Changelog

0.3.0#

1 October 2026. Teach a model your own decisions, and twelve fixes, one of them for privacy.

New: teach a model your own decisions#

• The Train page fine-tunes any of the eleven models on a spreadsheet of your past decisions, on this computer's GPU. It works out the questions from the file, recommends a model with a time estimate, trains it, and shows the result on examples it never saw, with one it used to get wrong. Use it now opens the new model in the Playground; Compare on Evaluate opens Evaluate with the held-out examples and both models ready. Train.
• Nothing worse is ever kept. Each training replays general decisions so the model keeps what it knew, measures how far its general answers move, and is kept only if it is better on your held-out examples, not worse on general questions it never trained on, not answering everything the same way, and identical once saved and reloaded. Fine-tuning and its safeguards, with results on every model.
• Trained models appear with the others (Playground, Models, Evaluate and your code), cost a small file of changes on disk and no extra memory, and move between computers with Export and Import a trained model.
• Training is safe to leave running. One training at a time, in its own process, within a memory limit set for its model, at low priority; it waits when other programs need the memory, can be paused and continued, and restarts itself if the GPU stops responding.
• The training API under /api/training and /api/finetunes. Training and trained models.
• CLM 8B answers the same request the same way every time. It embedded several texts together, so an answer could change by up to 0.06 depending on what was cached; it now embeds each text on its own.
• Intern-Decision 4B and Lev load straight onto the GPU, without a temporary second copy of the weights in memory (about 9 GB on a GB10).

Fixes#

• A sensitive file is never kept. An image, audio or video variable marked sensitive was stored with the decision: its contents, its file name and its content hash. The model now reads it and History keeps only a keyed hash, as for every other sensitive value. A queued background decision no longer holds inline file data either, and temporary copies are removed even when a decision fails or is cancelled.
• Background decisions no longer report a false save failure. Each one came back with a history_write_failed warning although it was saved, and failed outright when the studio was set to refuse unsaved decisions.
• A template's default model is no longer shown as unable to run it. A variable that may hold more text than a model reads is now a note on that model, not a problem that greys it out.
• A version that narrows a variable is marked breaking. A new list of allowed values, a lower maximum, a higher minimum or a new pattern can refuse values callers already send; such versions were classed as wording changes.
• A variable with a default is never required. Adding a default with a patch left required: true behind.
• Cancelling a background decision stops the model load it started, unless another request is waiting for that model or you loaded it yourself.
• Evaluate recommends an act threshold only within 50% to 99%. It could suggest a lower one that the Playground's slider cannot show. A model that never reaches 50% certainty gets its own message.
• Save as template in a fresh window names and saves the model the Playground uses, instead of saving no default model.
• The template id and alias fields are checked in the browser again. Their pattern was invalid in current browsers, so the check was skipped; the server always validated.
• Leaving the Playground straight after opening it no longer throws an error.
• A second studio started with BASAL_NO_DOWNLOADS=1 answers download requests with the reason instead of an error 500.
• The environment check reports a GPU that is out of memory as that, not as "PyTorch failed to import".

Release on GitHub

0.2.1#

30 September 2026. The decisions endpoint is easy to find, and so is what gets saved.

• The API page opens on the studio API. POST /v1/studio/decisions comes first, and its quick start runs a starter template. Before, it sat behind a fourth tab and looked missing.
• Every decision endpoint says "Saved to History", in every format (the studio API, TypeSafe, OpenRouter and Vercel), and the page says how to keep nothing: send "store": false. Each saved call's id comes back in the x-basal-decision-id response header.
• The interactive reference at /docs documents the studio API's request bodies, with examples you can run from the page, the main History filters and each endpoint's path parameters. Try it out works; before, it offered no body to fill in.

Release on GitHub

0.2.0#

30 September 2026. Templates and history: decisions you make once can be reused, and every decision is kept, so you can see how they went and improve them.

• History replaces Activity. Every decision is kept on this computer, from the Playground or from your code through any API format: the situation, the questions, the answers, the model and the timing. Filter by template, model, answer or time; open the decisions that asked a person; label the right answers; rerun one on another model; pin the ones to keep. Decisions are kept for 30 days by default, and the History page sets what is kept and for how long.
• Templates are decisions you reuse on any model: the questions, variables that fill in the situation (a message, a plan tier, a screenshot), and a default model and settings. Every save is a new version. Code can pin support-triage@3, follow an alias such as support-triage@production, or take the latest. Callers can add questions, add options or skip questions where the template allows it.
• Compare versions on real traffic: how the answers are spread, how often each version acts on its own, its accuracy on the decisions you labelled, and how the answer changed for the same input.
• Test examples: inputs with their right answers, added from History in one click.
• In the Playground, Templates opens a template to fill in its variables and run it, Save as template keeps a decision you built, and Edit questions with Save version improves a template. Keep in history turns saving off for your own experiments.
• 30 starter templates, one for each Playground example.
• The studio API (/v1/studio) beside the TypeSafe-compatible /v1/systemone: templates, decisions with history, feedback, test examples and settings.
• /v1/systemone is unchanged. Responses are byte for byte what they were, and the official TypeSafe SDKs pass every conformance test. Calls are now saved to History too; send "store": false or X-Basal-Store: 0 to keep nothing. Variables marked sensitive are used for a decision and never written to disk.
• Safer by default. Web pages from other sites open in your browser can no longer send decisions to the studio.
• Your Activity log is imported into History the first time 0.2.0 starts.

