Skip to content

Latest commit

 

History

History
62 lines (49 loc) · 2.62 KB

File metadata and controls

62 lines (49 loc) · 2.62 KB

evals requirements

Context

The corpus produces and validates the root AGENT-SPEC.md. The vision is defined in vision.md.

Assumptions

  • A01 Real harnesses: Claude and Codex may require different overlays and hooks while implementing the same agent contract.
  • A02 Mixed judges: Deterministic, cross-model, and LLM judges are all valid when their grading behavior is explicit and tested.

Acceptable Tradeoffs

  • T01 Bounded coverage: Sequential, cost-bounded evidence is preferable to an unbounded all-provider run.
  • T02 Evidence before prose: AGENT-SPEC.md may lag a proposed idea until executable evidence is accepted.

Requirements

Must make every cell reproducible

  • R01 Tracked cells: Every maintained eval is an executable, inventoried cell with its declaration, fixture, task, judges, and cleanup behavior in version control.
  • R02 Explicit models: Every provider launch declares its model and effort; no eval inherits an expensive personal default.
  • R03 Real hooks: Harnessed cells use the actual Claude or Codex hook behavior they claim to evaluate. Agent hooks and personas are baked into the fixture rather than resolved from mutable external defaults.
  • R04 Valid judges: Judges discriminate the intended outcome and have mutation evidence showing that relevant bad outcomes fail.
  • R05 Hermetic fixtures: Fixtures contain every hook and persona used by their launched agents and declare every other dependency that affects the result.

Must bound execution

  • R06 Bounded provider use: Agent count, timeouts, concurrency, stop conditions, and provider usage are explicit before a paid run.
  • R07 Cleanup: Every run terminates its launched agents and detects leftover processes.
  • R08 Durable receipts: Runs record enough state to audit, resume, and skip completed work safely.
  • R09 Visible cost: Model, effort, usage, and observed cost are visible per eval and, when runs are batched, per batch.

Must derive the spec from evidence

  • R10 Executable specification: AGENT-SPEC.md is extracted from accepted executable scenarios; the corpus proves ideas before the prose declares them supported.
  • R11 Fast authoring: A new agent idea can become a validated, runnable eval by adding declarative scenario content rather than changing the runner or building one-off infrastructure.
  • R12 Universal run model: Every agent and ordinary agent run can be represented as an eval with explicit inputs, hooks, model, lifecycle, evidence, cleanup, and receipts, and executed by the current runtime.