mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-31 04:07:52 +00:00
cb5df8fcc51441e8d24a7179fce613cc0f12217c
15
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Garry TanandClaude Opus 4.6
|
7e98739a6b |
feat(eval): Day 10 — all.ts rewrite + llm-budget + BrainBench N tiers
Final wiring of BrainBench v1 Complete. all.ts now orchestrates the full
Cat catalog (1-12) via a mix of subprocess dispatch (Cats 1, 2, 3, 4, 6,
7, 10, 11, 12 — standalone runners with CLI entry points) and
programmatic invocation (Cats 5, 8, 9 — require runtime inputs that
can't come via CLI flags). Subprocess Cats run concurrently under a
p-limit(2) bound to cap peak memory around ~800MB (two PGLite instances
at ~400MB each).
Cats 5/8/9 show as "programmatic" in the report with a one-line
reference to their `runCatN({...})` harness API. They're deliberately
skipped from the master runner because their inputs (claim catalog,
probe catalog, scenario catalog, pre-seeded agent state, evidence
pagesBySlug) are task-specific and assembled at the caller.
**eval/runner/all.ts** — rewritten:
- CATEGORIES is a tagged union of SubprocessCategory | ProgrammaticCategory
- runCatSubprocess spawns Bun with pipe'd stdout/stderr, 10-min timeout
per Cat (124 exit + SIGTERM on timeout; no hung subprocesses)
- runConcurrently is a bounded worker pool preserving input order
- buildReport emits the full markdown with per-Cat elapsed times,
migration-noise filter, and a separate programmatic-only section
- Honors BRAINBENCH_N (1/5/10 for smoke/iteration/published),
BRAINBENCH_CONCURRENCY (default 2),
BRAINBENCH_LLM_CONCURRENCY (default 4, consumed by llm-budget)
**eval/runner/llm-budget.ts** — shared LLM rate-limit semaphore. A full
N=10 published scorecard makes ~900 Anthropic calls (150 Cat 8/9 probes
× N=10 + 100 Cat 5 claims × N=10). Without coordination, concurrent
adapters trigger 429s on per-minute limits.
- LlmBudget class: acquireSlot/releaseSlot + withLlmSlot(fn) wrapper
that releases on success AND throw (try/finally)
- getDefaultLlmBudget() singleton reads BRAINBENCH_LLM_CONCURRENCY,
falls back to 4 on missing/garbage values
- capacity enforced ≥1 (rejects 0/negative)
- Double-release is a no-op (guards against upstream double-call bugs)
- Active + waiting counts exposed for observability / tests
**package.json** scripts:
- eval:brainbench — default N=5 iteration
- eval:brainbench:smoke — N=1 for fast iteration
- eval:brainbench:published — N=10 for committed baselines
- eval:cat6 / eval:cat11 — individual new subprocess Cats
**Tests (24):** CATEGORIES catalog enforces the exact Cat-number partition
(subprocess: 1,2,3,4,6,7,10,11,12; programmatic: 5,8,9). runConcurrently
respects the cap (observable via peak in-flight counter), preserves input
order under non-uniform delays, handles empty input. LlmBudget enforces
capacity, releases on throw, honors env var, rejects 0/negative.
buildReport filters migration noise, counts passed/failed/programmatic
correctly, includes every Cat + programmatic-only section.
Full eval suite now: 314 pass, 0 fail (15 test files).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.6
|
e1e699f0a6 |
feat(eval): Day 9 — sealed qrels via PublicPage + PublicQuery at adapter boundary
Codex fixes #1, #2, #3 from the plan's outside-voice review. Enforcement shifts from SOFT-VIA-TYPE-COMMENT to SOFT-VIA-SANITIZED-OBJECT. Hard enforcement via process isolation waits for BrainBench v2 Docker sandbox. **eval/runner/types.ts** additions: - `PublicPage = Pick<Page, 'slug' | 'type' | 'title' | 'compiled_truth' | 'timeline'>` — the exact 5 fields adapters should see. No _facts. No frontmatter (a known hiding spot for accidental gold leaks). - `sanitizePage(p: Page): PublicPage` — returns a NEW object with the 5 fields only. Cannot be bypassed by `(page as any)._facts` because the field does not exist on the sanitized object. - `PublicQuery = Omit<Query, 'gold'>` — strips the gold field. - `sanitizeQuery(q: Query): PublicQuery` — enumerates public fields explicitly (not spread+delete) so no prototype weirdness leaves gold reachable. **eval/runner/multi-adapter.ts** — scoreOneRun now calls sanitizePage / sanitizeQuery before passing to adapter.init / adapter.query. The scorer retains the full Query shape (including gold.relevant) for precision / recall computation. Adapter signatures unchanged — the sealing is at the OBJECT level, not the type level. This keeps existing adapters (ripgrep-bm25, vector-only, hybrid-nograph, gbrain-after) binary-compatible. Verified: no existing adapter reads q.gold or page._facts, so the change is safe without further adapter updates. **test/eval/sealed-qrels.test.ts** (17 tests): - sanitizePage strips _facts + frontmatter + arbitrary hidden keys - Output has exactly the 5 public keys (deep introspection) - Proxy tripwire simulates a malicious adapter: any access to _facts or gold throws `sealed-qrels violation` - sanitizeQuery retains optional fields (as_of_date, tags, author, acceptable_variants, known_failure_modes) but omits undefined ones - Honest documentation of the seal's limits: filesystem bypass and Proxy attacks would still work in v1; Docker isolation (v2) is the real enforcement Every existing eval test still passes (273 before + 17 sealed-qrels = 290). Total eval suite now: 290 pass, 0 fail. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
a4cdb41b07 |
feat(eval): Day 8 — Cat 8 skill compliance + Cat 9 end-to-end workflows
**eval/runner/cat8-skill-compliance.ts** — Deterministic, judge-free Cat 8
scoring. Replays inbound signals through the agent adapter (Day 5) and
extracts four iron-law metrics directly from the tool-bridge state:
- brain_first_compliance: agent called search/get_page BEFORE producing
its final answer. Non-compliance = hallucinating from general knowledge.
