OneDroid

An AI agent hands you 50,000 lines overnight. You can't read it all. So don't: hold it to what it promised. Turn each promise into a check, seal the checks before the work, and keep them from the builder. It can't game checks it never sees. Then publish the verdict with its boundary.

Shown on a decades-old Fortran library: 100 sealed checks from its user guide, 91 held, 9 failed, every check shown able to fail, and the boundary published. https://onedroid.ai/case-study-code-vs-user-guide?utm_source=bluesky&utm_medium=social&utm_campaign=CV-BS1

Testing used to be the expensive part, so it got squeezed. We measured an independent tester next to an agent that builds: same ticket, 3 runs each. Quality tied (13/13 hidden checks). Builder $1.01 vs $1.97 a run. With the tester's one-off session: $2.78. Same order of magnitude.

OneDroid Argus now has a GitHub app. When a deployment succeeds, it runs your end-to-end checks against the running system, on your own infrastructure, and puts the verdict on the deployed commit as a check. It never reads your code: checks, deployments and metadata only.

Bird, plane, or nothing? We held a drone camera's off-the-shelf detector (YOLOv8n) to plain-word requirements in a simulated scene. Near hawks: seen in 8.5% of frames vs 80% required. Red. An airliner: 95.6%, first within 366 ms, never called a bird. Green.

When a coding agent runs on a developer's machine, who keeps the record? We built one into OneDroid Synapse: prompts, tool calls and model requests, in order. Secrets shown redacted. Reading the evidence needs a reason, and that read is logged too. On our dev environment, not released yet.

Most tests check a system while everything around it works. OneDroid Argus can also make it fail on purpose while a check runs: stop a deployment, slow a connection, cut it. Only what the operator lists can fail. Every fault is undone, and a failed undo is shown. Two minutes from the tester guide.

An AI agent built a small game and wrote its own tests. All 110 passed. Then OneDroid Argus ran 30 checks from a separate tester. The builder never saw them. 28 passed. 2 found real defects. Told only which two checks failed, the builder fixed both. Next version: 30 of 30.

The OneDroid Argus runner is open source, under MIT. It runs sealed checks next to your system, and argus certificate verify checks a certificate against the chain without asking us.

We let a model play the browser game our AI agent built. Jev, a decision model from TypeSafe, answers one question per tick: left, straight or right. 20 games, 300 ticks each, the same seeds for every player.

Limits: one game, one model, 300 ticks, so a capped game says how far Jev got, not how far it would go. One of the 20 games, replayed move by move in the browser, is below.

We let a model play the browser game our AI agent built. Jev, a decision model from TypeSafe, answers one question per tick: left, straight or right. 20 games, 300 ticks each, the same seeds for every player.

An open source demo puts an AI agent's tool server in front of trading data, with a guardrail on three symbols. We ran 7 checks with OneDroid Argus, sealed before the run, 6 generated from its FINOS CALM architecture file. 5 passed, 2 failed.

What held: the declared MCP connection, and the guardrail on search (LSE:AAPL through, all three symbols refused). The set was recorded on Base Sepolia and in Bitcoin before the run, the verdict after, and the running image digests matched the declared ones.

An open source demo puts an AI agent's tool server in front of trading data, with a guardrail on three symbols. We ran 7 checks with OneDroid Argus, sealed before the run, 6 generated from its FINOS CALM architecture file. 5 passed, 2 failed.

Your team ships software built by AI. Every time, the AI says: done, all tests passed. The catch: the AI wrote those tests itself. It's marking its own homework. So who checked the work? 45 seconds, no jargon.

We ran the same 26 checks against OTRS 6.0.30 and its fork Znuny 7.3.7, three times each. A change log tells you what its authors remembered to write down. Running both tells you what changed. Six differences:

The emoji is the one to look at: the old release was the reference, and it is the one that fails. Same on both: six error paths, customer permissions, pending, close, queue move, search, sign-in response time.

We ran the same 26 checks against OTRS 6.0.30 and its fork Znuny 7.3.7, three times each. A change log tells you what its authors remembered to write down. Running both tells you what changed. Six differences:

Your AI agent wrote the code, then the unit tests, then the integration tests. All green. Who checked the work? The agent that did it. Unit vs integration vs OneDroid Argus in 50 seconds. The difference isn't what is tested. It's who wrote the test.

A unit test and an integration test come from the builder, who knows every expected answer. An Argus check comes from a tester who isn't the builder, and the builder's credential can't read it. It runs the real deployed system, nothing mocked.

An AI agent built this game in 55 minutes and it passed all 110 of its own tests. Then a tester it could not talk to ran 30 checks against it. 28 passed. The game is Dewgrub. The interesting part is the 2.

OneDroid Argus ran 30 checks from a separate tester session: 28 of 30, two real defects in the game's small service. The builder got two check names and nothing else, found both causes, fixed them tests first. Next version, same checks: 30 of 30.

An AI agent built this game in 55 minutes and it passed all 110 of its own tests. Then a tester it could not talk to ran 30 checks against it. 28 passed. The game is Dewgrub. The interesting part is the 2.

5 open-source repos for running agentic coding loops, and where each one checks the agent's work. Facts from each README, 6 Oct. 1. Ralph for Claude Code (MIT): loops until done, exits only on completion indicators AND an explicit EXIT_SIGNAL. Tied to the claude CLI for now.

5. Dark Factory (ours, Apache-2.0, 3 stars): a build accepts only raw evidence, and "a blocked or skipped check is not a pass." Stops and asks a human if a tool is missing.

5 open-source repos for running agentic coding loops, and where each one checks the agent's work. Facts from each README, 6 Oct. 1. Ralph for Claude Code (MIT): loops until done, exits only on completion indicators AND an explicit EXIT_SIGNAL. Tied to the claude CLI for now.

2. mini-swe-agent (MIT): ~100 lines of python, >74% on SWE-bench verified. Bash is its only tool. 3. OpenHands (MIT): one control center for many agents. Quick-start gives the agent your whole filesystem.

A dashboard that is green everywhere tells you nothing. OneDroid Argus has five pages, each one a question: Is it OK? What is tested? How much load can it take? Can I prove it? What happened? Here two of them answer "not measured" and "not yet". 100 seconds.