDoctor coverage deferred to commit 9's E2E because doctor calls process.exit). Full eval-contradictions suite: 226/226 across 15 test files. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval-contradictions): build-contradictions-fixture script (T2) Local-only operator script for building the privacy-redacted gold fixture used by the precision/recall test (deferred to v0.34 when probe data informs the labeling). Runs against the user's REAL brain via the local gbrain engine config; never auto-run in CI. Flow: 1. Read --queries-file (JSONL); spin up engine via loadConfig + toEngineConfig + createEngine + connectWithRetry. 2. Run the contradiction probe with --no-cache and a stubbed judgeFn that captures candidate pairs without spending tokens. 3. Interactive prompts (skipped under --non-interactive): for each candidate, the operator labels y/n/skip + severity + axis. 4. Apply the v0.33.0 fixture-redact passes (slug rewrite, name placeholders, monetary obfuscation, PII scrubber). 5. Pre-commit safety gate: every text field passes isCleanForCommit; anything that fails gets a [REDACT?] sentinel + an _operator_review marker on the JSONL line, and the script exits 1 so the operator can't accidentally commit unredacted output. Audit comment block at the top of the JSONL records every redaction the session made (slug→placeholder, name→placeholder, monetary multiplication) so reviewers can see what was changed. Usage: bun run scripts/build-contradictions-fixture.ts \\ --queries-file FILE.jsonl \\ [--top-k N] [--judge MODEL] [--max-pairs N] [--output PATH] \\ [--non-interactive] Output defaults to test/fixtures/contradictions-eval-gold.jsonl. Typecheck clean; redactor + isCleanForCommit guard tested separately in test/eval-contradictions-fixture-redact.test.ts. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * test(e2e): real-Postgres E2E for contradiction probe (v0.33.0, T1) Required-on-DATABASE_URL E2E covering Postgres-specific behavior that PGLite can't exercise. Six surface areas, 12 cases total. All pass on fresh pgvector/pgvector:pg16: 1. Migrations v51 + v52 apply cleanly; both tables exist in information_schema; Wilson CI columns are REAL; composite PK on eval_contradictions_cache includes prompt_version + truncation_policy (Codex outside-voice fix pinned at the schema level). 2. JSONB round-trip on Postgres: writeContradictionsRun + loadTrend preserves nested object shapes (regression guard against the v0.12 double-encode bug class). Confirmed via jsonb_typeof = 'object', not 'string'. 3. P2 cache with real now(): lookup/upsert round-trip, expired rows hidden from lookup, sweepContradictionCache deletes them, and different prompt_version is a separate cache key. 4. M5 trend semantics: TIMESTAMPTZ ordering DESC is stable on real PG; days-window filter via ran_at >= cutoff correctly excludes/includes backdated rows. 5. find_contradictions MCP op end-to-end: empty case returns "No probe runs" note; populated case returns latest run findings with slug substring + severity filters applied. Verified locally against pgvector:pg16 on port 5434 — all 12 cases pass. Skips gracefully when DATABASE_URL is unset per gbrain E2E convention. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * v0.33.0 feat: brain-consistency probe + doctor + MCP + dream-cycle wire-up VERSION 0.32.0 → 0.33.0. package.json + CHANGELOG.md + llms-full.txt synced. Headline: gbrain learns to detect its own integrity drift. - new command: gbrain eval suspected-contradictions [run|trend|review] - new MCP op: find_contradictions(slug?, severity?, limit?) - new doctor check: contradictions (paste-ready resolution commands) - new dream-cycle hook: synthesize reads prior contradictions per slug - new schema: v51 (eval_contradictions_cache) + v52 (eval_contradictions_runs) - 6 new engine methods (listActiveTakesForPages, write/load run, P2 cache trio) Codex outside-voice review folded in: - Command name "suspected-contradictions" (was "contradictions" — describes what the tool actually does, not what it pretends to evaluate) - judge_errors first-class output (not silent stderr — biased denominator) - prompt_version + truncation_policy in cache key (prompt edits cleanly invalidate prior verdicts) - Wilson 95% CI on headline % + small_sample_note when n<30 - Query-conditioned judge prompt (sees user's query, not just two chunks) - Deterministic sampling for prevalence metric (stable cache hit-rate) Decision criterion for the bigger swing (chunk-level revises field): Wilson CI lower-bound: <5% → source-boost + recency-decay + curated pages handle the load 5-15% → operator's call >15% → plan for v0.34+ New docs: - docs/contradictions.md (architecture, severity rubric, action criteria) - docs/eval-bench.md extended (nightly cadence + trend workflow) - skills/migrations/v0.33.0.md (post-upgrade agent instructions) Full test suite green at the cut: - 226 hermetic unit tests across 15 files (eval-contradictions-*) - 12 real-Postgres E2E (DATABASE_URL=...; verified locally on pgvector:pg16) - typecheck clean - build:llms regenerated and the test/build-llms.test.ts gate passes Plan reference: ~/.claude/plans/system-instruction-you-are-working-hashed-dewdrop.md Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: regen llms-full.txt for v0.32.6 rename --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
6 lines
277 B
JSON
6 lines
277 B
JSON
{"query": "what is acme-example MRR"}
|
|
{"query": "what role does alice-example hold at acme-example"}
|
|
{"query": "did widget-co-example raise a Series A"}
|
|
{"query": "what is the canonical thesis on remote-only startups"}
|
|
{"query": "what is the latest take on AI agents in 2026"}
|