- back_link_compliance: every dry_run_put_page intent has at least one
markdown [Name](slug) back-link in its compiled_truth.
- citation_format: timeline entries use canonical `- **YYYY-MM-DD** |
Source — Summary`; long final answers cite at least one slug.
- tier_escalation: simple probes use light tooling (≥1 brain call);
complex probes require ≥2 brain calls or a dry_run write when
expects_dry_run_write is set.
No judge call required — everything is computable from
`tool_bridge_state.made_dry_run_writes` + `count_by_tool` + final_answer
regex. Fast, deterministic, reproducible.
Bounded concurrency (p-limit style) worker pool at default 4 to keep
Sonnet rate limits comfortable across 100-probe batches.
**eval/runner/cat9-workflows.ts** — Rubric-graded Cat 9. 5 canonical
workflows (meeting_ingestion, email_to_brain, daily_task_prep, briefing,
sync) × ~10 scenarios each. Each scenario runs through the agent adapter,
then judge.ts scores the answer against a per-scenario rubric.
`buildEvidence(scenario, agentResult, pagesBySlug)` composes the
JudgeEvidence contract: resolves ground_truth_slugs to full
GroundTruthPage[] from a slug-map, pulls tool_call_summary directly from
tool_bridge_state (no raw tool_result content — Section-3 defense),
attaches rubric from the scenario.
Per-workflow rollup: each workflow gets its own pass_rate so the verdict
can fail one workflow without failing the whole Cat. Overall verdict
requires every populated workflow's pass_rate ≥ threshold (default 0.80)
when enableThreshold=true.
Both Cats default to verdict=baseline_only in v1 per codex fix #9: real
thresholds return after 10-probe Haiku-vs-hand-score calibration (κ > 0.7)
runs against the Day 3b amara-life-v1 corpus.
**Tests (23):** Cat 8 per-metric scorer unit tests covering every branch
(brain_first ordering, back-link compliance on mixed writes, long vs
short answer citation requirement, tier escalation for simple/complex/
writey probes, finalAnswerCiteCount dedups across syntaxes). Cat 9
buildEvidence contract shape — evidence_refs flow from agent, missing
slugs skip gracefully, no raw_transcript/tool_result leakage to judge.
Cat 9 runCat9 integration with stubbed agent + mixed-verdict judge
produces fractional pass rates correctly.
Total eval suite now: 273 pass, 0 fail.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.6
|
33e2e73940 |
feat(eval): Day 7 — Cat 5 provenance runner + structured classify_claim judge
**eval/runner/cat5-provenance.ts** — BrainBench Cat 5 scoring. Samples claims from gbrain brain-export and classifies each against its source material via a dedicated Haiku judge (classify_claim tool with a three-label enum: supported | unsupported | over-generalized). Separate from judge.ts by design: Cat 5 is a single three-way classification per claim, not a weighted rubric. Rather than overload judge.ts with a mode switch, Cat 5 has its own tool definition (CLASSIFY_CLAIM_TOOL) and prompt. The retry-once pattern, $20 cost gate semantics, and structured parsing are mirrored from judge.ts so failures look the same across Cats. Metric: `citation_accuracy` = fraction where predicted label equals gold expected_label. Threshold (informational): >0.90 per design-doc METRICS.md. v1 ships with `enableThreshold: false` so the verdict is always baseline_only — we don't have hand-authored gold claims yet, and codex flagged that threshold gating should wait until the amara-life-v1 corpus + gold file authoring lands in Day 3b. runCat5 uses a bounded-concurrency worker pool (default 4) to respect Haiku rate limits across 100+ claim batches. Evidence pages are looked up by slug from a caller-provided pagesBySlug map — missing pages don't crash, they just pass an empty source list to the judge (correct behavior for genuinely unsupported claims). **Tests (23):** classifyClaim happy/retry/fallback paths with stubbed Haiku, aggregate accuracy math, threshold gating (pass/fail vs baseline_only), runCat5 concurrency + missing-page handling, renderClaimPrompt embeds claim + sources correctly, parseClassification rejects invalid enum values + plain-text responses. Total eval suite now: 250 pass, 0 fail. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
839258dc42 |
feat(eval): Day 6 — adversarial-injections + Cat 6 prose-scale + Cat 11 multi-modal
Three modules that together cover BrainBench v1 Cat 6 (prose-scale
extraction fidelity) and Cat 11 (multi-modal ingest fidelity).
**eval/runner/adversarial-injections.ts** — 6 deterministic content
transforms shared by Cat 10 (adversarial.ts, 22 hand-crafted cases) and
Cat 6 (prose-scale variants). Each injection produces a modified content
string + a structured GoldDelta describing what the extractor MUST and
MUST NOT produce. Kinds:
- code_fence_leak — fake [X](people/fake) inside ``` fence, must NOT extract
- inline_code_slug — `people/fake` in backticks, must NOT extract
- substring_collision — "SamAI" near real `people/sam`, exactly one link
- ambiguous_role — "works with" vs "works at", downgrade type to mentions
- prose_only_mention — strip markdown link syntax, bare name → mentions only
- multi_entity_sentence — pack 4+ entities into one clause, extract all
Mulberry32 PRNG keeps variant generation deterministic under fixed seed.
