FIG. I · THE RECEIPT 30 TASKS · ONE RUN
FOR PEOPLE WHO SHIP CODE THEY DIDN’T WRITE

Your AI says it works. AdaptOrch checks.

Run AI-written patches in a clean project copy. Separate code failures from runner failures. Get one readable receipt.

Start free with GitHub Book Shadow Audit PoC

Starter runs on your own machine, with your own API key. Your code goes to the AI provider you already use — and nowhere else.

Already have a coding agent? Wire it up in one line. 

claude mcp add --transport http adaptorch https://adaptorch.com/mcp

Cursor, Codex, Gemini CLI, VS Code, Windsurf →

Not checked
    Checked
      30 ledger tasks, one run. 8 caught 1 we got wrong Download the evidence JSON

      Every square is one recorded synthetic task. Green passed the recorded check; the red outline marks the gate’s own miss.

      In this 30-task synthetic ledger, verification moved 8 tasks from fail to pass and moved 1 the other way. Both arms and the evidence JSON is public.

      Works beside the tools you already use

      • Cursor
      • Claude Code by Anthropic
      • GitHub
      • GitHub Actions
      • Docker

      Compatible workflow surfaces; logos are not endorsements.

      • RUNS ONyour machine
      • YOUR KEYalways
      • WORKS WITHany AI agent
      • RAW DATApublished
      FIG. II · TWO USERS one execution record

      Different users. Same shipping question.

      Engineering teams need faster reviews. Solo builders need a clear pass, block, and next step. Both need the same execution evidence.

      FOR ENGINEERING TEAMS

      Stop reviewing runner noise

      Separate code failures from install, network, and runner failures before CI and review time disappear.

      See the execution trail →
      FOR SOLO BUILDERS

      Know whether the AI patch is ready

      Read Pass, Needs review, or Block without understanding every line. Send the receipt back to your coding agent as fix context.

      PASSRan in a clean environment BLOCKAction and fix context required
      FIG. II-B · PRODUCT PREVIEW connect → run → receipt

      From AI patch to one review queue.

      Connect a repository, run the patch in isolation, then open the generated run ID and receipt.

      1. 01Connect GitHub or your existing agent workflow.
      2. 02Run before-and-after checks in a clean copy.
      3. 03Review the verdict, changed files, and replay link.
      Run the 10-second walkthrough →
      AdaptOrch Runs dashboard showing queued, running, and completed verification runs
      Actual AdaptOrch Runs interface · product screenshot
      FIG. I-B · PRODUCT RECEIPT illustrative example · one page

      This is the whole product. One page, every time.

      Not a dashboard you have to learn. Not a score you have to trust. A slip that tells you what the AI touched, whether it ran, and what looked wrong — in the order you would ask.

      • 01Every file it touched, including the ones nobody asked for
      • 02The test result before the change and after it — that gap is the proof
      • 03A verdict in words, not a number you would have to interpret
      EXAMPLE RECEIPT · run_demo_a7f3c1SANITIZED
      CHANGED 3 files, 41 lines src/table.py+18 −4 tests/test_io.py+19 −0 config/prod.yaml+4 −0 · unasked
      RAN before the changeFAIL after the changePASS 12 tests · 0 skipped41s
      FLAGS 1 file outside the stated scope
      VERDICT Look at this one

      correctness_claim: false — it ran and this is what happened. We do not claim it is right.

      Open the technical workflow Copy · run · classify · receipt
      FIG. III · WHY CHECKING IS POSSIBLE copy · run · receipt

      Code is the only thing an AI makes that can be graded automatically.

      Ask an AI for an essay and nobody can prove it’s right. Ask an AI for code and you can just run it. It works or it doesn’t. There is a right answer and a machine can find it. AdaptOrch is built on that one fact.

      So we don’t ask your AI whether it did a good job. We run what it wrote and read the result. And we don’t run it once — we solve the same task a few different ways and compare. Where the attempts disagree is where bugs live.

      SANDBOX · run_a7f3c1ISOLATED
      SOURCE your reporead-only uv.lockpinned
      COPY /work/a7f3c1/1,284 files networkoff writes outside /workdenied
      YOUR FILESUntouched

      The box is thrown away when the run ends. Nothing it wrote survives.

      EXECUTION · run_a7f3c141s
      BEFORE test_io.py::test_roundtripFAIL 11 passed · 1 failed19s
      AFTER test_io.py::test_roundtripPASS 12 passed · 0 failed22s
      RUNNER installok environment faults0
      EVIDENCEBroken, then fixed

      Failing before and passing after is the strongest thing software can show. A green suite on its own is not.

      EXAMPLE RECEIPT · run_demo_a7f3c1SANITIZED
      CHANGED src/table.py+18 −4 tests/test_io.py+19 −0 config/prod.yaml+4 −0 · unasked
      RAN beforeFAIL afterPASS
      FLAGS 1 file outside the stated scope
      VERDICTLook at this one

      correctness_claim: false — it ran and this is what happened. We do not claim it is right.

