Your agents obey instructions nobody has checked.

I find every sentence in your agent instructions that is no longer true, prove each one against your actual repository, and leave behind a CI gate so the lies cannot come back. You are not buying a report. You are buying a permanent check — the report is its first run.

The problem

Code that lies gets caught by a compiler. Agent context that lies gets executed. Every team that has adopted coding agents has accumulated a layer of standing instructions — CLAUDE.md, AGENTS.md, Cursor rules, memory files — that agents obey literally, at scale, every session, and that nothing on earth verifies. Each sentence was true when it was written. Drift is a property of time, not of discipline.

I know because I audited my own. The most disciplined repository I operate — a production system carrying real payments, run the way the flagship case study describes — had four live instances. Your context layer has never been checked, and mine had four.

What a finding looks like

This is the drift report format, populated with the real findings from that audit. Every claim is cited to file and line, verified against the code before it is reported, and carries a fix rather than a complaint. A suspected finding that turns out to be false is labelled as such with its reasoning — never silently dropped.

Summarised, those rows are a format. Read the excerpt itself — four findings at full depth, with the claims quoted from the file, the evidence cited to line, the assertions left behind, and the tool's real output on the repository. Including the finding no assertion can express, and the one whose check passes while the sentence is still wrong.

What you get

Seven deliverables, named, so that done is not a matter of opinion.

1. The drift report. Every claim in the context layer that is currently false, cited to file and line, severity-rated, each with a concrete fix.
2. The assertion suite. A .groundtruth.jsonc committed into your repository — your own sentences turned into executable checks. This is the deliverable that outlives the engagement.
3. The gate. groundtruth running on every pull request, failing the build when a context claim goes false. Two lines of YAML.
4. The contradiction pass. Every context file read against the others for direct contradictions, not just individually false statements. The tool cannot do this yet; a person does — which is why this is an audit and not a download.
5. The context-economics read. Which standing rules are load-bearing, which are never cited, and what to prune.
6. The read-out. One recorded hour walking the findings, ending in a sequenced remediation plan your team can execute without me.
7. The thirty-day re-check. The gate re-run a month after delivery, with a short note on what came back and why.

Scope, and fit

In scope: the agent-context layer of one repository, or up to three inside one organization — every instruction file, rules file, memory file and decision log an agent reads, plus the CI configuration that is supposed to enforce them. Out of scope: general code, security and performance review, and writing your documentation for you. Saying that plainly is what keeps a fixed price fixed.

It fits teams of roughly ten to five hundred developers where agents are genuinely in the daily loop, at least one context file is older than about three months, more than one person edits it, and the team controls its own CI. It is honestly the wrong purchase for a two-person startup with a week-old context file — there is no drift yet — or for a team without CI, which cannot receive half the deliverable.

Timeline and price

Terms: half on signature, half on delivery of the read-out. Ten working days is deliberately generous — a fixed date I always hit is worth more than a fast one I might miss. What I need from you: read access to the repository (or a working session inside your environment if access is the blocker), a named counterpart for kickoff and read-out, and permission for the two pull requests that carry the assertion suite and the gate.

Founding-client terms, through October: if the read-out does not name at least three things worth fixing, the second half is waived and you keep the assertion suite and the gate. Given what the best-kept repository in my own portfolio contained, I consider that a safe promise — but you should not have to take my word for it.

Why the evidence is public

The tool is open source and installable in a minute, the method is written up as a case study with counted figures, and the receipts come from a production system with real users and real money moving through it. The audit costs less than one day of a ten-person engineering team, and far less than a single incident caused by an agent confidently acting on an instruction that stopped being true.