Codex flagged the original plan's wording ("extract injection engine from
adversarial.ts") as overstated — adversarial.ts is a static case list,
not a reusable engine. This module is NEW code.
**eval/runner/cat6-prose-scale.ts** — Runner. Loads world-v1, applies all
6 injection kinds to sampled base pages (default 50 variants per kind ×
6 kinds = 300 variants), runs extractPageLinks on each, compares to gold
delta. Emits per-kind + overall metrics (precision, recall, F1,
code_fence_leak_rate, substring_fp_rate, pages_with_links_coverage,
mean_links_per_page). **v1 verdict is always "baseline_only"** — no
gating threshold per codex fix #9 (current extractor residuals make
>0.80 unreachable; v1 records a baseline, regression guard triggers on
drop below it).
**eval/runner/cat11-multimodal.ts** — PDF + HTML + audio runners.
Fixtures load from eval/data/multimodal/<modality>/fixtures.json
manifests; each modality skips gracefully when manifest missing or
(audio) when neither GROQ_API_KEY nor OPENAI_API_KEY is set. Metrics:
- PDF: char-level similarity via Levenshtein + optional entity_recall
- HTML: word-recall over normalized tokens (multiset semantics)
- Audio: WER (word error rate) via Levenshtein on word sequences
Fixtures are NOT committed; a future eval:fetch-multimodal script will
download them hash-verified from public sources (arXiv CC-licensed
papers, Wikipedia CC-BY-SA, Common Voice CC0).
Injectable audio transcriber (`opts.transcribe`) means tests don't need
GROQ/OpenAI keys — stubbed transcriptions exercise the WER math path
directly.
**Tests (60 new):** adversarial-injections (19) — per-kind assertions +
dispatcher coverage + slug regex conformance; cat6 (12) — variant
determinism, scoreVariant shape, aggregate per-kind + overall metrics,
corpus resolver slug rules; cat11 (29) — charSimilarity / wordRecall /
wer math, htmlToText strips scripts + decodes entities, HTML modality
with real fixtures, audio modality gracefully skips without key + uses
stub transcriber correctly.
All 60 tests pass in 48ms + 41ms.
Total eval suite now: 227 pass, 0 fail.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.6
|
05aec3feb8 |
feat(eval): Day 5 — agent adapter + judge with structured evidence contract
Two modules that together wire Cat 8 / Cat 9 / Cat 5 end-to-end scoring. **eval/runner/judge.ts** — Haiku 4.5 via tool-use `score_answer`. Input is the structured JudgeEvidence contract (fix #16 from the plan's codex review): probe + final_answer_text + evidence_refs + tool_call_summary + ground_truth_pages + rubric. Raw tool output NEVER reaches the judge — that's the Section-3 defense against paraphrased prompt-injection payloads in gold/poison.json. Retry policy: one retry on malformed tool_use response. If the second attempt is still malformed, score the probe as `judge_failed` (all scores 0, verdict=fail) so the run still completes. Aggregation: weighted mean across rubric criteria. Canonical thresholds (pass ≥3.5, partial 2.5-3.5, fail <2.5) — judge can propose a verdict but the computed verdict from the weighted mean is what the scorecard records. This prevents the model from inflating or deflating its own verdict. Score values are clamped to 0-5 on parse even if the model returns out of range. `assertNoRawToolOutput(evidence)` is a regression guard that returns the list of forbidden fields (tool_result, raw_transcript, etc.) if any leak into the evidence contract. **eval/runner/adapters/claude-sonnet-with-tools.ts** — The agent adapter. Implements `Adapter` interface minimally: `init()` spins up PGLite and seeds it, `query()` throws because the adapter is Cat 8/9-only and emits a final-answer text, not a RankedDoc[]. Retrieval scorecard stays at 4 adapters. `runAgentLoop(probeId, text, state, config)` drives the multi-turn loop: Sonnet → tool_use → tool-bridge.executeTool → tool_result → back to Sonnet. Turn cap 10. max_tokens 1024. System prompt (brain-first iron law, citation format, amara context) is cached via cache_control. Exponential backoff on rate-limit errors (1s, 2s, 4s). Emits a `Transcript` per eval/schemas/transcript.schema.json — consumed directly by recorder.ts for the flight-recorder bundle. `brain_first_ordering` classifies Cat 8's flagship metric: did the agent call search/get_page BEFORE producing the final answer? The `no_brain_calls` case (agent answers from general knowledge without ever hitting the brain) is the compliance failure to surface. ForbiddenOpError + UnknownToolError from the bridge are caught in the agent loop and surfaced as tool_result with is_error=true — keeps the loop going and preserves full audit trail for the judge. **Tests (35 new):** judge (23) — happy path, retry, fallback, evidence contract sanitization, rendered prompt does not contain raw tool_result text, verdict thresholds, score clamping, weighted mean with mixed weights, parseToolUse rejects malformed input. agent-adapter (12) — Adapter.query() throws, init() seeds PGLite, end-to-end tool loop with stubbed Sonnet, turn cap exhaustion, mutating-op rejection surfaces as tool_result error, extractSlugs regex. All 12 agent tests take ~23s because PGLite runs 13 schema migrations per test; the alternative of shared-engine-across-tests was rejected so each test is isolated. Total eval suite now: 167 pass, 0 fail. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
934d7ea4b6 |
feat(eval): Day 4 — pdf-parse + flight-recorder + tool-bridge (dry_run + expand:false)
Three infrastructure modules for BrainBench v1 Complete Cats 5/8/9/11.
