Garry Tan
c2938f62fb
The catch that mattered (testing + maintainability, independently): --llm extraction metrics were computed per fixture but never reached any scoreboard cell — dead matched_any_gold/stored_rows aggregation fields proved the unfinished wiring. Now aggregated Σ-style into the write-back cell when llm is on, pinned by a harness-level stubbed-transport test. Also: RUN-scoped --llm BudgetTracker (a per-invocation cap multiplied by fixture count — ~$550 worst case — now one tracker, exhaustion aborts the run loudly); continuity loop collapsed to one prep per pair + read-only reader per harness (the writer replay was provably observable-effect-free and orderings rebuilt byte-identical brains — 90 preps → 15, runtime ~12s → ~7s, identical scores; baseline counts updated 24 → 12 honestly); factKeywordProbe escapes ILIKE metacharacters; validator rejects unpassable slug-less retrieve-turns; ci-gate fails HARD on a broken ref instead of silently running ungated; privacy year-scan covers the gold dir; e2e tests self-sufficient (shared artifacts in beforeAll) with a minimal-env --llm gate; round4/cellKey single-sourced; explicit return in eval.ts dispatch; stale comment + type anchor fixes.
12 KiB
Evaluation Metric Glossary
Auto-generated from src/core/eval/metric-glossary.ts. Do not edit by hand. Run bun run scripts/generate-metric-glossary.ts to regenerate.
Every metric gbrain eval * and gbrain search stats reports has a plain-English explanation here. Industry terms are preserved verbatim so users searching the literature find what we report.
Retrieval Metrics
Precision at k (P@k)
Key: precision@k
Plain English: Of the top k results the engine returned, what fraction were actually relevant? High precision means few junk results in the top of the list.
Range: 0..1, higher is better. P@10 = 0.7 means 7 of the top 10 results were on-topic.
Recall at k (R@k)
Key: recall@k
Plain English: Of all the relevant results that exist in the brain, what fraction did the engine find in its top k? High recall means few missed answers.
Range: 0..1, higher is better. R@10 = 0.81 means out of every 100 questions, the right answer was in the top 10 for 81 of them.
Mean Reciprocal Rank (MRR)
Key: mrr
Plain English: On average, how far down the list is the FIRST relevant result? An MRR of 1.0 means the first hit is always right; an MRR of 0.5 means it's typically at rank 2.
Range: 0..1, higher is better. Computed as the average of 1/rank-of-first-relevant-result across all test queries.
Normalized Discounted Cumulative Gain at k (nDCG@k)
Key: ndcg@k
Plain English: Like precision@k, but the engine gets MORE credit for putting good results near the top than near rank k. A perfect ordering scores 1.0; a totally random ordering scores near 0.
Range: 0..1, higher is better. nDCG@10 above 0.65 is the common "ship it" threshold for hybrid retrieval on technical corpora.
Retrieval-Quality / Evidence Metrics (NamedThingBench)
Hit rate at 1 (Hit@1)
Key: hit@1
Plain English: Fraction of queries where the right page is the very first result. NamedThingBench hard-gates title-substring Hit@1 >= 0.95 and alias Hit@1 >= 0.98 — a query that is a page's name or title phrase should land it at rank 1, not "somewhere in the top 10".
Range: 0..1, higher is better.
Hit rate at 3 (Hit@3)
Key: hit@3
Plain English: Fraction of queries where the right page is in the top 3 results. NamedThingBench requires the multi-chunk-dilution family to hit 1.0 — a page with one strong chunk among many weak ones must never be buried.
Range: 0..1, higher is better.
Average rank-1 match score
Key: avg_rank1_score
Plain English: The mean base (pre-boost) retrieval score of the TOP result across recent searches, from gbrain search stats. It is NOT a labeled accuracy number — it is a drift signal: if this trends DOWN over time, retrieval quality is regressing (the early warning that would have caught the duplicate-page incident before a human did).
Range: 0..1. Watch the trend, not the absolute value; pair with the <0.6 / 0.6-0.85 / >=0.85 bucket counts for shape.
