The brain page

The brain page at /brain/<id> is where you land after sending a brain to practise. It’s a state machine with three faces:

• 🎒 Practising — training is in flight
• 🎓 Ready — training finished successfully; the brain is usable
• 😅 Stumbled — training failed, was cancelled, or was skipped

The page subscribes to a server-sent-events stream so progress updates in real time without polling.

Practising state

What you see while a run is queued, running, or draft:

• Hero — ”🎒 Your brain is practising”
• Progress card:
  • Progress bar — fills as latestStep / totalSteps. Indeterminate scan animation while warming up.
  • “Step 3,200 of 10,000” counter
  • “Latest loss: 0.241 — lower is better. The brain learns by lowering this number.”
  • Live log — last 8 lines from the trainer, color-coded: green (good), red (bad), grey (note)
• ⏹ Cancel practise button — at the bottom of the progress card, subtle red

Cancelling

The cancel button is destructive — clicking it opens a confirmation modal:

Cancel practise? The brain will stop training where it is. Any wallclock time spent so far won’t be recovered, and the run will be marked cancelled. You can always start a new practise run from the same prior + brain — none of your wizard choices are lost.

[Keep practising] · [⏹ Yes, cancel]

Confirm and the run status flips to cancelled and the page transitions to the Stumbled state. If you were training on Vast.ai (Marathon mode), the instance is not terminated automatically — head to the Vast.ai dashboard and destroy it if you don’t want to keep paying.

Ready state

What you see when the run completes successfully:

• Hero — ”🎓 Your brain can <primary capability>” (e.g. “Your brain can predict a number”). If multiple capabilities, the lede mentions secondaries: “It also learned to show confidence.”
• What this brain can do card — capability chips with hover descriptions. Derived from the prior slug (see Capabilities reference for recovery details).
• How well it learnt card — one line per goal, green ✓ + the goal’s title. Below the goals, a “Training summary: final loss 0.241, practised for 5m 12s.”
• Try it on your data card — the inline predict form (see below)
• Open the advanced view footer link — drops to /projects/<id> for the full developer surface

Try it on your data

The form is shape-aware. For a regression brain it shows:

• Examples (with answers) — drop or paste a CSV, one row per example. Last column is y.
• Questions (predict for these) — drop or paste a CSV, one row per query. Same columns as examples minus the answer.

For classification: same shape, last column is the label (0 or 1).

Each field has:

• A drop zone (file picker + drag-and-drop)
• A count chip that turns amber/red if the row dimensions don’t match what the model expects
• An ”✨ or use a sample to see how it works” button that fills with synthetic data of the right shape
• An expand-to-modal icon for editing larger grids row-by-row

Click Predict and the predictions render in a table beneath:

Question	Prediction
[1.2, 3.4]	5.1
[2.1, 0.8]	4.3

For classification, the columns are x and label (with probability). For unsupported shapes (causal discovery, more than inlineMaxCols features), the inline form is replaced by a link to the advanced Try-it at /projects/:id/runs/:runId.

Stumbled state

What you see when the run is failed, skipped, or cancelled:

• Hero — ”😅 Practise hit a snag”
• Failure summary — one sentence in plain English:
  • cancelled → “Practise was cancelled before it finished.”
  • skipped → “The brain didn’t run because: <reason>. Usually a missing dependency on the runner.”
  • failed → “Something went wrong during practise. The technical details below will say what.”
• Failure detail code block — the error message or skip reason
• CTAs:
  • ↻ Try again (primary) — re-executes the same run; same prior + model + evals + hyperparams
  • See the technical details — link to the advanced run-detail page

Auto-fill recovery after navigation

The wizard state (selected capabilities) lives in-memory only — it’s empty when you reload /brain/<id> an hour later. To still show capability chips, the brain page looks up the prior’s slug in the catalog (capabilitiesForPriorSlug(slug)) and recovers the wrapper’s capabilities tags. Single-prior brains work cleanly; mixture brains (starter-mix-* slug) show no chips for now.

Routing

State	Status field on the run	Hero hue
practising	queued, running, draft	amber (#fbbf24)
ready	completed	green (#34d399)
stumbled	failed, skipped, cancelled	red-ish (#f87171)

The hue cascades into the ambient backdrop + progress bar + CTA buttons so the page’s mood matches the state.