      FIG. IV · FOUR LIES told with a straight face

      Four lies your AI tells with a straight face.

      None of these are made up. Ship AI code for a week and you will meet at least two.

      1. #01“I FIXED IT”

        It fixed nothing

        That test was already passing before the fix. We run it before and after. Broken first, working after — that is a fix. Anything else is a story.

      2. #02“ONE FILE”

        It touched six more

        Agents wander. We list every file the change actually touched and flag the ones nobody asked for.

      3. #03“ALL GREEN”

        It deleted the test

        Or wrapped it so it can never fail. A green check you got by deleting the check is the most expensive green there is.

      4. #04“YOUR BUG”

        Your internet dropped

        Most tools count a dead network as a broken change. So you hunt a bug that was never there. We tell you which one it was.

      Open measured evidence Two runs, raw data, and the result that went against us
      FIG. V · MEASURED EVIDENCE measured run + scenario estimate

      One measured run.

      Plate 00 is a measured paired run — baseline 86.7% against verified 100.0% within one session on 30 ledger tasks, +13.3pp (95% CI +3.3 to +26.7). It is a different experiment from the receipt at the top of this page: that one is verifier_gate_ledger (73.3% → 96.7%, compared across separate arms), this one is paired_confirmatory_ledger (same units inside one session). Both run 30 ledger tasks, which is why two different pairs of numbers appear — both blocks are in the evidence JSON under those names. The aggregate paired verdict did not reach significance and the family gate did not pass, so read Plate 00 as directional, not proven. Validate against your own repositories before making a deployment decision.

      Measured paired run on 30 ledger tasks: baseline 86.7 percent, verified 100.0 percent, paired delta +13.3pp, 95 percent CI +3.3 to +26.7. The aggregate paired verdict did not reach significance and the family gate did not pass, so read this as directional, not proven. Two of three families saturated by baseline. Model gemma-4-31b on Cerebras, deterministic seeds.
      PLATE 00 · MEASURED — PAIRED CONFIRMATORY · OPEN FULL SIZE ↗
      FIG. VI · THE MAJORITY TRAP 20 puzzles · three models

      You already ask three AIs. It makes the answer worse.

      When an answer matters, everyone does the same thing: ask a second model, then a third, and take whatever two of them agree on. We ran that exact procedure on twenty ARC-AGI-2 puzzles. The vote got half. The best single model got seven in ten — for less than half the money.

      • 01The vote fixed none of the best model’s mistakes
      • 02It overturned four answers that were already right
      • 03Two weaker models can outvote a stronger one

      We publish this because it points the wrong way for us. Fanning a question out to more models is what our own engine used to do, and we retracted that claim the day this run finished. On this run the vote fixed nothing and broke four answers that were already right. Agreement between the two cheaper models was rare — 4 of the 20 puzzles — and right on 3 of those 4. Every one of those counts is in the evidence JSON under arc_agi2_majority_vote.

      Download the evidence JSON and count it yourself

      ARC-AGI-2 · 20 PUZZLES2026-08-26
      BEST ONE claude-opus-4-814 / 20 spend$16.91
      THE REST qwen3.7-max9 / 19 gpt-5.6-luna6 / 16
      ALL THREE majority vote10 / 20 spend$35.81 mistakes fixed0 right answers broken4
      VERDICT Twice the cost. Worse.

      n = 20 · McNemar exact p = 0.125 — directional, not proven.

      FIG. VII · PATCH INSPECTOR interactive example · apply → classify → review

      Watch us check one AI change, step by step.

      Pick what your AI just did. We copy your project, run the change, work out whether a failure was the code or the machine, and write the receipt.

      WHAT THE AI DID Candidate patch modifies astropy/table.py to repair issue behavior.
      1. RUN ITDocker harness apply: pass
      2. LIST THE CHANGESRealized git diff generated
      3. CHECK ITFlagged: repaired, but never reproduced
      VERDICT Needs Review

      The patch applies cleanly, but we cannot call it correct. The test evidence came back inconclusive, and the shadow check flagged repair risk with no reproducible run behind it.

      FIG. VIII · PACKAGING shadow evidence → PR preflight

      Free while you try it.
      $39 when it saves you a weekend.

      Starter is free forever and runs on your own machine. Paid plans add a hosted dashboard, shared projects, and exports. Audit work for companies lives on the enterprise page.

      Bring your own key (BYOK) is required on every plan, including Starter. In plain terms: you use your own AI key, we do not resell model calls, and your AI provider bills you directly for the tokens. No key, no run — the API answers 401 and nothing executes.

      Optional ensemble fan-out calls 2–5 of your own models in one run. You are still charged one call from your plan quota, but your provider bills the tokens for every member — a 3-member run uses roughly 3× the tokens of a single-model run. Agreement and consensus figures on the receipt are measurements, not an accuracy guarantee.

      AdaptOrch runs generative AI models. Everything it hands back is AI-generated, labelled as such in the API response, and needs a human to look at it before you rely on it.