Create-safety hint (evidence contract)
Key: create_safety
Plain English: A result's answer to "is this page already in the brain — safe to NOT write a new one?" Derived from the strongest evidence, NOT a raw score: exists (alias_hit / exact_title_match / high_vector_match — do not duplicate), probable (solid keyword match — prefer updating), unknown (weak match — look closer). An agent keys its don't-duplicate decision off this, which is what prevents the incident's duplicate-stub class.
Range: enum: exists | probable | unknown
Set-Similarity / Stability Metrics
Jaccard similarity at k (set Jaccard @k)
Key: jaccard@k
Plain English: How much do two result lists overlap? Compare the top k slugs from the captured baseline against the current run; Jaccard@10 = 1.0 means perfect agreement, 0.0 means zero overlap.
Range: 0..1, higher = more stable. Below 0.5 on a stable corpus means retrieval changed significantly.
Top-1 stability rate
Key: top1_stability
Plain English: Fraction of queries where the #1 result is the same between two runs. The most aggressive stability check — small ranking shifts that don't change the top answer don't hurt it.
Range: 0..1, higher = more stable. Above 0.85 typically means safe-to-merge for retrieval changes.
Statistical-Significance Metrics
p-value (paired bootstrap)
Key: p_value
Plain English: How likely the observed difference between two modes is just noise. Lower = stronger evidence the difference is real. We compute paired bootstrap with 10,000 resamples and Bonferroni correction across the 12 comparisons (3 modes × 4 metrics).
Range: 0..1, lower = stronger signal. Below 0.05 is the common "statistically significant" threshold; below 0.01 is strong evidence.
95% Confidence Interval (CI)
Key: confidence_interval
Plain English: The range we're 95% sure the true value falls inside, given the sample we measured. Narrower CI = more reliable estimate. Computed via bootstrap resampling.
Range: Two-tuple [low, high]. If 0 is inside the CI for a Δ, the difference isn't statistically significant.
Operational / Cost Metrics
Cache hit rate
Key: cache_hit_rate
Plain English: Fraction of searches that reused a recent cached answer instead of running fresh. Higher hit rate = lower latency + lower LLM spend, but stale results may slip through if the threshold is too loose.
Range: 0..1, higher generally better. 0.7-0.9 is the sweet spot for a busy brain; above 0.9 may indicate the similarity threshold is too loose.
Average results returned
Key: avg_results
Plain English: Mean number of search-result rows the engine returned per call. Should be near the active mode's searchLimit unless the brain is small or the budget is dropping results.
Range: 0..searchLimit. Far below searchLimit suggests budget pressure or sparse retrieval.
Average tokens delivered
Key: avg_tokens
Plain English: Estimated tokens (chars / 4) in the chunk text returned per search call. The direct measure of how much context an agent loop is paying for each search.
Range: 0..tokenBudget. Approximates OpenAI tiktoken count for English; off by ~5-10% for Anthropic and worse for non-English.
Cost per query (USD)
Key: cost_per_query_usd
Plain English: Sum of LLM + embedding API charges for one search call. Includes Haiku expansion call (tokenmax mode only) + embedding cost + downstream answer-model cost if measured.
Range: 0..unbounded. Conservative mode is typically <$0.001 per call; tokenmax with answer-gen can exceed $0.01.
p99 latency (ms)
Key: p99_latency_ms
Plain English: 99th percentile wall-clock time per search call. The latency that 1% of users see — long-tail experience, not the average.
Range: 0..unbounded. Warm-cache hits should be <50ms; tokenmax with expansion can exceed 200ms due to the Haiku call.
Result-Sizing Metrics
Autocut signal
Key: autocut.signal
Plain English: Which signal autocut used to size the result set. 'rerank' means it found a real score cliff in the cross-encoder rerank scores and cut there; 'none' means no trustworthy cliff (no reranker, <2 scored results, or the gap was too small) so it returned the full list.
Range: 'rerank' | 'none'. 'none' is not a failure — it means autocut declined to cut because the signal didn't justify it.
Autocut gap ratio
Key: autocut.gap_ratio
Plain English: The size of the largest score drop autocut found, as a fraction of the top result's score. A gap of 0.40 means the score fell by 40% of the top score at the steepest point. Autocut cuts there only when this clears the sensitivity threshold (autocut_jump, default 0.20).
