Skip to content

Validator Cost

The validator line is what makes an artifact defendable instead of just produced. We don’t judge — we are math and code. The referee applies a declared rulebook and throws flags; it never renders a quality opinion and it is never a judge model.

Grounded in the live deterministic referee

Section titled “Grounded in the live deterministic referee”

This isn’t aspiration — it’s how the live eval lane works today. The DefendableCloud executor (app/executor.py) runs the checkable rules against the agent’s JSON output deterministically:

  • JSON-schema field checks — type, value-in-set, required keys on array items, expected value.
  • MATH re-derivation — each calculation’s result is re-derived from its own inputs and formula and compared within tolerance. That’s the real math referee.
  • Structured rule DSL — declared primitives, applied verbatim.

Rules whose condition is free-text English fall through to a checklist a human applies as a binary (satisfied / raise flag). Any model-backed check is advisory only — it never produces proof and never sets the verdict. Verdict tiers come from the flags: honey (pass), jelly (risk), propolis (fail).

The cost shape: deterministic Python checks are compute-bound on owned GPU at roughly $0.0001/pass; advisory model-backed checks run around $0.0005/pass; across a ~12-check chain the blended cost is about $0.0002/pass.


🐝 Operator-grade · books and records · to the shed.