elizaos--eliza/packages/benchmarks/lifeops-quality/README.md

lifeops-quality — recorded-baseline LifeOps benchmarks (#10723)

Two keyless, deterministic quality benchmarks over real LifeOps code paths, with committed baselines and ratchet-style regression gates (mirrors packages/benchmarks/recall-bench).

Lanes

Lane	Code under test	Corpus	Gate
triage	plugins/plugin-inbox/src/inbox/triage-classifier.ts — classifyMessages's prompt-build → parse → validate → fail-closed path	56 labeled inbox items (triage/corpus.ts) + a committed fixed-quality mock model (triage/fixtures.ts, 7 deliberate errors)	per-class precision/recall + accuracy + macro-F1 floors (budgets.json), plus an EXACT match against baseline.json (the pipeline is fully deterministic)
timeliness	plugins/plugin-personal-assistant/src/lifeops/scheduled-task/scheduler.ts — processDueScheduledTasks on a real PGlite runtime with an injected clock	two 4-day 2026 DST windows (spring-forward + fall-back), 5-minute cadence, cron/once/interval across NY/Berlin/Kolkata/Sydney/UTC (timeliness/corpus.ts)	missed/duplicate/early/occurrence-mismatch fire counts must be exactly 0; max/mean fire-time deviation under budgets.json ceilings

The mock model in the triage lane simulates a model of fixed quality — the benchmark measures the classifier pipeline, not production model quality. Each batch's response rotates through every parse envelope the classifier accepts (plain object, fenced legacy array, <think>-prefixed, column format, alias keys), so a regression in any accepted shape shows up as missing/failed rows.

The timeliness oracle (timeliness/oracle.ts) never calls the production cron walker: once/cron expectations are hand-authored UTC instants (cross-checked against Intl tzdata by the unit lane), and interval expectations derive from the re-anchor-on-fire contract.

Commands

bun run test              # unit lane: metrics, oracle, corpus/fixture/baseline invariants (fast, no DB)
bun run bench             # both gates (real classifier + real scheduler tick over PGlite)
bun run bench:triage      # triage gate only (~10s)
bun run bench:timeliness  # timeliness gate only (~3min, ~2,300 real ticks)

Measured runs land in results/*.json (gitignored; uploaded as a CI artifact). CI: .github/workflows/lifeops-quality-bench.yml — keyless, runs on PRs touching the corpus, the classifier, the scheduled-task spine, or core trigger scheduling, plus nightly.

Editing the corpus, fixtures, budgets, or baseline

• Corpus/fixtures changed? Re-record baseline.json in the same change: run the gates, copy the measured blocks from results/*.json. The unit lane cross-checks baseline.json against both the corpus×fixtures score and budgets.json, so a stale baseline fails fast.
• Budgets are calibrated so ONE additional triage misclassification (or any scheduler fire defect) trips the gate — triage/corpus.test.ts proves this calibration. When a run beats a floor with headroom the gate prints a ratchet prompt; tighten budgets.json rather than letting headroom rot.
• The timeliness maxDeviationMs ceiling is the tick cadence (300000ms) — do not loosen it; a larger ceiling would tolerate a real lateness bug.
• The fall-back window deliberately has no cron inside the repeated 01:00–01:59 hour: core's ambiguous-hour behavior is pinned separately and must not be baked into this gate's expectations.