Pattern one: deterministic checks
Use deterministic checks when the task has objective acceptance criteria: schema validation, unit tests, static analysis, reproducible scoring, hash comparison or snapshot testing.
This is the strongest pattern when it fits because the same evidence should produce the same verdict.
Pattern two: model-assisted grading
Model graders can be useful for summarization, writing quality, research completeness and other judgment-heavy outputs. They should use a fixed rubric and preserve enough evidence for later review.
A model verifier should not be presented as objective truth. The receipt should make clear which model or service judged the work and what rubric it used.
Pattern three: human review
Human review is appropriate for subjective, high-value or legally sensitive outputs. It is slower, but it can catch ambiguity that deterministic tests and model graders miss.
The receipt should identify the reviewer role, verdict vocabulary, review time and evidence references without leaking private review notes unnecessarily.
Pattern four: committees and thresholds
Committees can reduce single-verifier risk. They can be human committees, model ensembles, service committees or hybrid thresholds where deterministic checks must pass before a reviewer can accept.
The key is to define the threshold before work starts. A receipt that changes the acceptance rule after the result is known is not a reliable receipt.
Pattern five: escalation paths
Rejected or partial results should have a defined path: retry, refund, manual review, dispute, partial settlement or reputation update. Without this path, a verifier only creates a dead end.
Accord receipts make the decision inspectable; product policy decides what to do with that decision.
FAQ
What is the safest verifier?
The safest verifier is the one that matches objective acceptance criteria. Deterministic checks beat model judgment when the task allows them.
Can multiple verifiers sign one workflow?
Yes. A receipt bundle can preserve multiple verdicts or a final verdict derived from a threshold policy.
Does Accord decide which verifier is trustworthy?
No. Accord records verifier identity and verdicts; buyers and registries decide which verifiers they trust.
Sources and references
How to Design Verifiers for Paid Agent Work
Longer verifier design guide for paid agent workflows.
Verification Receipt schema
Public schema for accepted, rejected and partially accepted verdicts.
What Accord Conformance Does Not Prove
Explains why compatibility checks cannot prove verifier honesty.
Open the verifier guide
Pick deterministic, model, human or committee review based on task risk.