Inputs
What the worker is given, and what it is not permitted to go looking for.
Design Pattern · September 2026
The reviewable unit of AI work is not the model and not the prompt. It is a written contract stating inputs, sources of truth, output shape, allowed actions, review gate, memory rules, and failure behaviour.
Under the pillar Zero Neo
A worker contract is short, plain, and specific enough that a person who did not write it can tell whether the worker complied.
What the worker is given, and what it is not permitted to go looking for.
Which system wins when two systems disagree, stated in priority order.
The exact artifact produced, so downstream steps can be built against it.
What it may change, and the boundary it may not cross unattended.
Who reviews, when, and what a rejection does to the work in flight.
What it retains between runs, for how long, and what it must forget.
What it does when it cannot complete: stop, escalate, or return a partial result marked as partial.
Models change, prompts drift, and vendors deprecate. The contract is the thing that can be reviewed, versioned, argued about in a meeting, and held constant while the implementation underneath it is replaced.
It is also the only durable answer to what did the AI do. A log tells you what happened. The contract tells you what was supposed to happen, which is the half of the comparison most deployments never wrote down.
Source priority, output shape, failure behaviour, memory discipline, and edit authority. Each of them looks like an implementation detail during a demonstration and becomes the entire problem in production.
Edit authority is the sharpest of them. A worker permitted to correct a record it also reports on has removed the only independent check in the loop.