Benchmarks & Verification

Orrin separates code verification (tests, every commit) from capability benchmarks (measured on long autonomous runs) from run evidence (sealed capsules and run reports).

Code verification

• make verify is the standing gate: lint (ruff), type checks, import/layering contracts, module size limits, and the pytest suite (1,300+ tests across tests/brain, tests/goals_test, tests/memory, tests/runtime, tests/supervisor_tests, tests/observability_tests, tests/llm).
• make coverage runs the suite under coverage and gates against a recorded floor (.coverage-floor) via brain/scripts/coverage_ratchet.py — coverage can only ratchet up.

Capability benchmarks

brain/benchmarks/ is the single home for benchmark specs (what each tests + success criteria) and the harness that collects evidence and scores them. Two kinds:

• Passive — measured just by running the autonomous loop with sampling on (e.g. B1 bounded memory, B2 affect-driven switching).
• Scenario — need a seeded test goal or specific flags (e.g. B3 offline planning, B4 satiety closure, B5 self-repair), seeded with seed_scenario(...).

Run them by launching Orrin with ORRIN_BENCHMARK=1 (per-cycle sampling + auto-eval). B3 requires the LLM off to prove offline planning. Results and the claims-vs-evidence ledger live in docs/Capability, Benchmarks & Evidence/.

Run evidence

• Life capsules (brain/evidence/life_capsule.py) — each run can be sealed into a self-describing .orrinlife.zip with raw streams, cleaned tables, a queryable SQLite DB, computed metrics, and a claims ledger. See Existence and Lifecycle.
• Run reports — every staging run gets a dated folder under docs/Behavioral Evaluation & Runtime Diagnostics/demo_runs/ with findings, audits, and a pass/fail verdict against the acceptance gate in docs/NEXT_RUN_TESTS.md.
• brain/symbolic/benchmark.py scores symbolic reasoning against ground truth continuously.

Code pointers

• Makefile — verify, coverage, coverage-update
• brain/benchmarks/__init__.py — specs + harness + seed_scenario
• docs/Capability, Benchmarks & Evidence/BENCHMARKS.md — the benchmark ledger
• docs/NEXT_RUN_TESTS.md — the current run acceptance gate