Limits first

If your agent's record says it refunded 20 dollars, does that make it so?

No. Arcaeon checks that every row of a record hashes to the next, and whether two records of the same work agree. With a pin, it also shows the rows up to the pinned head have not changed since the pin. It cannot tell you the record was right. This page puts what it cannot prove first, because that is the part people skip.

Where this comes from

It started on a 911 dispatch floor. A call's timeline gets read long after the call, by people who were not there. On that floor, knowing what a record cannot tell you matters as much as knowing what it can.

What it cannot prove

The record and the witness

That a record was never rewritten. Anyone who can write the file can still change it. An edit to one row breaks the chain, and the check names the line. Rewrite every row from the edited one onward and the chain checks out again, unless a pin holds the old head.
That the rows are true. The chain records whatever was written, true or not. To tie a row to a fact someone can fetch again, store the fact's digest in the row.
That nothing was cut off the end. Cut the newest rows off and what is left still checks out. No chain catches that alone. A pin does, as of the last pin.
Who wrote a row. That is data in the row, not a signature. Someone who rewrites the whole log from the first row can rewrite that too.
A second, unrelated party. The hosted witness is one witness, and we operate it. A pin there is not a stranger vouching for you.
Who ran the witness at the time of a pin. Today you take our word for it. That stays true until we publish a custody record, and we have not yet.
That the daily anchor checked anything. It is a clock, not a party. It fixes when the pins existed, not whether they were right.
That you can rerun the session. That takes the model, the tool server and their state at the time. You get the request and response digests instead, so you can check given bytes are the ones that crossed.

Two records of the same work

MATCHED means two recorders agree. A colluding agent side and tool side can write two agreeing tapes of a lie.
A call that crossed neither recorded seam leaves no row on either tape.
Without a pin, both tapes cut short or rewritten in agreement still match.

The evidence pack, the second reader and the mandate gate

A pack is not a statement that any law or standard is met, and it does not show the agent behaved well. It shows what was written was not changed after pinning.
When two readers read a sentence the same way, that does not make the claim right. Two models from one family, or one vendor, agree more for that reason alone.
The mandate gate records who a mandate speaks for and does not check it: the proxy cannot see who is behind the agent. A call that goes around the proxy is not checked.
vet and badge report what their own checks found in the bytes they read. They are not a review by a person and not a safety certification.

What it can prove

Nothing was changed in the middle. An edit, a deletion or a reorder breaks every later link, and the check names the first broken line.
Nothing was cut after a pin. Once a witness holds your head, a shorter or rewritten log no longer matches it. The longest gap between pins is your real exposure.
Two tapes line up, or where they do not. Each step both tapes should hold comes back MATCHED, MISSING or ALTERED, with the fields that differ.
What could not be looked at. A missing, unreadable or empty file is COULD NOT LOOK. It is never a pass.

Watch it break

Two rows, checked, then one word changed on line 1 and checked again:

arcaeon demolocal, no network

Check the record: VERIFIED, 2 rows, every link holds.

Now someone opens the file and changes one word on line 1:
  20 dollars becomes 200 dollars. Nothing else moves.

Check the record again. The link on line 1 no longer fits:
BROKEN: line 1: chain mismatch

The check names the line. It does not say which of the two amounts was the real refund. That part is a person's call.

The full list

The package carries the long version, word for word from its own README, in docs/WHAT_IT_CAN_AND_CANNOT_PROVE.md. It comes in the next release. Each answer word is explained on the what it proves page.

Get it for your AI