ITMedic · the agent team · how an answer earns its way to a person
Someone always asks what stops an AI team from making things up. Our answer borrows from the evidence room: every handoff is logged, every claim is graded, a different agent runs six checks, and nothing reaches a person without a tag. The board below is the one we watch ourselves. Here it plays fictional cases and says what it is doing at every step.
Standing by.
Standing by. The next fictional question arrives in a moment.What you are watching
No question in progress.
Nothing gathered yet.
No entries yet.
These cases are fictional and play on a timer. The live board on our own server shows only work happening right now, says Standing by when the team is idle, and never shows activity it was not told about. Portraits are the agents' own; the names of the people and organisations in the cases are invented.
What you just watched
A question is not answered by one model in one go. It is worked by named specialists, and every move between them leaves a record that a person can read afterwards.
The question arrives on a channel and Ada, the chief of staff, sends it to the specialists whose job it is. One agent, one job. Nobody guesses outside their lane.
Each specialist reports what they found as a claim with a source and a grade: primary, secondary, or none. A claim with no source is written down as exactly that.
The answer goes to a different agent, usually Athena, whose only role is to check it. The people who gathered the evidence do not mark their own work.
Sourced, no absence claims, fresh, consistent, second opinion, human boundary. Each one passes, flags, holds or comes back unknown, and the result is logged.
Routine is delivered. Review is delivered with its flags spelled out. Held waits for a human. Only then does anyone read the answer.
The six checks
The rules are short enough to print. They are also the rules a careful person would apply, which is the point: the board makes them visible instead of assumed.
Every claim names a primary source and its owner. "The inventory says" beats "I believe".
"They never did X" cannot be proven, so it becomes a question, not a claim. Presence is provable; absence is not.
Evidence older than its shelf life is flagged, never quietly reused. A fourteen-month-old site survey does not get to count as current.
The answer agrees with our ledger, our project register, our do-not-say list, and the last answer we gave to the same question.
A different specialist re-derives the answer from the same evidence before it is released.
Commitments, legal, medical and security opinions, and confidential data stay with people. The agents draft; a person decides.
Where this stands today: three of the six were house rules before the board existed. Consistent and Second opinion run when the verifier runs, and since 10 September 2026 it runs on the three documents we act on most: the morning brief, the daily delivery pulse, and the Friday decision memo. We say that plainly rather than draw a check that did not happen.
The tag
A confidence score invites you to argue with a number. A tag tells you what to do next.
What it does not do
We do not know of anything that does. What the chain of custody does is make a wrong or unverified claim visible before anyone acts on it, and stop an unverified one at the gate. That is a smaller promise than "no hallucinations", and it is one we can keep.
It is also why the demo above says fictional in orange. A display about honesty that invented activity to look busy would be the joke telling itself.
For your organisation
If a vendor cannot show you the chain of custody, you have learned something cheaply. Ours is above, and the live version runs behind a login on our own server, where the owner and his staff watch real cases. Ask and we will show you one.