      PLAN POSITIONING MONTHLY INCLUDES ACTION
      01STARTER For local evaluation $0/month Runs on your machine, your API key Full evidence log for every run Reports only — never edits your code Run locally →
      02PROSOLO BUILDERS For solo builders and AI agent developers $39/month Hosted dashboard, no setup Export any receipt as PDF or JSON Timeline of every patch and check Starter rule templates Choose Pro →
      03TEAM For startup engineering teams $149/month Shared projects Role-scoped review views See whether the code or runner failed JSON evidence export Choose Team →
      04SHADOW AUDIT POC For B2B teams adopting AI coding agents Directcontract 2–4 week technical validation 1–3 repositories · 30–100 candidates Full evidence trail and runner health Final review presentation Book PoC →
      05ENTERPRISE PRIVATE RUNNER For regulated or private-code environments Directcontract Private runner · on-prem or VPC SSO / RBAC Custom rule sets GitHub / GitLab workflow integration Talk to us →

      † AdaptOrch reports what it found. It never edits your code and never swaps in a different answer for you. Anything that would change a result stays switched off until we have paired evidence that it actually helps.

      Estimate your review load Optional calculator · two inputs
      FIG. IX · ROI review · CI · runner waste

      One bad merge costs an evening. Two cost a weekend.

      Move the sliders to your own numbers. This is the time you spend re-reading AI code that was never going to work.

      #01 · MONTHLY REVIEW HOURS 67.5 Your inputs, with no assumed catch rate
      #02 · THE PLAN THAT FITS Pro Based on the numbers you just entered
      #03 · RECEIPTS TO REPLAY 45 One receipt per checked change
      FIG. X · WHAT WE DON’T CLAIM the honest part

      We cannot prove your code is right.

      Nobody can — not us, not your AI, not the person reviewing your pull request. Any tool that tells you otherwise is selling you something.

      What we prove is narrower and real: it ran, here is exactly what it touched, here is what the tests said before and after, and here is how to run the whole thing again yourself. That is less than magic. It is still more than you have right now.

      We also publish the checks we got wrong. On that synthetic ledger, verification improved 8 task outcomes and regressed 1. Both arms are in the file you can download.

      Read the formal claim boundary

      FIG. XI · FAQ claim boundaries first

      Questions people actually ask.

      If anything on this page promises more than we can show you, tell us. We will fix the page.

      Q.01Is this another AI that writes code for me?

      No. It checks the AI you already use. Keep Cursor, Claude Code, Codex, or whatever you have. AdaptOrch sits next to them and grades their work.

      Q.02Can you prove my code is correct?

      No, and neither can anyone else. What we prove is narrower and real: it ran, here is exactly what it touched, here is what the tests said, and here is how to run it again yourself.

      Q.03Will it change what my AI does?

      No. It watches and reports. It does not rewrite your code, pick a different answer, or merge anything on your behalf.

      Q.04How long does a check take?

      On the 30-task ledger run we publish, the checked path took a median of 60 seconds and 126 seconds at the 95th percentile (gemma-4-31b on Cerebras — a synthetic suite, not your repo). Your number is mostly your test suite: we add a sandbox run and the check on top of whatever your tests already cost. It runs while you keep working and tells you when the receipt is ready.

      Q.05What if the test fails because my machine is broken?

      We say so, in those words. A failed download is not a failed change, and we refuse to let you waste an hour finding that out yourself.

      Q.06What counts as the AI's fault?

      A crash, a broken import, a syntax error, or a test that used to pass and now does not — where the trail leads back to the change itself.

      Q.07Do you ever peek at the answer?

      Never. Official results, hidden tests, and reference solutions are banned as inputs. A checker that has seen the answer key is not checking anything.

      Q.08Does it work with Claude Code, Cursor, or Codex?

      Yes, and with anything else that produces a code change — including your own scripts. We take the change, not the tool.

      Q.09What is actually on the receipt?

      Every file the change touched, whether it ran, what the tests said before and after, anything suspicious we noticed, and a link that replays the whole check from scratch.

      Q.10Can I run it on my own machine?

      Yes. Starter is self-hosted and free. Companies that need it inside their own network can run a private version.

      Q.11Is my private code safe?

      On Starter we never see it — it runs on your machine. On paid plans, reports store fingerprints and redacted text by default, not your raw source.

      Q.12When will it block a bad merge by itself?

      Only once it can show paired evidence — broken before, working after — on your own code. Until then it reports and you decide. We would rather be useful than impressive.

      FIG. XII · BOOK DEMO audit first · merge later
      START HERE

      Stop hoping. Start checking.

      Point AdaptOrch at one change your AI made today. Read the receipt. If it tells you something you did not already know, keep it.

      • Works with any coding AI
      • One readable page per change
      • Never changes your code
      • Every check can be run again
      • Free on your own machine
      REQUEST§ 01 · SHADOW AUDIT

      Prefer email? ingeng2004@gmail.com

      LOG IN

      BOOK POC START FREE