Research agents that have to show their work.

ARES

You give it a question. It reads papers and code, runs experiments in a sandbox, and writes up what it found. Before you see any of it, the system splits its own answer into separate statements and records which source each one rests on.

Then it checks them. A statement that fails is held back or marked unconfirmed. It is not quietly reworded.

In the recorded runs on the demo, every one of fifteen claims was declined. That is the system working, not failing.

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.

$ grep -c demichelis2025gannon paper/*.tex
0
$ grep -c parker2024gannon paper/*.tex
0

The manuscript is one I co-authored. The system was pointed at it after publication and had no part in writing it.

Three lines, in the order they have to arrive.

The agent exists and runs today. The two lines under it are what turn a good answer into one somebody can sign off on.

1. The agent. Reads, reasons, runs code, reports. Runs locally today. Hosted so a group can use it without me is the next work.
2. The verification layer. The four checks above, and the record of which claim rested on which source. Two checks run; two are written, tested, and waiting on a calibration soak before they go on by default.
3. The execution record. Every computation recorded so a stranger can repeat it: environment, network, exit code, digests of the code and its output. Recorded today, self-reported. Hardware-backed attestation is later work.

Beyond that: the same checking applied to work our own agents did not produce. We are not building that yet and will not claim a date for it.

When it runs code, it keeps the receipt.

Enough detail to be repeated by someone who does not trust the result.

It is a record, not a proof: the machine is reporting on itself. A receipt that overstated what it was would defeat the purpose.

People whose work costs them something if it is wrong.

A lab getting a paper ready, an engineering team that has to defend a number, an agency scientist whose sign-off carries their name. They are also the only users who can tell us whether a refusal was correct.

The plan is an annual licence for a group of five to fifteen researchers, priced against the salary of the person whose sign-off it defends. Not per run: charging by the run invites a comparison against a cheaper unverified answer, and that is a comparison we lose every time it is made.

No users, no revenue, and most of the checking switched off.

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.

The first thing we did with it was aim it at our own paper.

1. Ran five audit passes against a solar-wind manuscript I co-authored, already presented.
2. 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.
3. It also fabricated two citations, giving confident line numbers in a source file that does not exist.
4. The cause was mechanical: a quote was being stored as evidence without anyone checking the quote appeared in the source it was credited to.
5. We fixed that and re-ran the same five passes. Fabricated formulas went from six of seven to zero of seven.