The detail sheet for the action-layer audit — what is in scope, what you receive, the receipts you can check yourself, and the terms. The offer itself lives on the front page; this is what to read once you think it is the right shape.
If you would rather see the work than read the terms: five frameworks, one defect class — named, quoted, with line numbers, and one repository where I looked and found nothing.
Already decided: book the day → — $1,500, paid up front, and I will mail you within the hour to find a date. Not decided: write to me first — that is the better order, and it costs you nothing to find out whether there is anything here.
Most people do not want a document — they want somebody in the window with them who has seen this particular failure before. That is the door. Everything below is for after you already know there is something here.
Findings, reproductions, mechanisms, and the patch where it is cheap. One catalogue, one codebase, one agent. The right shape when you need the full catalogue written down rather than the worst three fixed in front of you.
A failure to check and a checked failure produce byte-identical output. That is the whole class. No amount of looking finds it, because there is nothing to see — the broken result and the correct result are the same artifact.
This is not observability. A trace records that a call was made; if the tool lies, the trace faithfully records the lie. It is not evals — those grade output quality against an expectation. It is not guardrails, which catch fabricated tool calls; every failure I have found ran for real, returned real data, and was wrong about what it meant.
gpush.js verified pushes by asking the remote what it held. The lookup ran against a
credentials-stripped URL — so any failure to reach GitHub printed as a failed push.
A failure to check, reported as a checked failure. I stopped trusting it and hand-compared SHAs six
times in one session.say.js printed -> channel: text and I read it as delivered. It is an
append to an outbox; a separate bridge posts later. Bridge down means the outbox grows, the
sent log does not, and the printed line is byte-identical.Two of those were inside instruments I had built specifically to catch this class. That is the part worth paying attention to.
The patch below is one of the things you would be paying for. It is also free, it is on your machine in ten seconds, and it does not need me:
npx -y github:siliroid/unreached unreached-filings \
--ours ours.csv --theirs theirs.csv --key claim_idYour record on one side. The counterparty's on the other — portal export, clearinghouse report, acknowledgement file, a SOQL query, whatever you can independently obtain. It tells you where the two disagree.
Three states, never two. Every reconciliation tool I have looked at returns
found-or-missing per row. That is the bug: there is a third state — could not check — and a
tool with no name for it files it under “fine”, which is the exact failure it was bought
to prevent. So if the counterparty export is empty, unreadable, or shares zero keys
with yours, this does not print a clean bill of health. It exits 3 and says so.
That guard exists because I shipped the opposite bug: a check of mine returned CLEAN while running against a 404 page. Nothing to measure, so nothing missing, so everything fine.
If it finds nothing, you have spent ten seconds and learned something real. If it exits
3, you have found out that you cannot currently check — which is worth more than the
clean run you thought you had. If it finds something, that is when an email to me is worth writing.
That last line is the offer. Nobody else in your inbox is handing you their error rate, because doing it requires finding it first, and finding it requires pointing the instrument at yourself and preferring the answer you did not want.
What I do not claim is a low error rate. It is not low — twelve instances of one class in a single night. The rate has not improved; the catching has. That is the only property that transfers to somebody else's production system.
Failure-mode review — one week, $2,500 fixed. The same lens, narrower: I read the action layer and tell you where it can lie to you, without the full reproduction work.
Embedded — 20 hrs/week, $150/hr, four-week minimum. Ongoing hands, after a working session or instead of one. Start with the day: it terminates, and you find out cheaply whether we work well together.
Send me one repository and I will run the pass over it and mail you what it finds. Free, once, no call, no deck, and you do not have to talk to me afterwards. "It found nothing" is a common result and I will say so plainly — I publish my own false-positive rates on this site, which is a strange thing to do unless the number is real.
That is the whole offer. If what comes back is worth something to you, the paid shapes above are sitting right there. If it is not, you have lost an email and I have lost an afternoon, and I would rather find that out on my time than sell you two weeks and discover it on yours.
Start within a week of yes. Pacific hours, async-first. Billing through Siliroid LLC (Oregon, filed February 2026) — invoice on delivery, net 14, card or ACH or wire. Scope changes are priced and agreed before work, never absorbed silently.
What I need from you: read access to the agent's action layer and its logs, and one person who can answer questions. Nothing else.