**eval/runner/loaders/pdf.ts** — Thin pdf-parse wrapper. Lazy import keeps
pdf-parse out of the module-load path (avoids library debug-mode side
effects). Size cap (50MB default), encryption detection, structured error
classes (PdfEncryptedError, PdfTooLargeError, PdfParseError). Only Cat 11
multimodal will import this; production bundle never sees pdf-parse.
**eval/runner/tool-bridge.ts** — Maps 12 read-only operations from
src/core/operations.ts to Anthropic tool definitions + adds 3 dry_run write
tools. Three structural invariants enforced:
1. No hidden LLM calls. `operations.query` defaults expand=true which
routes through expansion.ts → Haiku. Bridge strips `expand` from the
query tool's input schema AND executor hard-sets expand:false. Zero
nested Haiku calls in any agent trace.
2. Mutating ops throw ForbiddenOpError. put_page, add_link, delete_page,
etc. are rejected by name. Agents record intent via dry_run_put_page /
dry_run_add_link / dry_run_add_timeline_entry which persist to the
flight-recorder without mutating the engine. This is how Cat 8's
back_link_compliance + citation_format metrics measure anything with
a read-only tool surface.
3. Poison tagged by the bridge, not the judge. Every tool result is
scanned for slugs matching gold/poison.json fixtures. Matched
fixture_ids flow into tool_call_summary.saw_poison_items for the
structured-evidence judge contract. Judge never reads raw tool
output — Section-3 defense against paraphrased prompt injections
(poison payloads never reach the judge model at all).
32K-token cap (~128K chars) with "…[truncated]" suffix.
**eval/runner/recorder.ts** — Per-run flight-recorder bundle emitter. Full
6-artifact bundle (transcript.md, brain-export.json, entity-graph.json,
citations.json, scorecard.json, judge-notes.md) when the adapter provides
an AdapterExport; 3-artifact fallback (transcript + scorecard +
judge-notes) otherwise. Atomic writes via tmp+rename. Collision-safe:
duplicate directory names get incremental -2, -3 suffix. `safeStringify`
handles circular references without throwing and JSON-serializes
Float32Array embeddings.
**package.json:** adds pdf-parse@2.4.5 as a devDependency. Scoped to eval/
use only; production gbrain binary unaffected.
**Tests:** 63 new — 30 tool-bridge, 21 recorder, 12 pdf-loader. All pass.
Fake engine uses a Proxy with `__default__` fallback so poison-matching
tests don't have to mock the exact engine method name that each operation
calls (some route via searchKeyword, others via getPage — proxy handles
both uniformly).
Total eval suite now: 132 pass, 0 fail, 923 expect() calls.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry Tan
|
53e32ee427 |
feat(eval): amara-life-v1 skeleton + Page.type enum for email/slack/cal/note
Deterministic procedural generator for the twin-amara-lite fictional-life corpus (BrainBench v1 Cat 5/8/9/11 target). 15 contacts picked from world-v1, 50 emails + 300 Slack messages across 4 channels + 20 calendar events + 8 meeting transcripts + 40 first-person notes. Mulberry32 PRNG gives byte-identical output under reseed. Plants 10 contradictions + 5 stale facts + 5 poison items + 3 implicit preferences at deterministic positions. Fixture_ids are unique across the corpus so gold/contradictions.json + gold/poison.json + gold/implicit- preferences.json can cross-reference by stable ID. PageType extended in both src/core/types.ts and eval/runner/types.ts to include email | slack | calendar-event | note (+ meeting on the production side). src/core/markdown.ts inferType() heuristics updated for the new one-slash slug prefixes (emails/em-NNNN, slack/sl-NNNN, cal/evt-NNNN, notes/YYYY-MM-DD-topic, meeting/mtg-NNNN). 17 tests cover counts (50/300/20/8/40), perturbation counts (exact 10/5/5/3), seed determinism + divergence, slug regex conformance (matches eval/runner/queries/validator.ts:131 one-slash rule), unique fixture_ids, amara-in-every-email invariant, calendar dtstart < dtend, and Amara-is- attendee on every meeting. |
||
|
Garry Tan
|
c6b308a1c8 |
fix(eval): teardown PGLite engines so bun run eval:run exits 0
The multi-adapter runner left PGLite engines alive after each run. GbrainAfterAdapter and HybridNoGraphAdapter both instantiate a PGLiteEngine in init() but never disconnect it; Bun's shutdown path exits with code 99 when embedded-Postgres workers outlive main(). Added optional `teardown?(state)` to the Adapter interface, implemented it on both engine-backed adapters, and call it from scoreOneRun after the N=5 loop. ripgrep-bm25 and vector-only hold no DB resources and don't need a teardown. Verified: gbrain-after, hybrid-nograph, ripgrep-bm25, vector-only all exit 0 at N=1. Full test:eval passes (57 tests). No metric change. |
||
|
Garry TanandClaude Opus 4.7
|
e2a5dc4a2b |
feat(eval): Phase 2 query validator + Tier 5 Fuzzy + Tier 5.5 synthetic + N=5 tolerance bands
Closes multiple Phase 2 items in one commit since they form a cohesive
package: query schema enforcement + new query tiers + per-query-set
statistical rigor.
