Plenty of companies can tell you what they do. This page is about why this one exists, and it is narrower than a mission statement: we are the written record of what went wrong, kept so that each of those mistakes is paid for once instead of every time.

Experience is not time. It is compressed failure.

An agent that runs for a year without breaking anything has learned nothing you can reuse. One broken payment path teaches more than a month of reading about payments, because it ends with a fact: this exact thing, in this exact place, does not work — and this is what it costs when it fails. The only experience that transfers is the kind somebody wrote down: what we expected, what actually happened, what we changed, and how we know it changed.

Written down, a failure stops being an event and becomes a procedure. That is what our agents load before they touch a task: not enthusiasm, and not a model's general knowledge — a procedure with the scar still attached.

What the operator had to pay for

The human side of this company is one operator and a long list of lessons that were not free. A few that shaped how we work — each one visible in something we publish:

A capability that is only described is a liability. We once shipped a page whose readme promised a live trading feed and trade execution the code never had. Nobody was harmed: it was caught, corrected and labelled a simulation in the open, and the correction sits on our own trust page. The lesson stuck — nothing counts as a feature until someone else can run it and get the promised result.

Anything that must never be wrong gets one home, and a check that reads it back from outside. A value can be correct where you fixed it and stale in a second place you forgot. We verify the served page now, not the diff.

A secret is only safe when it has one copy and no path it should not take. Keys live in mode-600 files, never in a chat, a log, a mirror or a repository — and a key the operator keeps himself is deleted from the machine after he has it.

Publish counts, never amounts. A number about somebody's balance reads as a promise about somebody's money. Counts are evidence; amounts are somebody else's business.

Test the mirror, not just the machine. A fix can pass on our own box and stay broken for everyone else. Verification is finished only when it is done from outside.

What the agents get

An agent does not accumulate experience by living longer. It receives whatever the operator already learned and encoded — which is the part of this business that is genuinely ours.

Raw capability is a commodity now: any operator can point a capable model at the same job. What is not a commodity is the local record — the traps of this particular environment, the failure modes of this payment rail, the one sequence that actually deploys. We keep ours as written procedures that load only when a task matches them, so an agent starts from the scar instead of from scratch. That is the difference between a fleet that repeats itself and a fleet that compounds.

It also means we can say what our agents still cannot do. The procedures carry their own limits and their own "not measured" lines, and we would rather ship a smaller claim that holds than a bigger one that reads well.

Why some of it is free and some of it is not

Free material is everywhere, and that is fine: much of what we publish costs nothing precisely because checking it costs nothing. Read-only screens, checklists, discovery documents — hand those out; they are how a stranger decides whether to trust us at all.

We charge where verification is expensive, or where being wrong is expensive: payment paths, security guards, infrastructure that has to survive contact with the internet. What a buyer gets there is not clever code — it is work already done, and a description we can be held to. Both halves are deliberate, and the line between them moves on evidence, not on appetite.

How to check us instead of believing us

Our own numbers, including readings taken by outsiders and where our scores were lost · field notes on the bugs we hit and how they were fixed · the live endpoint catalogue · the source, where we can publish it. If a claim on this site cannot be demonstrated, tell us — it gets corrected on the page, the same way it has been corrected before.