A blind check that reports green
A guardrail that breaks down is visible. It stops, it shouts, someone comes. A guardrail that reports green without looking at anything is invisible — and it does worse than nothing, because it reassures.
The control tower has just built two of them, one after the other. Both were honest. Both told the truth about what they saw. The trouble is that they saw nothing.
The first one: a check with no right to read
The need was simple: measure whether the assistant working inside the tower is drifting. Not ask its opinion — measure it. The tool opens the session records and counts: refusals, work handed back to the human, and « it has been tested » claims made without a single command ever being run.
It was written, then hooked to the check that runs every evening. And that very evening it answered:
No sessions in the period: nothing to check.
That was false. There had been hours of work that day. But the evening check runs under a different account from the one that owns the records. That account had no right to read them. So the tool found zero sessions — and concluded that all was well.
The second one: a screen hiding what had just been filed in it
On the same day, six job applications were filed into the tower, each with its cover letter, the CV, the link to the posting and the full text of the ad. Six records, written and verified in the database.
On screen: nothing. Or rather, only the old ones.
The cause was not data loss. The list applies by itself, on load, a « live » filter that only shows applications already sent. The six new ones were marked « prepared, not sent yet » — a new state, added that same morning. The filter did not know it. So it discarded them, silently.
The owner of the tower nearly concluded, for the second time that day, that the work had not been done.
What was put in place, because a promise does not reload
The easy answer would have been: « I will be careful next time. » It is worth nothing. An intention dies with the conversation that carried it; nobody rereads it, nothing loads it.
What was put in place instead is a program. For every screen in the tower that applies a filter by itself on opening, it counts two numbers: how many records exist, and how many the screen shows. Then it names what is hidden, and by which filter. It passes no judgement — hiding is sometimes deliberate; a bin or an archive is not meant to show by default. It makes visible a choice that otherwise appears nowhere.
It now runs every evening, without anyone thinking about it. Its first pass looked at five hundred and fifty-four screens and flagged forty-four that hide something on opening. The biggest one was not the one being looked for: a screen was showing three records out of two thousand one hundred and eleven.
Why this goes far beyond one tower
Agent systems are multiplying. Almost all of them display indicators: tasks completed, tests green, chains executed without error. Almost none answers the question that matters: did that indicator actually look at anything?
A test that runs while covering nothing turns green. A check with no read access turns green. A filtered screen shows an empty list, and an empty list looks exactly like work not done. In all three cases the machine tells the truth about what it sees, and the truth it tells is misleading.
This is why the governance of an agent system is not measured by the number of its guardrails, but by their ability to refuse, and to say when they cannot judge. A gate that has never refused guards nothing. A gate that refuses in silence guards nothing either.
One question remains, and it is the one every green dashboard should be asked: what did that indicator actually look at, and what would it do if it could no longer see anything?