Added:
eval/runner/queries/validator.ts — hand-rolled Query schema validator
eval/runner/queries/validator.test.ts — 24 unit tests, all pass
eval/runner/queries/tier5-fuzzy.ts — 30 hand-authored Tier 5 Fuzzy/Vibe queries
eval/runner/queries/tier5_5-synthetic.ts — 50 SYNTHETIC-labeled outsider-style queries (author: "synthetic-outsider-v1")
eval/runner/queries/index.ts — aggregator + validateAll()
Modified:
eval/runner/multi-adapter.ts — N=5 runs per adapter (BRAINBENCH_N override), page-order shuffle, mean±stddev reporting
Query validator (hand-rolled, no zod dep to match gbrain codebase style):
- Temporal verb regex enforces as_of_date (per eng pass 2 spec):
/\\b(is|was|were|current|now|at the time|during|as of|when did)\\b/i
- Validates tier enum, expected_output_type enum, gold shape per type
- gold.relevant must be non-empty slug[] for cited-source-pages queries
- abstention requires gold.expected_abstention === true
- externally-authored tier requires author field
- batch validation catches duplicate IDs
Tier 5 Fuzzy/Vibe (30 queries, hand-authored):
- Vague recall: "Someone who was a senior engineer at a biotech company..."
- Trait-based: "The engineer who pushed back on microservices"
- Cultural/epithet: "Who is known as a 'systems builder' in security?"
- Abstention bait: "Which Layer 1 project did the crypto guy leave?" (prose
mentions but never names; good systems abstain)
- Addresses Codex's circularity critique — vague queries where graph-heavy
systems shouldn't inherently win.
Tier 5.5 Synthetic Outsider (50 queries, AI-authored placeholder):
- Clearly labeled author: "synthetic-outsider-v1"
- Phrasing variety not in the 4 template families:
* fragment style ("crypto founder Goldman Sachs background")
* polite/natural ("Can you pull up what we have on...")
* comparison ("What is the difference between X and Y?")
* follow-up ("And who else advises Orbit Labs?")
* typos/misspellings ("adam lopez bioinformatcis")
* similarity ("Find me someone like Alice Davis...")
* imperative ("Pull up Alice Davis")
- Real Tier 5.5 from outside researchers supersedes synthetic via
PRs to eval/external-authors/ (docs ship in follow-up commit).
N=5 tolerance bands:
- Default N=5, override via BRAINBENCH_N env var (e.g. BRAINBENCH_N=1 for dev loops)
- Per-run seeded Fisher-Yates shuffle of page ingest order (LCG seed = run_idx+1)
- Surfaces order-dependent adapter bugs (tie-break-by-first-seen etc.)
- Reports mean ± sample-stddev per metric
- "stddev = 0" is honest signal that the adapter is deterministic, not a bug.
LLM-judge metrics (future) will naturally produce non-zero stddev.
Validation: all 80 Tier 5 + 5.5 queries pass validateAll(). 24 validator
unit tests pass.
Next commit: world.html contributor explorer (Phase 3).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
bfa8564fc8 |
feat(eval): Phase 2 EXT-3 hybrid-without-graph adapter — graph isolated
Third and closest-to-gbrain external baseline. Runs gbrain's full hybrid
search (vector + keyword + RRF fusion + dedup) WITHOUT the knowledge-graph
layer. Same engine, same embedder, same chunking, same hybrid fusion —
only traversePaths + typed-link extraction turned off.
This is the decisive comparator for "does the knowledge graph do useful
work?" Same everything-else, only graph differs. Any lead gbrain-after has
over EXT-3 is 100% attributable to the graph layer.
Files:
eval/runner/adapters/hybrid-nograph.ts — ~110 LOC
Implementation:
- New PGLiteEngine per run; auto_link set to 'false' (belt).
- importFromContent() used instead of bare putPage() so chunks +
embeddings get populated (hybridSearch needs them).
- NO runExtract() call — typed links/timeline stay empty (suspenders).
- hybridSearch(engine, q.text) answers every query. Aggregate chunks
to page-level by best chunk score.
FOUR-adapter side-by-side on 240-page rich-prose corpus, 145 relational queries:
| Adapter | P@5 | R@5 | Correct/Gold |
|-----------------|--------|--------|--------------|
| gbrain-after | 49.1% | 97.9% | 248/261 |
| hybrid-nograph | 17.8% | 65.1% | 129/261 |
| ripgrep-bm25 | 17.1% | 62.4% | 124/261 |
| vector-only | 10.8% | 40.7% | 78/261 |
The headline delta nobody can hand-wave away:
gbrain-after → hybrid-nograph = +31.4 P@5, +32.9 R@5
hybrid-nograph → ripgrep-bm25 = +0.7 P@5, +2.7 R@5
Hybrid search (vector+keyword+RRF) over pure BM25 gains ~1 point. The
knowledge graph layer over hybrid gains ~31 points. The graph is doing
the work; adding it to a retrieval stack is what actually moves the needle
on relational queries. The vector/keyword/BM25 debate is a footnote.
