For anyone putting AI-written material in front of a board
[ILLUSTRATIVE: SCENARIO] A founder put an AI writer in front of the board. One slide had a number nobody could trace back to a source — the model had smoothed a guess into a fact, and it read exactly like every other line on the page. Someone asked "where’s that from," and there was no good answer. The deck was pulled before it did real damage, but the trust in everything else the AI had touched that quarter went with it — the good analysis got the same suspicious look as the fabricated number, because nothing on the page told them apart.
No pricing on this page. Engagements are scoped and priced once we understand the work.
Sourced, estimated, proposed, or unproven — a fact and a guess never look the same twice. Every "the agent says it's done" gets checked by something other than the agent that did the work, before anyone calls it finished. Every commit an AI agent makes gets stopped at the door if it breaks a rule the codebase can't survive breaking. Every decision that matters gets tested against the handful of things the company has agreed can never be argued around, before it ships.
The outcome: the board sees a labeled deck instead of a plausible one. You can point at any number and say exactly where it came from — and stop pointing where there's nothing to point at. AI output stops being trusted or distrusted as a blanket; it earns trust per claim, the way any one employee's track record would.
A person, a number, a quote, an AI guess, or a recommendation gets tagged at the point of writing — never left to read as a bare, unattributed fact. Proof: this page — every claim above is either checkable against a cited source or explicitly marked illustrative.
Before an AI agent's "task complete" claim is accepted, a separate check — not the same context that wrote the code — runs syntax, lint, and test verification and returns pass or fail with a transcript. This runs in our own build today.
An AI agent can't hardcode a path or drift a naming convention into a codebase we run without the commit itself refusing to land. Three mechanisms, all running today: a drift-detection hook, a path registry, and a database registry — each replacing an ad hoc guess with one lookup table.
Before a product or architecture decision is treated as settled, it's checked against a short, named list of non-negotiable requirements — a document the org actually gates work against, not a values poster. A human approves every addition to that list; the mechanism never self-approves.
This is the mechanism itself, not a screenshot of it. Nothing you paste is sent anywhere — the labeling runs as JavaScript in your own browser tab, and the rules it applies are listed beside it, not hidden inside a black box.
Nothing labeled yet. Paste something on the left and press "Label it" — or just start typing; nothing here reaches a server either way.
A pattern firing lights up here as it fires on your text — this is the same taxonomy this page cites throughout, not a black box.
Pattern check, not proof — nothing you paste leaves your browser.
The labeling taxonomy, the verification gate's code and test suite, the drift-detection hook configuration, and the decision-gate agent definition plus a registry template — as files, to drop into your own repo and vault. Not included: us operating any of it on your material, or tuning beyond what ships generically.
We run the verification loop against your repo's actual goals and the labeling protocol against a batch of your AI-generated content, and hand back a labeled record: pass or fail per run, a marked-up version of your content, and a short report naming what needs a source it doesn't have. A scoped audit, not a standing subscription.
We wire all four components into your existing repo and workflow: writing the first verification contracts for your codebase, configuring the drift hooks against your real patterns, standing up your own decision registry, and training your team to run and extend all of it after we leave. A scoped install with a start and an end.
The first step is the same for every mode, and it is not a call. Send a company URL and one sentence about what is stuck, and what comes back is a reply from a person.
A person reads every one of these, and replies within three business days.
We store what you send here to reply to it. We do not add you to a list, because we do not have one. The full detail is on the privacy page.