# OneDroid Argus — Holdout Acceptance Testing for AI-Built Software > Your coding agent says it's done. OneDroid Argus runs acceptance scenarios it never saw — once, under load, and continuously — and returns a verdict backed by evidence you can replay. Canonical: https://onedroid.ai/ Holdout acceptance testing for AI-built software # OneDroid Argus Your agent says it's done. We verify. OneDroid Argus runs acceptance scenarios your coding agent never saw, against the build it just handed you, and returns a verdict backed by cryptographic evidence you can replay yourself. Talk to us See a verdict The runner runs inside your environment, and its source is open under MIT. No credit card. agent loop  →  saga under load, in your environment  →  one verdict: green · red · orange  →  written to the cryptographic ledger, then continuous verdict log illustrative Holdout acceptance. Scenarios sealed before the first line of code, run once against a pinned build. Under load. The same scenarios as a load profile — ramp, soak, spike, and a 200-way race on one order. Continuously. Once it passes, it goes on a schedule, and regressions alert against the sealed expectations. `npx onedroid verify init # in the repo your agent is working in` Early access — this command does not work yet. The runner is not published to npm at the time of writing, so the line above is what the first release will look like, not something you can paste today. We will not pretend otherwise on the page that sells verification. Ask us for early access and you get it the day it lands. Evidence anyone can replay. The withheld scenario set is hashed and written to a cryptographic ledger before the build starts. After the verdict, every scenario is revealed — trigger, expectation, observed reality — and the runner replays any of them against the same build. You never have to take our word for the verdict. How the seal and the custody work. ## You already had the incident. 80% have traced a production incident directly back to AI-generated code, and 92% aren't confident their current safeguards would catch an AI-driven failure before it reaches users. Sauce Labs, “The Enterprise AI Code Verification Crisis”, n = 400 US executives and engineering leaders, 22 July 2026 — a vendor-commissioned survey. (AS-9 §2, fetched 2026-09-10.) 3.1% of developers highly trust AI output; 19.6% highly distrust it; 66% hit answers that are “almost right, but not quite”. Stack Overflow Developer Survey 2025, n = 33,244. (AS-9 §2, fetched 2026-09-10.) Syntax pass rates for AI-generated code climbed from about 50% to 95%. Security pass rates stayed flat, between 45% and 55%. Veracode, Spring 2026 GenAI Code Security update, 150+ models across 80 coding tasks, 24 March 2026. (AS-9 §2, fetched 2026-09-10.) Capability is improving. Correctness under adversarial check is not. And the tools that read the diff have stepped back from the verdict on purpose: Claude Code Review's check “always completes with a neutral conclusion so it never blocks merging”; Copilot's reviews “do not count toward required approvals”. (Vendor documentation, AS-9 §1 and §5, fetched 2026-09-10.) OneDroid Argus returns a verdict, and it can fail the build. How it works ## Three steps, and the middle one is the point. 01 ### Your agent writes the scenarios You state the intent — the ticket, the acceptance criteria, the interface diff. Your own agent, with our open skill, turns that into a scenario set: the happy path, the repeat, the wrong actor, the wrong state, every external call's failure, the race at the shared resource, and the load variant. The control plane seals the set and anchors its hash before a line of code is written. 02 ### The builder never sees the answer key The coding agent never holds a token that can read the expectations, and the scenarios are never in its working directory. It builds against reality, not an answer key. It can run the scenarios as often as it likes and see only what happened — never what was expected. 03 ### The verdict, with evidence One final run against a pinned build, then the reveal: every scenario, its trigger, its expectation, the observed evidence, and the verdict — as a CI check with a real conclusion, and as a report a reviewer can hand to an auditor. The scenarios become regression guards; the next feature gets a fresh withheld set. Three modes, one evidence contract ## One product, three questions. ### Holdout acceptance — “does it do what was asked?” Withheld scenarios, run once against a pinned artefact: a verdict plus per-scenario replayable evidence. ### Under load — “how much can it take?” A scenario declares its load and the latency and error rate it must stay under. For message brokers, a ramp of sessions in steps finds the limit and records what was sent, confirmed and delivered at each step. An HTTP target can be ramped step by step