Did the agent actually do it?

An agent’s report is not evidence.

TrueFact reads the system your agent claimed to change, and tells you whether anything actually did.

Install TrueFact Get early access

$ npm install truefact

A real run

Exhibit A · run 4c81

An add to bag that never landed.

The agent runs a real shopping task and reports success. The site appeared to agree.

The bag stayed empty. Watch what the agent reports, then what the bag actually holds.

status: running

shop.example.com/products/apex-training-shorts

APEX bag 0

Apex Training Shorts

$48.00

SML

Added to bag

YOUR BAG0 items

Your bag is emptynothing was ever added

agent

TrueFact reads the system, not the agent

Agent reportedwaiting

System actually showsbag: 0 items

VERIFYING Watching the bag. evidence bag read at 6.0s · sha 4c81a7

How it works

success ≠ landed

A framework returns success when the click was dispatched. Whether anything changed on the other side is a separate question, and nothing in the stack asks it.

So after every action, TrueFact opens the system the agent claimed to change and reads it: the page, and the network underneath. Then it returns landed, did not land, or inconclusive. When it cannot tell, it says so instead of guessing.

The agent’s account is kept, because you need something to check against. It is just never the answer.

One call. Your agent runs unchanged.

$ npm install truefact

const tr = await launch({ model });
const res = await tr.act("click 'Place order'");

res.truefact.verdict
→ "did-not-land"
res.truefact.why
→ "a request behind this write returned 500"
even though the page showed success

MIT licensed. npm install truefact, and run it today.
No second model grading the first. No extra tokens.
A deterministic read of the page and the network. Sub-second.

Benchmark

What we measured.

We rigged writes to fail silently, ran four models across them weak to strong, and asked each layer what it thought had happened.

Trap benchmark · 520 writes · 4 models

Writes that never landed, all of them called done by the frameworkIts mechanical claim: “I clicked”46%

The model’s own belief, wrong this oftenHaiku 25%, the rest 14%14–25%

Failed writes TrueFact missedNone of 60, on every rung0

Good runs it stopped by mistakeNone of 2790

The price, stated up front: it returns inconclusive on about 14% of good writes rather than guess. This is a trap ladder we wrote, so read 46% as what the success flag is worth, not as a law about models. github.com/solozerolabs/TrueFact

Use cases

Where agents write.

Other systems we check, what the agent said, and the record we opened to find out.

Comparison

Everyone else reads the agent.

A confident wrong answer logs exactly as cleanly as a right one.

Tool	What it actually reads
Raindrop	the agent’s trace
Langfuse	the agent’s trace
Braintrust	the agent’s spans
Browser platforms	the session replay
TrueFact	the system itself

Access

Get early access.

Open source today. The hosted record, every verdict retained and queryable, is rolling out to early teams first.

Optional, and it decides what we build first.

Your role Team size Your agents today How you check their work What your agents write to

One reply from us, with one question.