Release on GitHub

0.1.5#

30 September 2026. Updates always show.

• No stale interface after an update. The studio tells the app window, and any browser, to check for newer scripts and styles every time; unchanged files cost only a quick check. Before, a window could keep running the previous version's files after an update.
• The desktop app keeps the same address. When port 8420 was taken, the app picked a random port on every launch, and the window lost what it saves per address: your draft, the theme and choices already made. It now reuses the port it used last time (8420 when free) and remembers it.

Release on GitHub

0.1.4#

30 September 2026. Every model comes with its own examples, each run on its model and checked to give the right answers.

• Choosing a model shows what it is made for. CLM 8B opens on a web agent choosing its next action, Lev on a question with 77 banking intents, Kev 4B on a refund policy with an exception, Laya Multilingual on a message in Japanese, Jev-Omni on a receipt photo.
• In the Playground, choosing a model while an example is showing swaps in that model's own example; Undo brings the previous one back. Anything you have written yourself is never replaced.
• The Examples menu opens with "Made for" the selected model, and the placeholders follow the model. A blank decision shows what the model is made for and its examples.
• The Models page shows "Made for" and Try it links for each model.

Release on GitHub

0.1.3#

30 September 2026. Start a new decision from scratch.

• New, next to Examples (and Blank decision at the top of the Examples menu), clears the situation, questions, attachments and answers, and keeps your model and settings. Undo brings everything back.
• A blank Playground shows three steps (describe the situation, add a question, press Decide), ticked off as you go.
• Pressing Decide before adding a question takes you to Create new Decision and highlights it.

Release on GitHub

0.1.2#

30 September 2026. Fixes the "Choose models to download" sheet on macOS.

• The model list is visible in Safari's engine. On 0.1.1 it collapsed to a thin strip showing half of one row, so you could not tick models. The list and the other scrolling areas now start at their content's height in every browser engine.
• Tested in real Safari on macOS 14 and 15 on every interface change, through Apple's safaridriver.
• The Learn page no longer assumes a GPU.

Release on GitHub

0.1.1#

30 September 2026. Fixes first-run setup on macOS.

• Setup finishes on macOS. On 0.1.0 it stopped at "Installing the model libraries" with error: File not found: /Users/<you>/Library/Application, because uv split a file path at the space in "Application Support". Setup now names the PyTorch build directly, with no file path involved. The same could have affected Windows user names with a space.
• Setup reads PyTorch's version and the final device check without being confused by warnings PyTorch prints.
• The Details log quotes commands correctly, and Copy details no longer repeats the error.
• A real first-run install into a folder named "Application Support" now runs on macOS, Windows and Linux on every installer change.

If setup failed on 0.1.0, install 0.1.1 over it and run setup again; it continues where it stopped.

Release on GitHub

0.1.0#

30 September 2026. The first release: run open decision models on your own computer.

• Eleven models: Julia 1, Laya, Laya Multilingual, Laya Typed-Decisions, Kev 0.5B, Kev 4B, Lev, GLiNER2.5 Decide, Intern-Decision 4B, CLM 8B and Jev-Omni.
• Six kinds of question: pick one, rate on a scale, yes or no, pick any, put in order, estimate a number.
• Pages: Playground, Models, Evaluate, Activity, API, Learn and System.
• The API: TypeSafe's Jev API at http://127.0.0.1:8420/v1/systemone, plus OpenRouter's and Vercel AI Gateway's formats. The official TypeSafe SDKs work unchanged.
• Installers for macOS (Apple Silicon), Windows, and Linux x64 and ARM64. The first launch checks the hardware, asks where models should run, and installs the matching engine: PyTorch for CUDA, Apple Metal, Intel XPU, ROCm or the processor.
• Tested: 14 of 14 API conformance checks, and 21 of 21 end-to-end checks with every model answering all six question types on an NVIDIA GB10.

Known limits: the macOS and Windows builds are not code-signed, Intel Macs are not supported, and AMD GPUs (ROCm) are experimental.

Release on GitHub