The prevailing method for evaluating AI agents is broken because it collapses three distinct questions into a single, unreliable vibe. A new framework from Brennan Zambo, creator of the Agent Execution Receipt (AER) standard, argues that you must separate record verification, quality scoring, and outcome analysis to get a true picture of agent performance. The core premise is simple: a hash match proves the work ran, but it says nothing about whether the work was good.

Step Zero: Define Done Before You Start

Before an agent executes a single tool, you must write the evaluation criteria. If you invent the rubric after the output arrives, it will fit the result like a glove, masking errors. The framework requires defining the artifact, necessary properties, and budget constraints upfront. This pre-written list serves as the claim list for verification and the rubric for quality scoring, ensuring that every property claimed by the agent has a corresponding step that leaves a verifiable record.

Step One: Verify the Execution with Cryptographic Proof

The first question is purely mechanical: did the work actually happen as claimed? This is answered by recompute, not opinion. Take a live example from November 22, 2026: a live_price call returned BTC at $59,710.86 USD. The record contains 669 canonical bytes and a published output_hash of sha256:3e7d6beea897b25f6d09d02a7113c473a04fb54c5749e66169120458e75584e8. By hashing the canonical bytes yourself, you can confirm the result was not altered after recording. This process, available at zambo.dev/verify without an account, closes the first question with arithmetic instead of trust.

Step Two: Score the Output Against the Rubric

Once execution is verified, you must judge quality. A hash match proves the system saved what it claimed, but it does not prove the result was correct. In a contrasting example, a doc_intel call analyzed a MAPPING.md document, minting a receipt with 3,580 canonical bytes. While the hash verified the analysis ran, the actual evaluation required a human to read the 1,360-word output and score it against the pre-defined criteria. The framework explicitly notes that "plausible but unverified" is a valid score, distinguishing rigorous evaluation from mere testimonials.

Step Three: Record the Verdict in the Ledger

Evaluations that live in chat threads disappear when disputes arise. The framework mandates attaching the verdict to the job's own record, specifying which steps verified, which scored what, and which were not checked. Acceptance is a recorded act, ensuring that the evaluation history is immutable. This prevents the common failure mode where a "pretty good" feeling dissolves into ambiguity during later audits.

Key Takeaways

  • Verification is free and requires no account; anyone can recompute hashes at zambo.dev/verify.
  • A verified receipt proves integrity, not correctness; a glitched price feed still verifies if the glitch was recorded.
  • Criteria must be written before the run to prevent post-hoc rationalization of poor outputs.
  • The Zambo endpoint runs 153 tools with 37 on the default surface, leaving a checkable record for every call.

The Bottom Line

Stop letting your agents grade their own homework with vibes. If you can't recompute the hash, you don't have a record, you have a story. And if you haven't written the rubric before the run, you're just making excuses after the fact.