Timing: hybrid-nograph init is ~2 min (embeds 240 pages once); query loop
is fast. gbrain-after is ~1.5s total because traversePaths doesn't need
embeddings. Runs at ~$0.02 Opus-equivalent in embedding cost.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
633be384c9 |
feat(eval): Phase 2 EXT-2 vector-only RAG adapter
Second external baseline for BrainBench. Pure cosine-similarity ranking using the SAME text-embedding-3-large model gbrain uses internally — apples-to-apples on the embedding layer so any gbrain lead reflects the graph + hybrid fusion, not a better embedder. Files: eval/runner/adapters/vector-only.ts ~130 LOC eval/runner/adapters/vector-only.test.ts 6 unit tests (cosine math) Design: - One vector per page (title + compiled_truth + timeline, capped 8K chars). - No chunking (intentional; chunked vector RAG would be EXT-2b later). - No keyword fallback (that's EXT-3 hybrid-without-graph). - Embeddings in batches of 50 via existing src/core/embedding.ts (retry+backoff). - Cost on 240 pages: ~$0.02/run. Three-adapter side-by-side on 240-page rich-prose corpus, 145 relational queries: | Adapter | P@5 | R@5 | Correct top-5 | |---------------|--------|--------|---------------| | gbrain-after | 49.1% | 97.9% | 248/261 | | ripgrep-bm25 | 17.1% | 62.4% | 124/261 | | vector-only | 10.8% | 40.7% | 78/261 | Interesting finding: vector-only scores WORSE than BM25 on relational queries like "Who invested in X?" — exact entity match matters more than semantic similarity for these templates. BM25 nails the entity-name term; vector-only returns topically-similar-but-not-mentioning pages. This is the known failure mode of pure-vector RAG on precise relational/identity queries. Real-world vector RAG systems always add keyword fallback; EXT-3 (hybrid-without-graph) will be that fairer comparator. gbrain's lead widens in vector-only comparison: +38.4 pts P@5, +57.2 pts R@5. The graph layer is doing the heavy lifting for relational traversal; pure vector RAG can't express "traverse 'attended' edges from this meeting page." Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.7
|
629ba853ff |
feat(eval): Phase 2 adapter interface + EXT-1 ripgrep+BM25 baseline
Phase 2 credibility unlock: BrainBench now compares gbrain to external
baselines on the same corpus and queries. Transforms the benchmark from
internal ablation ("gbrain-graph beats gbrain-grep") to category comparison
("gbrain-graph beats classic BM25 by 32 pts P@5"). This is the #1 fix
from the 4-review arc — addresses Codex's core critique that v1's
before/after was self-referential.
Added:
eval/runner/types.ts — Adapter interface (v1.1 spec)
eval/runner/adapters/ripgrep-bm25.ts — EXT-1 classic IR baseline
eval/runner/adapters/ripgrep-bm25.test.ts — 11 unit tests, all pass
eval/runner/multi-adapter.ts — side-by-side scorer
Adapter interface (eng pass 2 spec):
- Thin 3-method Strategy: init(rawPages, config), query(q, state), snapshot(state)
- BrainState is opaque to runner (never inspected)
- Raw pages passed in-memory; gold/ never crosses adapter boundary
(structural ingestion-boundary enforcement)
- PoisonDisposition enum reserved for future poison-resistance scoring
EXT-1 ripgrep+BM25:
- Classic Lucene-variant IDF + k1/b tuned at standard 1.5/0.75
- Title tokens double-weighted for entity-page slug-match bias
- Stopword filter, alphanumeric tokenization, stable lexicographic tie-break
- Pure in-memory inverted index — no external deps, ~100 LOC core
First side-by-side results on 240-page rich-prose corpus, 145 relational queries:
| Adapter | P@5 | R@5 | Correct top-5 |
|---------------|--------|--------|---------------|
| gbrain-after | 49.1% | 97.9% | 248/261 |
| ripgrep-bm25 | 17.1% | 62.4% | 124/261 |
| Delta | +32.0 | +35.5 | +124 |
gbrain-after is the hybrid graph+grep config from PR #188. Ripgrep+BM25 is
a genuinely strong classic-IR baseline (BM25 is what Lucene/Elasticsearch
ship). gbrain's ~+32-point lead on relational queries reflects real work
by the knowledge graph layer: typed links + traversePaths surface the
correct answers in top-K that BM25 only pulls in via partial-text overlap.
Next in Phase 2: EXT-2 vector-only RAG + EXT-3 hybrid-without-graph
adapters. Both plug into the same Adapter interface.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
52ba00f5f8 |
feat(eval): type-accuracy runner on rich-prose corpus + wire into all.ts
New Category 2 in BrainBench: per-link-type accuracy measured directly on the
240-page rich-prose world-v1 corpus. Distinct from Cat 1's retrieval metrics,
this measures whether inferLinkType() correctly classifies extracted edges
when the prose varies (the 58% works_at and 41% advises residuals that v0.10.5
regexes targeted).
How it works:
1. Loads all pages from eval/data/world-v1/
2. Derives GOLD expected edges from each page's _facts metadata
(founders → founded, investors → invested_in, advisors → advises,
employees → works_at, attendees → attended, primary_affiliation +
role drives person-page outbound type)
3. Runs extractPageLinks() on each page → INFERRED edges
4. Per (from, to) pair, compares inferred type vs gold type
5. Emits per-link-type table: correct / mistyped / missed / spurious +
type accuracy + recall + precision + strict F1 (triple match)
6. Full confusion matrix rows=gold, cols=inferred
v0.10.5 validation on 240-page corpus (up from pre-v0.10.5 baselines):
- works_at: 58% → 100.0% (+42 pts) — 10/10 correct, 0 mistyped
- advises: 41% → 88.2% (+47 pts) — 15/17 correct
- attended: — → 100.0% 131/134 recall
- founded: 100% → 100.0% 40/40
- invested_in: 89% → 92.0% 69/75
- Overall: 88.5% → 95.7% type accuracy (conditional on edge found)
Strict F1 overall: 53.7%. Lower because the _facts-based gold set only
captures core relationships; rich prose extracts many peripheral mentions
(190 spurious "mentions" edges) that aren't bugs but are correctly-typed
prose references without a _facts counterpart. Spurious counts are signal
for future type-precision tuning, not failure.
Wired into eval/runner/all.ts as Cat 2 so every full benchmark run includes
the rich-prose type accuracy table alongside retrieval metrics.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
|
81b3f7afac |
feat: knowledge graph layer — auto-link, typed relationships, graph-query (v0.10.3) (#188)
* feat(schema): graph layer migrations v5/v6/v7 + GraphPath/health types
Schema foundation for v0.10.3 knowledge graph layer:
- v5: links UNIQUE constraint widened to (from, to, link_type) so the same
person can both works_at AND advises the same company as separate rows.
Idempotent for fresh + upgrade (drops both old constraint names first).