Range: 0..1, higher = a sharper cliff (more confident cut). Below the autocut_jump threshold → no cut.
BrainBench — Cross-Harness Memory Conformance
Know-to-ask failure rate (BrainBench)
Key: know_to_ask_failure_rate
Plain English: Of the conversation turns where memory SHOULD have surfaced something unprompted, the fraction where nothing relevant was injected. This is the thesis failure mode every agent harness shares: the agent can't ask for what it doesn't know it forgot — the memory layer has to volunteer it.
Range: 0..1, LOWER is better. 0.15 means memory stayed silent on 15% of the turns where it had the answer.
False-fire rate (BrainBench)
Key: false_fire_rate
Plain English: Of the turns where memory should have stayed SILENT, the fraction where it injected anyway. The anti-gaming companion to the know-to-ask rate — "always inject" would ace one and bomb the other. Silence beats noise.
Range: 0..1, LOWER is better.
Push precision (BrainBench)
Key: push_precision
Plain English: Of everything the memory layer volunteered into context, what fraction was actually relevant to the turn? Micro-averaged over injected pointers, so a 3-pointer turn weighs three times a 1-pointer turn — the way a token budget experiences it.
Range: 0..1, higher is better.
Push recall (BrainBench)
Key: push_recall
Plain English: Of everything that SHOULD have been volunteered (the gold pointers), what fraction actually was? Pointer budgets cap this by design: a seam that may inject only 1 fragment cannot reach full recall on a 3-entity turn — that constraint is what the per-harness rows measure.
Range: 0..1, higher is better.
Write-back fidelity (BrainBench)
Key: write_back_fidelity
Plain English: Of the facts stated in a conversation, what fraction survived the PRODUCTION conversation→memory pipeline (segmentation, insertion, dedup) and are findable afterward with the right entity attached? Measures the write path users actually run, not a test-only insert.
Range: 0..1, higher is better.
Provenance accuracy (BrainBench)
Key: provenance_accuracy
Plain English: Of the facts that survived write-back, what fraction carry correct provenance — the right source tag, session id, and origin page? A fact you can't trace is a fact you can't trust, audit, or expire.
Range: 0..1, higher is better.
Cross-session continuity rate (BrainBench)
Key: continuity_rate
Plain English: A decision is recorded in one session and persisted through the production write path; a different harness asks about it later on the same brain. What fraction of those decision probes were recalled — by pointer injection or stored-fact lookup? This is the continuity-that-survives-the-harness-hop moat, measured.
Range: 0..1, higher is better. Scored per reader harness (the v1 write path is harness-independent, disclosed in docs/eval/BRAINBENCH.md).
Source-isolation violations (BrainBench)
Key: source_isolation_violations
Plain English: Count of injected pointers that belong to a source other than the active one. Cross-source leakage is gbrain's must-never-violate invariant (a missed source filter is a data leak), so this gates at ZERO — any baseline, any run.
Range: 0..n, count. MUST be 0; any value above 0 fails the gate.
Average injected tokens per turn (BrainBench)
Key: avg_injected_tokens
Plain English: Estimated tokens of volunteered context per replayed turn (chars/4 heuristic). The intrusion-budget diagnostic: two seams with equal precision can differ 3x in how much context they spend to get it. Reported, not gated, until calibration data exists.
Range: 0..n tokens, judgment call — lower is cheaper, but starving the agent has its own cost. Non-gating.
Extraction recall (BrainBench --llm)
Key: extraction_recall
Plain English: With the real LLM extractor running (instead of the deterministic gold extractor), what fraction of the gold facts did it actually extract and persist? Only scored in --llm runs — the hermetic CI gate never calls a model.
Range: 0..1, higher is better. Absent in deterministic runs.
Extraction precision (BrainBench --llm)
Key: extraction_precision
Plain English: Of everything the real LLM extractor persisted, what fraction matches a gold fact? Low precision means the extractor invents or over-extracts — junk memory that pollutes future recall.
Range: 0..1, higher is better. Absent in deterministic runs.
Coverage
Every metric printed by any gbrain eval * or gbrain search stats command resolves through getMetricGloss() in src/core/eval/metric-glossary.ts. Adding a new metric to the glossary REQUIRES updating this doc; the CI guard catches drift.