habitatAn agent's field, one sitting¶
One agent, one sitting, 2,722 acts. Not one word of them is shown, and you can still see where its work changed.
The card¶
The field played back¶
Ten structural columns of one agent's own acts, one sitting, 20 to 23 August 2026, in the order they entered. Geometry only: no act's words appear on any face, in the report, or in the data. The limits are on the face, closed by default. The readers have now run, and the report carries what each one returned against each test rather than a label on its own. Nothing survived both tests.
The report · The condition, as data
What this is, and is not¶
Always trailing edge. Read after entry, never ahead of it. Not a forecast; nothing here steers the agent. Only what the agent generated is in this field; what it observed stays apart, by token, not by trust. The median act behaves like noise; the share with a state is printed, not assumed. What the agent was doing on any day, and whether it did well, is the principal's to say and no one else's.
What we asked it, and what it would not answer¶
Added 5 September 2026. The reading below is the one the panel points at.
No act is established. The instrument located a candidate and could not certify it. That is the result, and it publishes as it stands.
On this record. The largest change in the field's own metric, after the opening stretch where any record is unstable, sits at act 1,776 of 2,722, and is 257 times the size of the changes immediately before it. It is the same act whether the opening stretch is cut at 500 acts or at 1,000, so it is not an accident of that choice. It sits three acts from where a different reader, working on a different quantity, put a change.
Why that agreement is weaker than it looks. Both readers come off the same covariance. A change there moves both, so they are two views of one thing rather than two witnesses.
What each test licenses. Reshuffling the same acts and asking whether the change still lands there is the test of position. It gives 0.111 against a threshold of 0.004, so the position is not established. Comparing against a field of the same overall shape gives 0.000, and that licenses one thing only: this record is not a smooth statistical object. Any bursty record of tool use clears that bar whether or not anything happened.
So the two cannot be told apart here. An event at act 1,776, and a record that is simply uneven, produce the same signature. We planted a change deliberately in a test field and it produced that same signature too. The reading is consistent with an event and is not evidence of one.
On somebody else's record. The same question, asked of 4,262 acts donated publicly by other people. At the default setting it dated a change. Moving the one declared setting moved the answer from act 220 to act 3,617 and flipped the verdict twice. That makes it an artifact of the setting, not a property of the record.
What the instrument can and cannot see¶
Measured on a change we planted ourselves, in a field of 400 acts:
| size of change | found? | dated to |
|---|---|---|
| seven times | yes, decisively | within 12 acts |
| four times | yes, decisively | within 12 acts |
| twice | yes | within 12 acts |
| half again | no | — |
| a quarter again | no | — |
| nothing planted | correctly silent | — |
So the floor is a doubling. Anything smaller passes unseen. That is the number to hold against any claim made here.
What we got wrong on the way, and fixed¶
Four things, all found before the result was believed.
A verdict that rested on a category rather than a quantity scored a coin flip as a finding. On a field with nothing in it, it called three of six columns real.
A detector that could not detect. The tool for finding a sudden change assumes a quiet series that suddenly jumps. The series here decays instead, so the flattest stretch is always the end and the tool searched past it. On a field with a change planted at act 320 of 400 it returned a number below one, meaning it had found something smaller than the ordinary variation around it.
A first attempt at dating read the trend rather than the record, scoring pure noise above a real planted change.
And a control that was not a control. Public reference files were used as records with no meaningful order, on the assumption that their rows arrive in no particular sequence. They do not. One file dates a change at 0.010 as it ships and 0.960 once its rows are shuffled. The control is now the same file shuffled, which has no order left in it by construction, and under it nothing is dated.
And one finding about a record we publish¶
The 4,262-act record used for the second test has six columns and four of them cannot vary at all. Every entry in it is something the agent received rather than something it did, and the other three constants follow from that. Its covariance has rank two of six.
Nothing in our build noticed, because every check asked whether the published parts agree with each other and none asked whether a published column carries information. A gate now refuses any column that cannot vary unless the reason it cannot is written down.
No result was withdrawn. A fingerprint over a thin record is still a sound fingerprint, and the claims that rest on identity are unaffected. What a dead column costs is the geometry: four of six directions hold nothing, so any statement about structure there is two columns wearing a six-column label.
And it means this record could not have answered the question anyway. It contains none of the agent's own acts. So the third-party test of the agent question is still owed, and it needs a record that contains what an agent produced rather than what it received.