Agent proposes, verifier gates

Why “mostly correct” fails

A model that is right 95% of the time sounds excellent until you run it a thousand times a day against a ledger. Then it is fifty wrong entries a day, arriving with the same confident tone as the correct ones, in a system where the cost of a wrong entry is not one-twentieth of the value of a right one.

Accuracy is the wrong frame for automation. The question is not how often the model is right — it is what happens on the occasions it is wrong, and whether the system notices before the consequence lands.

How the pattern works

The work is split by what each component is actually good at. Models are good at ambiguity, language, and synthesis. Deterministic code is good at being certain. The pattern uses each for what it is reliable at, and never asks the model to certify itself.

The agent proposes

It reads the ticket, drafts the entry, maps the fields, extracts the values, or writes the summary — the part that requires judgment across messy inputs.

The verifier gates

Schema validation, business rules, reconciliation against a source of truth, type and range checks, and tests. None of it involves a model. All of it can be reasoned about and unit tested.

Pass proceeds, fail holds

Passing output continues automatically. Failing output routes to human review with the reason attached, rather than being silently retried or silently shipped.

The gate is the audit trail

Because every output passes an explicit check, the record of what was checked and why it passed is a by-product of running the system, not extra compliance work bolted on afterwards.

What the verifier actually checks

The strongest verifiers are boring. They assert that a total reconciles, that an identifier exists in the system of record, that a date falls inside the period, that a required field is populated and typed correctly, that a proposed write is idempotent.

Where a deterministic check is genuinely impossible, the gate becomes a human one — but a narrow one, presented with the specific claim to confirm rather than the whole task to re-do. The aim is to spend human attention only where it is irreplaceable.

Where the pattern applies

It earns its keep anywhere the cost of a silent error exceeds the cost of a held item: finance-critical automation, data migration validation, knowledge extraction, semantic-layer query validation, report generation, operational triage, and internal tooling that writes to systems of record.

It is unnecessary where output is inherently reviewed by the person who asked for it — a drafting assistant a human reads before sending does not need a gate, because the human is one.