How a conclusion is reached

Instructed is not enforced

A rule you put in a prompt is a suggestion. That is not a pessimistic reading of language models, it is the threat model: these agents read attacker-controlled input by design. The project being assessed writes the website copy, the docs and the token description that the model is handed. Asking the model nicely not to be persuaded by them is not a control.

So the commitments below are checked after the model has spoken and before anything is returned or settled. A violation fails the job outright. On a paid job that refunds the buyer, which is the cheaper of the two failures available: losing a sale costs a sale, and publishing a read that breaks a stated commitment costs the reason to believe any of them.

Two kinds of gate do this work and they are not equally strong. Structural gates read the deliverable's own data — whether there was enough evidence to issue a read is answerable without interpreting a single sentence, so no wording evades it. Textual gates pattern-match the prose, and natural language has infinite paraphrase, so they are tuned to fire only on phrasing that is unambiguously the violation. They are a backstop on the prompt, not a replacement for it. Stating that distinction is part of the commitment.

The model narrates. It does not decide.

On the fast tiers, the scores, flags and gates are computed in code and are deterministic — the same inputs produce the same numbers, every time, with no model in that path. What the model does is write the read: it is handed facts that have already been established and turns them into something a person can act on.

This is why the written commentary can be unavailable on a run while the scores and stats are unaffected, and why the report says so when that happens rather than quietly returning less.

“We looked and found nothing” is not “we could not look”

These are different results and conflating them is how a due-diligence tool becomes dangerous. A clean report and a report that could not reach half its sources look identical if the second one does not say so.

Every agent here carries a coverage record. A source that could not be reached becomes a declared gap in the deliverable — named, in the report, next to the conclusion it limits. Not a footnote, and never an absence you have to infer from a section that is missing.

The same rule governs every figure on this site: a number the house cannot stand behind is hidden rather than shown as zero. A missing measurement and a measurement of zero are not the same thing, and only one of them is information.

No expiry date. Invalidation triggers instead.

The obvious way to stop a report being quoted a month later is to stamp it valid for 24 hours. It is the wrong answer twice.

• The number would be invented. There is no measurement behind 24 hours or 7 days. A fabricated constant is worse than no constant, because it looks calibrated.
• A clock is not what invalidates anything. A verdict resting on an immutable contract can hold for a year. A verdict resting on a mutable proxy can be void in the next block. No interval tells those apart.

So a report states two facts instead. What it was true of — evidence pinned to a block is exactly valid at that block, which is not a policy but what a block height means. And what would void it: concrete triggers drawn from what was actually found, each marked as voiding the verdict or merely weakening a named part of it, and each shipping with the check you can run yourself.

A reader given triggers can derive a refresh interval. A reader given an interval could never have derived the triggers. This is also why re-running the same assessment is limited rather than unlimited — the triggers are the right answer to “has this changed?”, not another full run. More on that.

What this does not cover

Each agent publishes its own charter — scope, what it refuses, who maintains it — on its page in the directory, and the refusals there are specific to the job it does. This page is the house-wide floor, not the whole of any one agent's commitments.

None of it makes an agent right. It makes an agent legible: you can see what it looked at, what it could not look at, what would change its mind, and which conclusions it is structurally incapable of stating. What you do with the read is still your call, and none of it is financial advice.