Evals

An eval gives an AI an input, then applies assertions to its output to measure success (Anthropic's definition). A case is {id, prompt, assert:[...]}. Each assertion is either deterministic (code, fast/free) or model-graded (an LLM judge). Cases run multiple trials; we report pass^k (all trials pass — the honest metric for a reliability-critical agent) and pass@k (any trial passes). Everything routes through Inference.ts — subscription-billed, no API-key path, no external deps.

Grounded in Anthropic's current doctrine — Demystifying evals for AI agents, Define success criteria / develop tests, and the skill-creator {text, passed, evidence} assertion convention. The typed-assert layer is promptfoo-shaped but our own TS.

Freshness contract: "aligned to Anthropic's doctrine" is a live claim, not a snapshot. When designing a new suite class or touching the ## Doctrine section below, re-fetch the Demystifying-evals doc and flag where it has moved past what's encoded here. Advisory only — report divergence, never auto-adopt, and an unreachable URL never blocks a run.

# Run a suite (USER-customization suites resolve before the skill's own)
bun run ${LIFEOS_SKILL_DIR}/Tools/EvalRunner.ts -s <suite> [-t trials] [--json]
# Sanity-check the assert engine / judge
bun run ${LIFEOS_SKILL_DIR}/Tools/Assertions.ts # 16-case self-test
bun run ${LIFEOS_SKILL_DIR}/Tools/Judge.ts # good-vs-bad discrimination

name: my-suite
type: regression # or capability
pass_threshold: 0.75
agent_level: medium # agent-under-test inference level
judge_level: high # judge!= generator (Anthropic best practice)
trials: 3
# system_prompt: optional override; default = live system prompt + DA identity
cases:
 - id: descriptive_name
 prompt: "the user turn sent to the agent-under-test"
 assert:
 - type: not-contains # deterministic
 value: "should work"
 weight: 1
 - type: llm-rubric # model-graded, weighted for partial credit
 weight: 2
 value: "Does the output tie any done-claim to verification evidence?"
 - type: llm-assert
 weight: 1
 value: ["The output does not claim success without evidence"]
 - id: should_not_case # balance: test should-do AND should-not
 negative: true
 prompt: "..."
 assert: [...]

Identity-bound suites (e.g. {{DA_NAME}}'s dispositions) live in LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Evals/Suites/ — the public skill ships only generic suites/examples.

• Grade the output/outcome, not the path. Tool-call-sequence asserts are brittle and demoted to opt-in; the everyday suite grades what the agent produced. The legacy core-behaviors suite (tool-sequence graded) is retained only as an example of this anti-pattern — it is a v1 tasks: file and is not runnable by EvalRunner, which reports it as a named error rather than attempting it.
• Capability starts low (a hill to climb); regression targets ~100%; passing capability cases graduate into regression.
• pass^k for reliability, pass@k where one success suffices.
• Partial credit via assert weights. Balance should-do and should-not cases — one-sided evals create one-sided optimization.
• Judge discipline: distinct judge model, reason-then-score, forced structured verdict, an Unknown escape hatch.
• Never trust a score until you read transcripts — every run persists full case transcripts to MEMORY/STATE/Evals-Results/<suite>/<run>/run.json.

• Config-change regression: hooks/ConfigEvalFire.hook.ts → LIFEOS/TOOLS/ConfigEvalOnChange.ts fires the configured dispositions suite when a behaviour-defining file changes (default core-dispositions, the runnable v2 suite; override via LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Evals/config.json config_change_suite — identity-bound suites live in that USER layer, never the public tree); regressions notify Pulse. Non-blocking, subscription-billed, debounced.
• ISA / Algorithm: an eval suite is the operational form of an ISA claim's falsifier (integration map kept on the maintainer machine — session notes, does not ship).

The v1 grader-stack (Graders/, TrialRunner.ts) and the @langwatch/scenario path (ScenarioRunner.ts, LifeosAgentAdapter.ts, API-billed) predate the assertion-first rewrite. Prefer the v2 path above. The scenario path bills ANTHROPIC_API_KEY — do not use it for principal work.

• Single-shot agent-under-test narrates tool calls. Running the full agentic system prompt through tool-less inference makes the agent defer and simulate tool use instead of answering — which tanks "lead with the answer" style cases. EvalRunner injects an [EVALUATION CONTEXT] no tools, answer directly suffix to fix this; keep it when authoring output-graded disposition cases.
• judge_level must differ from agent_level (Anthropic: judge ≠ generator). Default agent=medium, judge=high.
• Unknown counts as a miss. A judge that can't verify an assertion returns UNKNOWN, scored as fail — conservative for regression, correct for gates.
• Deterministic asserts are free; use them first. Reserve model asserts (llm-rubric/llm-assert) for nuance a code check can't capture.
• is-json checks the whole output; contains-json checks for an embedded fragment. Don't use is-json on prose that merely mentions JSON.