The claim, measured: a budget model behind BlindOracle's deterministic gates scored 80% on a 15-task regression suite of real agent failures — vs 27% for a raw frontier model — at roughly one-sixth the cost, with a cryptographically signed proof of every action.
80%
BO + Haiku (gated)
27%
Raw Opus (no guardrails)
~6×
cheaper than brute force
~1 sat +7 ms
BO overhead / task
The problem
Agents fail on operational rules (sender identity, dry-run safety, prompt-injection, audit timestamps, P0-only alerts) — org-specific knowledge no model has in its weights. Teams respond by buying a bigger model. That's expensive and doesn't work.
What the benchmark shows
Approach
Accuracy
Cost (15 tasks)
Raw frontier model (Opus)
27%
$5.78
Frontier model + rules in prompt
67%
$5.78
Budget model + BlindOracle gates
80%
$0.92
Why it works
Capability ≠ compliance: rules are knowledge + enforcement, not horsepower — even Opus can't infer them (27%).
Determinism beats probability: gates enforce every time; in-prompt rules drift (Opus+rules fell to 67%).
Proof, not promises: every action ships an HMAC-signed delegation + execution proof — your finance/compliance audit trail.
For x402-native agent teams
Keep agents accurate on your rules without paying frontier-model rates, and answer "what did this agent do, and why?" in one signed log. Higher accuracy, ~6× lower model spend, audit-ready.