It reads papers and code, runs experiments, and writes up what it found. Before you
see any of it, every statement is separated out and checked against the evidence it
claims to rest on. Statements that fail are held back, not quietly reworded.
A finding, and the check
Two references in a published paper that the paper never cites.
Asked which entries in a manuscript's bibliography appear nowhere in its text, it
returned two out of sixty-seven. You do not have to take its word for it, which is
the entire point.
The manuscript is one I co-authored. The system was pointed at it after
publication and had no part in writing it.
How it was tested
The first thing I did with it was aim it at my own paper.
Ran five audit passes against a solar-wind manuscript I co-authored, already presented.
It found three real defects: bibliography entries cited nowhere, a framing that
over-claimed, and a headline statistic no reviewer could have reproduced from what
we published.
It also fabricated two citations, giving confident line numbers in a source file
that does not exist.
The cause was mechanical: a quote was being stored as evidence without anyone
checking the quote appeared in the source it was credited to.
I fixed that and re-ran the same five passes. Fabricated formulas went from six
of seven to zero of seven.
What a run records
When it runs code, it keeps the receipt.
Every computation is recorded with enough detail to be repeated by someone who does
not trust the result.
Field
Recorded
Environment
docker · python:3.11-slim
Network
disabled
Exit code
0
Code digest
sha256:134a87c07d230296773c7d9117c7bb33…
Output digest
sha256:cd1f7fa2be52425b90bfd563bfe5df87…
Attestation
self_reported
Source
pinned to one commit for the whole run
The last row is deliberate. It is a record, not a proof: the machine is
reporting on itself, and hardware-backed attestation is later work. A receipt that
overstated what it was would defeat the purpose.
What it does not do
The honest state, in the same size type as everything else.
No hosted deployment. It runs locally and I run it for the first users myself.
No users outside me, and no revenue.
Twenty-six of twenty-eight feature flags are off by default, including the pass
that generates counter-arguments. A claim has never been marked contested against
real conflicting evidence.
One of four verification contracts is a working acceptance gate. One cannot pass
by construction until its evaluator is wired in, and one is a declared stub pending
a proof checker.
Citations on my own agents' path are still judged by shape rather than by
fetching the source, so a well-formed but invented DOI can still earn credit.
Closing that is the next two weeks.
3,498 tests pass on a clean checkout. That number is here because the
rest of this was written fast, by one person, in five months.