Output contract

Output contract

Four sections, fixed. The shape is not decoration: several perspectives collate only if they produce the same structure, and a review that reshapes it cannot be compared with the others.

## Moments

**1. <what happened>.** `observed`
Names a specific artefact and a specific failure or friction.

**2. ...** `inferred`
**3. ...** `constructed`

## What it would need

Per moment: what would have to be true, and whether the artefact provides it.
Cite the section, or state plainly that it does not.

## Anti-scenario

REQUIRED. One case where this is useless, or actively harmful, to me.
Not a caveat — a real case.

## Gaps

Each traced to a moment above, each stating its cost
(tokens / complexity / privacy / adoption / maintenance).

Moments

A moment needs three things:

1. A named artefact — a file, a ticket, a command, a conversation, an incident.
2. A specific failure or friction — what went wrong, or what was hard.
3. A grounding label.

Label	Means	Test
observed	Grounded in the briefing or the artefact	Could you point at the line?
inferred	A reasonable extension of something in them	Can you name what you extended?
constructed	You made it up	Say so plainly

Rejected as moments: "improving team velocity", "better developer experience", "reducing friction". These name no artefact and no failure, so nothing about them can be checked or acted on.

Accepted: "the Tuesday migration rollback, where nobody could say which config had been reviewed" — an event, an artefact, a specific gap.

constructed is legitimate. Labelling a constructed moment as observed is not, and it is the failure that makes an entire review uncheckable: once one label is decorative, a reader cannot trust any of them.

What it would need

The point is to make the finding checkable. Two forms:

• Provided — cite the section. §5.2 covers this via the role field.
• Not provided — say so plainly. Nothing in the spec addresses this.

A third form is acceptable and useful: cannot verify from here — a requirement that lives in code the review did not see, or spans documents. Say what a reader should check rather than broadening the search.

Vagueness here is where a review stops being useful. "It should handle this better" gives nobody anything to do.

Anti-scenario

The single highest-yield element, and the most likely to be skipped or faked.

A caveat is not an anti-scenario. "It might be slow at scale" hedges a positive review. An anti-scenario names a situation where you would not use the thing, or would regret having used it.

Good ones tend to have one of these shapes:

• The artefact works and that is the problem. "The record is exactly what it promises, and under a formal review it is a machine-readable admission we shipped on priors."
• It is right and useless. "By the time I could consult it I have already answered the customer."
• It creates an obligation. "Adopting this means claiming a control I cannot evidence."
• It is worse than nothing. "A well-formed wrong answer costs more to dislodge than no answer."

The confabulation floor. Because it is mandatory, an agent with nothing genuine to say will write a fluent, specific, unfalsifiable one — and nothing in the contract can distinguish that from a real one. This is the strongest argument for grounding: a briefing gives the anti-scenario something to be about.

Gaps

Each gap:

• Traces to a moment above. Untraceable gaps are wishlist items — the review has drifted into designing rather than reviewing.
• States a cost. Tokens, complexity, privacy, adoption, maintenance. A gap with no cost is a preference in a finding's clothes, and it is the reviewer asking someone else to pay for their taste.

Register

Fixed regardless of how adversarial the perspective is:

• No enthusiasm vocabulary — no "game-changer", "seamless", "powerful".
• No invented statistics or percentages.
• No fictional company names. Say "my team", "the client".
• No melodrama. State the objection plainly.

The contract is what keeps a sharply-dialled perspective from becoming a caricature. A maximally adversarial reader still owes three grounded moments with cited sections, and that requirement holds the voice in place.

Length

Around 600 words. The cap is load-bearing: it forces selection, and selection is where judgement shows. An unbounded review lists everything and prioritises nothing.