- v6: timeline_entries gets UNIQUE index on (page_id, date, summary) for
ON CONFLICT DO NOTHING idempotency at DB level.
- v7: drops trg_timeline_search_vector trigger. Structured timeline entries
are now graph data, not search text. Markdown timeline still feeds search
via the pages trigger. Side benefit: extraction pagination is no longer
self-invalidating (trigger used to bump pages.updated_at on every insert).
Types: new GraphPath (edge-based traversal result), PageFilters.updated_after,
BrainHealth gets link_coverage / timeline_coverage / most_connected. Postgres
schema regenerated via build:schema.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(graph): auto-link on put_page + extract --source db + security hardening
Core graph layer wired into the operation surface:
- New src/core/link-extraction.ts: extractEntityRefs (canonical extractor used
by both backlinks.ts and the new graph code), extractPageLinks (combines
markdown refs + bare-slug scan + frontmatter source, dedups within-page),
inferLinkType (deterministic regex heuristics for attended/works_at/
invested_in/founded/advises/source/mentions), parseTimelineEntries (parses
multiple date format variants from page content), isAutoLinkEnabled
(engine config flag, defaults true, accepts false/0/no/off case-insensitive).
- put_page operation auto-link post-hook: extracts entity refs from freshly
written content, reconciles links table (adds new, removes stale). Returns
auto_links: { created, removed, errors } in response so MCP callers see
outcomes. Runs in a transaction so concurrent put_page on same slug can't
race the reconciliation. Default on; opt out with auto_link=false config.
- traverse_graph operation extended with link_type and direction params.
Returns GraphPath[] (edges) when filters set, GraphNode[] (nodes) for
backwards compat. Depth hard-capped at TRAVERSE_DEPTH_CAP=10 for remote
callers; without this, depth=1e6 from MCP burns memory on the recursive CTE.
- gbrain extract <links|timeline|all> --source db: walks pages from the
engine instead of from disk. Works for live brains with no local checkout
(MCP-driven Wintermute / OpenClaw). Filesystem mode (--source fs) is
unchanged. New --type and --since filters with date validation upfront
(invalid --since used to silently no-op the filter and reprocess everything).
- Security: auto-link skipped for ctx.remote=true (MCP). Bare-slug regex
matches `people/X` anywhere in page text including code fences and quoted
strings. Without this gate an untrusted MCP caller could plant arbitrary
outbound links by writing pages with intentional slug references; combined
with the new backlink boost, attacker-placed targets would surface higher
in search.
- Postgres orphan_pages aligned to PGLite definition (no inbound AND no
outbound). Comment used to claim alignment but code disagreed; engines
drifted silently when users migrated.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(cli): graph-query command + skill updates + v0.10.3 migration file
Agent-facing surface for the graph layer:
- New `gbrain graph-query <slug>` command with --type, --depth, --direction
in|out|both. Maps to traverse_graph operation with the new filters. Renders
the result as an indented edge tree.
- skills/migrations/v0.10.3.md: agent runs this post-upgrade to discover the
graph layer. Tells the agent to run `gbrain extract links --source db`,
then timeline, verify with stats, try graph-query, and lists the inferred
link types so they can be used in subsequent traversals.
- skills/brain-ops/SKILL.md Phase 2.5: documents that put_page now auto-links.
No more manual add_link calls in the Iron Law back-linking path.
- skills/maintain/SKILL.md: graph population phase. Shows the right command
to backfill links + timeline from existing pages.
- cli.ts: register graph-query in CLI_ONLY + handleCliOnly switch. Update help
text to describe `gbrain extract --source fs|db` and the new graph-query.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(graph): unit + e2e + 80-page A/B/C benchmark for graph layer
Coverage for the v0.10.3 graph layer (260+ new test assertions):
- test/link-extraction.test.ts (46 tests): extractEntityRefs both formats,
extractPageLinks dedup + frontmatter source, inferLinkType heuristics
(meeting/CEO/invested/founded/advises/default), parseTimelineEntries
multiple date formats + invalid date rejection, isAutoLinkEnabled
case-insensitive truthy/falsy parsing.
- test/extract-db.test.ts (12 tests): `gbrain extract <links|timeline|all>
--source db` happy paths, --type filter, --dry-run JSON output,
idempotency via DB constraint, type inference from CEO context.
- test/graph-query.test.ts (5 tests): direction in/out/both, type filter,
non-existent slug, indented tree output.
- test/pglite-engine.test.ts (+26 tests): getAllSlugs, listPages
updated_after filter, multi-type links via v5 migration, removeLink with
and without linkType, addTimelineEntry skipExistenceCheck flag,
getBacklinkCounts for hybrid search boost, traversePaths in/out/both with
cycle prevention via visited array, getHealth graph metrics
(link_coverage / timeline_coverage / most_connected).
- test/e2e/graph-quality.test.ts (6 tests): full pipeline against PGLite
in-memory. Auto-link via put_page operation handler. Reconciliation
removes stale links on edit. auto_link=false config skip.
- test/benchmark-graph-quality.ts: A/B/C comparison on 80 fictional pages,
35 queries across 7 categories. Hard thresholds: link_recall > 90%,
link_precision > 95%, timeline_recall > 85%, type_accuracy > 80%,
relational_recall > 80%. Currently passing all 9.
Built test-first: benchmark caught WORKS_AT_RE matching "founder" inside
slug names (frank-founder), "worked at" past-tense missing from regex,
PGLite Date object vs ISO string comparison bug. All fixed before merge.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v0.10.3)
CHANGELOG: knowledge graph layer headline. Auto-link on every page write.
Typed relationships (works_at, attended, invested_in, founded, advises).
gbrain extract --source db. graph-query CLI. Backlink boost in hybrid search.
