Help and diagnostics
Settings › Help & diagnostics is where you gather what support needs to help you. Nothing on this page sends anything anywhere: every button makes a file or a clipboard copy, and you choose who gets it.
The diagnostic report
Save diagnostic report bundles the app's own logs, your system details - each graphics card by name with its size, how much memory is available now (the figure models are graded against), and what the bundled engine runs on: Metal, Vulkan, or the processor - and the operating system's crash records for the app into one readable text file. It never includes your conversations, memories, or keys - open it in any text editor and see exactly what you are sharing. Copy to clipboard gives you the same report to paste into a reply, and Show in folder finds the saved file.
If the app ever gets stuck starting, the loading screen offers to save the same report - no working app required.
The truth matrix
Run truth matrix loads every downloaded model on this computer and checks the app's claims against what really happens. It unloads your current model first, keeps the computer awake while it runs (macOS and Linux), shows its progress line by line, and writes a report file you can send back; the page says where it was saved, with Show in folder. Stop ends it early. Expect more than the "10-20 minutes" the page says on a machine with many models: the answer-quality step alone gives each model up to three minutes. The sections, in order:
| Section | What it proves |
|---|---|
| Components | The small background servers answer - the memory model returns vectors, the helper model grades a question - and the agent program is the one this app ships |
| Devices | The Vault, when this computer last backed up, how old this Vault's copies of your other devices are, the last check for their changes, and where this computer is listed for them |
| Page errors | The last ten errors the app's own screen code threw, as time, message and code location - never your words |
| Tool call | Each model graded for agent work calls one named tool when asked the way a program asks |
| Fit truth | Each model loads at the size its card promised, with its real footprint and speed beside the estimate |
| Grade drift | Grades read the same with a model loaded as with the card empty |
| Chat format | Each model answers in plain words, with no stray markers, and its thinking switch is honored, asked the way a chat turn asks |
| Sampling truth | The generation settings reach the engine: a greedy run repeats exactly, a hot one varies |
| Memory grades | How well up to four models extract and consolidate memory - the grade that decides which model does your memory work |
| Answer quality | Ten plain questions scored against a rubric, per model, three minutes each; a slower model is reported as too slow, not failed |
| Report/code mode | The classifier that tells a report or code request from ordinary chat, per model |
| Tune bench | The setups that Measure on this computer tries actually load and time |
| MLX | On Apple Silicon with the engine installed, the same chat-format check on MLX versions |
| Routing | The routing battery runs on your own models, decide-only: no model is loaded, nothing is generated, no credits are spent |
| Headroom | The grade's "runs at" context and the loader agree on one memory figure |
The fit column shows the app's own grade, so a model that runs split between the graphics card and main memory reads as such. The report says what its speed figures measure: writing at a near-empty context, and reading a short prompt where start-up cost dominates; the bench leg reads a long prompt too and labels both.
What the matrix sees, and what it does not
The matrix asks its own fixed questions - a science question, a feelings question under a companion persona, the quality rubric's ten, and the routing battery's query table. It never reads your conversations, your memory, or your AIs' knowledge. A test load starts only when there is room for it; a model graded too big is not loaded to confirm it when memory is short, and the report says so. The report names no device and no identity: your identity is described by count only, your other devices as "other device 1" and so on, a server of your own as "your own server", your home folder as ~, and anything that looks like a path or an address in a page error as <path>. It stays safe to share.
Other controls nearby
- Check for updates (Settings › Appearance) asks yourownai.net once a day for the latest version number. The request carries nothing about you - not even your current version - and nothing ever installs by itself.
- Reset to defaults removes all your AIs, conversations, and learned memory, and brings back the default AIs. Your Flowsta identity and downloaded models are kept. It cannot be undone, and it asks first.