Evidence¶
This page lists what has been tested and how. The engineering limits are in Limits.
Surfaces¶
| Surface | Kind of evidence | What was run |
|---|---|---|
| Web | Required CI and field runs | The full pipeline on every CI build, plus third-party OpenEMR runs |
| macOS desktop | Counted trials | 3 of 3 TextEdit runs with exact file contents checked; refused a case with two matching windows without changing either file |
| Windows desktop | Counted trials | 3 of 3 WinForms runs with the saved rows confirmed in SQLite; 3 of 3 refusals for stale and ambiguous targets |
| Linux desktop | Required CI and counted trials | A real GTK3 workflow in CI: three confirmed saves, three refusals for an ambiguous target, three for a stale one |
| RDP | Counted trials | 3 of 3 runs over a real network into Windows 11 through the Aardwolf client, with the files checked through guest tools |
| Citrix | Required CI and a stand-in | 3 normal and 3 drift runs on a stand-in with no web page structure. The result records ica_hdx_accepted=false: no run over a real Citrix ICA/HDX session has been counted. |
A surface that isn't listed hasn't been tested. The detail and the retained reports for each row:
Windows. The counted 20260717-candidate-56759c8-v2 WinForms run
completed 3/3 trials, an independent SQLite read confirmed 3/3 saves, and
stale-target and ambiguous-target controls each refused 3/3, with 0 silent
incorrect successes, 0 over-halts, and 0 model calls. The report also keeps
earlier diagnostic runs that were rejected; they aren't counted. It covers
the named task on a Windows 11 ARM virtual machine snapshot.
Report, first retained at Flow commit defafbae758a75c8e149d9693f2cffe1f2264b8c.
macOS. On one macOS 15.7.3 arm64 host, candidate b1b61a5 completed 3/3
TextEdit trials with the exact file bytes checked, and refused a two-window
ambiguity without changing either file. The original report still reads
status: failed, because cleanup warnings at close were counted as a batch
failure. A separate review keeps that result, checks that the test's
processes and temporary folder were gone, and accepts the save and refusal
evidence.
Review, first retained at Flow commit ca1b522cad215875f7471782283f8f8bb8e6c998.
Linux. The required CI job
linux-atspi-x11
at Flow commit 3de5fc67acf3024a621f812c5a6ed9be07fac335 (Actions run
30059807758, job 89378981573) ran a fresh GTK3
process per trial on Ubuntu 24.04 with X11: 3/3 exact file saves, 3/3
ambiguous-target refusals, and 3/3 stale-target refusals, with no model
calls. It covers that test app and CI image, and doesn't establish Wayland
or arbitrary third-party application support.
RDP. Aardwolf 0.2.14 into a Parallels Windows 11 virtual machine
completed 3/3 file saves through the Windows Run dialog, read back through
guest tools
(report).
Separately, a real FreeRDP round trip recorded, compiled, and replayed a
synthetic note task: 3/3 healthy saves and 3/3 drift halts, with no model
calls
(report, first retained at Flow commit
affedc5f1f0de533a0744deaa8e30a203c91c6b3).
That second run is a synthetic Linux app, not Aardwolf, a Windows app, or
Citrix ICA/HDX.
Citrix. The --backend citrix path binds the exact Workspace or Viewer
window and requires on-screen readiness text before a governed run. On a
stand-in with no web page structure it completed 3/3 healthy saves and 3/3
severe-drift halts with no model calls. The retained result records
code_readiness_accepted: true and ica_hdx_accepted: false: it is not a
counted real ICA/HDX batch.
Report, first retained at Flow commit
f6faac5b900b78cbda5980de0e983a9f987285ac.
Fault cases¶
Certification and qualification run six cases, three times each, against the exact bundle: one normal run and five where something breaks, such as a second copy of the save button or a missing record check. Each fault case passes only if the run stops before it can report a false success. A full qualification of the MockMed sample on 17 October 2026 passed all 18 trials in 71.2 seconds. See Certify and qualify a workflow.
Production workflow admissions¶
The public list of workflow admissions is production-workflow-admissions.json. As of 14 October 2026, every entry is a synthetic workflow for remote testing. None is a customer workflow.
The managed runner¶
The managed browser runner installs a pinned engine release, not the newest
one. As of 2026-10-17, the managed runner's runtime manifest pins Flow 1.34.0
at release commit 30fc60e55778a0e0f92b9776117cafcfe2512249, wheel SHA-256
56d32818989cb3a92830080ead39e10b08718c55e00988117f22cbfeaac98854, and sdist
SHA-256 a941a70015fe91897e655bfc9bf83e6cf8fbcf8bd4e4c7206ffb133c505203f8.
This is the configured pin in OpenAdapt Cloud's source, not a check of the
live service. Runs on your own computer use whatever version you install.
The managed runner is only for browser workflows; desktop, RDP, and Citrix
runs happen on a runner you control.
Still being proven¶
- A real Citrix ICA/HDX session.
- Recording Windows, RDP, and Citrix through the agent's Windows agent
server. This is on
mainonly and has been tested against mock servers only. - Production qualification of a customer workflow.