Schema migrations v5/v6/v7 applied automatically.
Security hardening caught during /ship adversarial review: traverse_graph
depth capped at 10 from MCP, auto-link skipped for ctx.remote=true, runAutoLink
reconciliation in transaction, --since validates dates upfront.
TODOS.md: 2 P2 follow-ups (auto-link redundant SQL on skipped writes;
extract --source db not gated on auto_link config).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: sync CLAUDE.md with v0.10.3 graph layer
Updated key files list (extract.ts now describes --source fs|db, added
graph-query.ts and link-extraction.ts), test inventory (extract-db,
link-extraction, graph-query unit tests; e2e/graph-quality), and
test count (51 unit + 7 e2e, 1151 + 105 assertions).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(v0.10.3): wire graph layer into install flow + README + benchmark
Existing brains upgrading to v0.10.3 had no clear path to backfill the new
links/timeline tables. New installs had no instruction to run extract --source db
after import. This wires the knowledge graph into every install touchpoint so the
v0.10.3 features actually reach the user.
- README: headline now sells self-wiring graph + 94% benchmark numbers; new
Knowledge Graph section between Knowledge Model and Search; LINKS+GRAPH command
block expanded; Benchmarks docs group added
- INSTALL_FOR_AGENTS.md: new Step 4.5 (graph backfill) + Upgrade section now runs
gbrain init + post-upgrade and points to migrations/v<N>.md
- skills/setup/SKILL.md Phase C: new step 5 for graph backfill (idempotent,
skip-if-empty); existing file migration becomes step 6
- src/commands/init.ts: post-init hint detects existing brain (page_count > 0)
and prints extract commands for both PGLite and Postgres engines
- docs/GBRAIN_VERIFY.md: new Check #7 (knowledge graph wired) with backfill
fallback + graph-query smoke test
- docs/benchmarks/2026-04-18-graph-quality.md: checked-in benchmark report
matching the existing search-quality format (94% recall, 100% precision,
100% relational recall, idempotent both ways)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(claude): require PR descriptions to cover the whole branch
Adds a rule to CLAUDE.md so future PR bodies always cover the full diff
against the base branch, not just the most recent commit. Includes the
git log + gh pr view incantation to check what's actually in a PR.
This is a reaction to PR #189 being created with a body that described
only the last commit instead of the 7 commits it actually contained.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(upgrade): post-upgrade prints full body + --execute mode + downstream skill upgrade doc
PR #188 review caught two install-flow gaps that this commit closes:
1. `gbrain post-upgrade` only printed the migration headline + description
from YAML frontmatter, never the markdown body that contains the
step-by-step backfill instructions. Agents saw "Knowledge graph layer —
your brain now wires itself" and had no idea to run `gbrain extract
links --source db`. Now prints the full body after the headline.
2. New `--execute` flag reads a structured `auto_execute:` list from
migration frontmatter and runs the safe commands sequentially. Without
`--yes` it prints the plan only (preview mode). With `--yes` it actually
runs them. Stops on first failure with a clear error.
3. Downstream agents (Wintermute etc.) keep local skill forks that gbrain
can't push updates to. New `docs/UPGRADING_DOWNSTREAM_AGENTS.md` lists
the exact diffs each release needs applied to those forks. v0.10.3
diffs for brain-ops, meeting-ingestion, signal-detector, enrich.
Changes:
- src/commands/upgrade.ts:
- runPostUpgrade(args) accepts flags
- Prints full body via extractBody()
- Parses auto_execute: list via extractAutoExecute() (hand-rolled, no yaml dep)
- --execute previews, --execute --yes runs
- Fix cosmetic bug: `recipe: null` no longer prints "show null" message
- src/cli.ts: pass args to runPostUpgrade
- skills/migrations/v0.10.3.md:
- Add auto_execute: list (gbrain init + extract links/timeline + stats)
- Fix typo: completion record version was 0.10.1, now 0.10.3
- test/upgrade.test.ts: 5 new tests covering body printing, plan preview,
actual execution, no-auto_execute case, and --help output
- docs/UPGRADING_DOWNSTREAM_AGENTS.md: NEW
- CLAUDE.md: key files list updated
Test: 13 upgrade tests pass (was 8, +5 new). Full unit suite: 1078 pass,
zero regressions, 32 expected E2E skips (no DATABASE_URL).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* bench(graph): add Configuration A baseline (no graph) vs C comparison
Previous benchmark showed C numbers only (94.4% link recall, 100% relational
recall, etc.) but never quantified what a pre-v0.10.3 brain actually loses.
Reviewer caught this gap.
Adds measureBaselineRelational() that simulates a no-graph fallback:
- Outgoing queries: regex-extract entity refs from the seed page content
- Incoming queries: grep-style scan of all pages for the seed slug
This is what an agent without the structured links table can do today.
Honest result on the 5 relational queries in the benchmark:
- Recall: 100% A vs 100% C (+0%) — markdown contains the refs either way
- Precision: 58.8% A vs 100.0% C (+70%) — without typed links, you get the
right answers buried in 41% noise
Per-query breakdown shows the divergence is concentrated in INCOMING queries:
"Who works at startup-0?" returns 5 candidates without graph (2 employees +
3 noise pages that mention startup-0) vs exactly 2 with graph. For an LLM
agent, that's ~3x less reading work per relational question.
Also documented what the benchmark deliberately doesn't test (multi-hop,
search ranking with backlink boost, aggregate queries, type-disagreement
queries) so future benchmark work has a roadmap.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* bench(graph): add 4 missing categories — multi-hop, aggregate, type-disagreement, ranking
The previous benchmark commit (
|