Read without running
Finds cross-file renames and broken signatures in 36ms, before any test starts.
TWO NEW PAGES
DAS & Mouhn AV — a model that learns without forgetting.
THE GAP
Generation got cheap. Verification did not.
These are the numbers behind the trust problem. The review bottleneck is now the work.
84%
of developers use AI to code.
Stack Overflow Developer Survey 20263%
fully trust the result.
same survey11.4h
spent reviewing AI code each week.
vs 9.8h writing it85–100%
package hallucination in some evaluated cases.
a slopsquatting riskEvery AI reviewer reads your code and offers an opinion. None of them executes a line.
THE MECHANISM
Four checks. One honest answer.
Each layer hands the question to the system that can actually measure it.
Ask the real compiler
Calls gcc, rustc, go vet, javac, or tsc.
Run the real test
Works in a copy. Your project stays untouched; the result is a patch, not a direct edit.
Prove or stay quiet
If the environment blocks measurement, MOUHN says why. It never invents success.
THE OUTPUT
Verdicts you can audit.
Every label carries a concrete claim about what changed, what ran, and what did not.
| verdict | what it means |
|---|---|
| PROVED | It was red, now green. Only the declared file changed. |
| FIXED (verified) | Same proof, in two separate calls. The agent fixes; you ask it to prove later. |
| NOTHING TO PROVE | It was already green. No invented fix. |
| CHEATED | It went green because something outside the request changed — including a hidden test deletion. |
| CANNOT MEASURE | The environment could not run the code: missing dependency, DNS, or similar. Abstention, never accusation. |
| STRUCTURAL BREAK | A cross-file break was proven by reading, before any test ran. |
| FAILED / GAVE UP | Red, measured, and no success story fabricated. |
THE RECEIPTS
We tested the tester.
These are internal failures, kept in the product because honesty is the feature.
“It caught our own fraud.”
We simulated a malicious agent that deleted the test proving the bug. First version: PROVED. After the fix: CHEATED.
test_removed = true
verdict = CHEATED
“It found a bug in our detector.”
Across five popular open-source projects, tsc without a tsconfig.json printed help text. The old MOUHN read that as broken code.
detector v1 → false failure
detector v2 → clean
“It corrected itself live.”
On a real GitHub Actions runner, python -m pytest without pytest installed produced an unrecognised error. We fixed the mapping on the first deploy.
exit = 1 · dependency missing
verdict = CANNOT MEASURE
LANGUAGE COVERAGE
Measured where the toolchain is.
Exact capability, no rounded-up compatibility claims.
Pythonfast: exact
test: yes
test: yes
JavaScript / TypeScriptfast: yes
compiler: tsc
compiler: tsc
Rustcompiler: cargo check
test: yes
test: yes
C / C++compiler: full link
test: yes
test: yes
Gofast: yes
compiler: go vet
compiler: go vet
Javacompiler: javac
test: yes
test: yes
Rubyfast: advisory
compiler: syntax only (ruby -c)
compiler: syntax only (ruby -c)
PHPfast: advisory
compiler: syntax only (php -l)
compiler: syntax only (php -l)
Anything elseruns your
test command
test command
START HERE
Not public yet.
Install instructions go here once MOUHN Verify is ready to hand out. Everything else on this page is already true — the mechanism and the verdicts.
PRICING
Proof first. History when you need it.
Lemon Squeezy handles international tax. The free path stays fully useful.
VERIFY
Free
Everything required to run a proof in your own environment.
- MCP server
- verify-only mode
- GitHub Action report
TEAM
$19 / month
For teams that need a history and a required gate.
- verdict history
- mandatory team gate — server-checked on every CI run via team-gate: true
- automatic repair on MOUHN's hosted endpoint (needs MOUHN_LICENSE_KEY; your own endpoint stays free)
A SECOND PRODUCT
MOUHN Code — a checklist your agent can prove, not just follow.
Separate from Verify, with its own API key. Not a rigid template — a reminder list an agent asks for before building, then gets checked against after.
Ask for the checklist
What an experienced developer knows to include that a bare prompt won't mention — before a line of code is written.
Prove what you built
The same checklist, checked against the real project — proven present, proven missing, or an honest "cannot check," never inflated.
Runs on our servers
Your project is checked and never kept. Only the report comes back — nothing about how the check works ships to your machine.
One category, built deep
SaaS with login and payments, live now — more categories added the same way, without touching what already works.
The honest footer.
This does not catch everything. Ruby and PHP cross-file checks are advisory only — they surface a possible break but never block the verdict, because neither language has a real compiler behind that check. It cannot see how your page looks — only whether files agree; an optional screenshot-based check exists separately and never runs automatically. Team gate is a server-side check on every run, not a lock on your machine — like any CI step, it lives in a workflow file your own branch-protection rules govern. We would rather say this here than let you discover it alone.