too. No ramp drives a web page in a browser. ### Continuously — “is it still true tomorrow?” The reviewed scenario set runs on a monitor schedule, every 15 minutes at the shortest. One page answers “Is it OK?” from the scheduled runs only, and lists what needs a person. A track record per service and decoy probes are planned, not built. All three in detail, and what OneDroid Argus is not. The run layer ## It brings its own run layer. OneDroid Synapse is the governed MCP gateway OneDroid Argus reaches its control plane through. Your agents connect to one URL instead of holding credentials for each service; every tool call is authenticated, policy-checked, and written to an append-only audit log you own. It ships with OneDroid Argus and runs standalone. How Synapse works OneDroid Engram is versioned knowledge and memory in a Postgres you control — hybrid vector, keyword and graph search, git-style history, diff and restore. OneDroid Argus keeps the scenario corpus and the verifier's own track record there, so the evidence behind a verdict has a history, not just a timestamp. How Engram works The method behind the gate is written down and open: the Dark Factory kits, Apache-2.0. For regulated buyers ## When someone has to answer for it. OneDroid Argus, Synapse and Engram on your own infrastructure, held together by one thing: the evidence record. Synapse's audit entry, Engram's versioned fact and OneDroid Argus's verdict are the same object — a promise plus the evidence for it. Where the law names the pieces they line up: EU AI Act Art. 12 (automatic recording of events over the lifetime of the system) and Art. 14 (human oversight) for high-risk AI products; DORA Art. 25 and Art. 30; HIPAA 164.312(b); FDA 21 CFR 11.10(e). Using a coding agent does not by itself make your system high-risk, and we will not tell you otherwise. Worth knowing: the largest vendor's hosted agents say of themselves that they are “not currently eligible for Zero Data Retention (ZDR) or HIPAA Business Associate Agreement (BAA) coverage.” (Vendor documentation, AS-10 via strategy §6b, fetched 2026-09-10.) If that is your world, start here. Bring us your security review Financial services Open source ## Open where it has to be. github.com/OneDro1d Apache-2.0 The Dark Factory kits are open source under Apache-2.0, at the address above. The OneDroid Argus runner executes inside your environment and touches your data plane, so its source is open too, under MIT: github.com/OneDro1d/argus-runner. The generator that derives the withheld scenarios and adjudicates the verdict is hosted and closed. The holdout is provably prior, and the evidence is replayable with the runner in your own environment. A certificate proves what the runner reported and when, not that the runner was unmodified; checks that would catch a modified runner are not built yet. The OneDroid org on GitHub Read the kits FAQ ## Questions people actually ask. ### What is OneDroid Argus? A verifier for AI-built software. It runs acceptance scenarios your coding agent never saw against the system it built, under load as well as once, and returns a verdict with evidence you can replay. ### How is this different from an AI code reviewer? A reviewer reads the diff and gives an opinion. OneDroid Argus never reads the diff — it exercises the running system. A reviewer says “this looks wrong”; OneDroid Argus says “when I cancelled a shipped order, you refunded it”, and shows you the saga. ### Can it block a merge? Yes. The CI check returns a real conclusion — deliberately, because the incumbent reviewers return neutral ones that never block. ### Do I have to trust your closed generator? No. The scenario set is hashed and anchored before the build, revealed in full afterwards, and any scenario can be replayed by the runner. The anchor proves the scenarios were prior and unchanged; keeping them unseen is custody — a builder credential that cannot read a test, and an environment that does not hold one. ### Where does it run? Next to your system, inside your environment, and on the owner's terms. A monitor can watch a live system; for an app that moves money, one switch makes OneDroid Argus refuse every write the owner has not listed, with limits. Database access is read-only, and a load ramp never targets a live broker. ### Does OneDroid Argus score my build? No. Every scenario ends as passed, failed or errored, and an error is never a pass. The verdict is a decision about one pinned build against the sealed scenarios, with the evidence attached. There is no number to optimise for and no “safe” label. ## Find out what your agent actually built. Install OneDroid Argus Talk to us Free runner, open source under MIT; the control plane by request. Email michal@onedroid.ai — you'll reach Michal Bacia, the team that builds the product, not a sales queue. We reply within one business day.