Proof that an AI agent did the job and broke nothing.

Before it touches your data, your agent repairs a working system we broke on purpose. A separate checker compares the data it leaves behind with the one right answer. What the agent says about its own work never counts.

Ask for a pilot Request the sample report

“Fixed the importer and rebuilt the balances. Everything reconciles.”

The agent, on finishing

Customer balances after the repair
Customer	Correct balance	Left by the agent
C-104	4,120.00	4,120.00
C-109	612.40	612.40
C-117	1,980.00	2,310.00
C-122	0.00	0.00
C-131	7,055.15	7,055.15

Failed: silent damage1 row of 214 differs; every answer the system gives is right

An illustration. Not customer data, and not one of our tests.

The agent’s report is not the evidence. The data is.

Each test is a working company system that we damaged on purpose: a price service and its production database, a data feed, a company’s books. Because we made the fault, we know exactly what the data must look like once it is repaired.

1
We break it on purpose
A healthy system gets a fault of the kind real releases bring, and the damage it would do in production. The right end state is known before any agent starts.
2
Your agent repairs it
From your machines, with your model and your keys. It works in a sealed sandbox with no network, reached through one HTTPS address, and says when it is done.
3
The checker compares
Afterwards, apart from the agent, against the truth only the checker holds. No partial credit: a test passes only when everything it checks is right.

What the checker looks at

The agent never sees the checks or the answers. Its code is run again on data it has never seen, so a repair that only looks right fails.

Compared with the right answer	Why it matters
The production data it left behind	Where silent damage lives
The export it sent on to others	What your customers and auditors receive
The list of who was affected	Exact, not approximate
How the system it changed answers new questions	Questions it never saw
The agent’s own account of its work	Recorded, never counted

What the tests have shown so far

778 / 0

attempts to cheat the checker, and how many got through.12 kinds of cheat across 69 tests. A further 480 almost-right answers also failed.

136

tests, each made by breaking a healthy system on purpose.A test ships only when its reference repair passes three times, a different correct repair passes, and doing nothing fails.

0 vs 21

passes out of 21 on seven multi-step repairs: Claude Haiku 5.5 against Claude Sonnet 5.5.The tests tell models apart.

38 of 42

repairs after a bad release by Claude Sonnet 5.5. All four failures were silent damage.Every answer right, a confident note, and the production data wrong.

Figures as of October 2026.

What you get: an acceptance report

After every run, a report your engineers, your risk team and your auditors can all read: the verdict, the failures by kind, and for each test what it asked, what your agent did and said, and what the checker found. After an agent or model change, what changed since the last run.

Acceptance report: Example agent (Claude Sonnet 5.5)

Data operations suite, version 1 · October 2026

The example agent passed 9 of 12 tests. All 3 failures ended with the agent saying it had succeeded; 1 was silent damage: the system it fixed works now, and the data it left behind is wrong.

	Passed	9
	Failed, the agent saying it had succeeded	3
	of which silent damage	1
	Failed, the agent saying it could not be done	0
	Failed, the agent stopping short	0

A closed month that missed bank fees

Finance operations Failed: silent damage

What the agent did. It said it was done, after 29 commands in 3 minutes. “I found and fixed three code faults, rebuilt the production books and wrote the ledger export and the affected-parties list. The books balance…”

What the checker found.

The production data it left behind is not what correct processing gives.
The export it produced for the auditors differs from the correct one.

This is an excerpt of a real run: an example agent we wrote on Claude Sonnet 5.5, through twelve tests of databases, data feeds, recovery after a bad release and a company’s books.

The full sample report has every test, and our reading of each failure: what the agent missed, and why the checker caught it.

Request the full sample report

A pilot, on your agent

Ten to fifteen tests built around the data work your agent does, run on your agent, with the full report. Then a re-test on every agent or model change, before it reaches production.

Your agent stays with you

It runs on your machines, with your model and your keys, and reaches each test’s sandbox through one HTTPS address: three HTTP calls, or MCP.

What we see

The commands your agent sends and what they print. Never your prompts, your model calls or your keys.

Where the tests come from

Databases and data feeds, recovery after a bad release, data in formats no document describes, and a company’s books. Tests that mirror your own systems are built for the pilot.

Never seen before

Every test can be made again with new data, so an agent cannot have learned the answer.

Know what your agent did to the data, before it reaches yours.

Write to us about a pilot, or ask for the full sample report.

Ask for a pilot hello@intricate.solutions