AI feature: automated feasibility check between designing and building an experiment

[jdpigeon]

A "computational thinking in the age of AI" prototype for custom experiments.

Flow

1. Design: Overview (question / hypothesis / methods), Conditions (names + keys), Trials (counts + order), Parameters. Each step has a Next → button; Parameters ends with Check feasibility →.
2. Feasibility Analysis: a blocking modal (no Escape, outside click, or ✕). The buttons unlock when the local model finishes. Revise my design returns to editing; Continue is the only way through, and it is always offered, so the student is never hard-blocked.
3. Build: Stimuli (image/sound folders + trial list, new step), Instructions, Preview.

Entering any build step with a design that changed since the last passed check reopens the modal. A green / yellow / red verdict chip in the design nav shows while the passed check still matches the design. It is saved with the workspace (params.feasibilityCheckedPrompt) and disappears on any edit that changes the analysis.

How it works

• Rules decide (src/renderer/utils/feasibility/rules.ts): keyword rules (most limiting wins) plus pitfalls (trials per condition, session length, movement, missing hypothesis). They are meant for advisor review.
• A small local model phrases it: Qwen3-4B-Instruct-2507 or Qwen3.5-2B (Q4_K_M GGUF, Apache-2.0), picked by RAM, with thinking off and a 200-token cap. With no model installed, the modal uses the rules' own wording.
• Process model: node-llama-cpp runs in an Electron utilityProcess (src/main/llm), forked lazily and killed on quit, behind llm:generate / llm:abort / llm:event. The worker is bundled via electron-vite ?modulePath.
• Trial counts are planned up front. balanceStimuliByCondition already samples ceil(nbTrials / conditions) per condition, so folder picks no longer overwrite counts (countPhases removed). Custom defaults are now 100 experimental / 10 practice.

Full design notes: docs/feasibility-coach.md. Background research on local models: Running a Local LLM (MiMo) Inside Brainwaves (Electron).md (MiMo was dropped as too slow for lab PCs).

Try it

Models are not bundled or downloaded by the app yet. Install one by hand:

cd ~/Library/Application\ Support/BrainWaves/models   # mkdir -p first
curl -L -C - -O https://huggingface.co/bartowski/Qwen_Qwen3-4B-Instruct-2507-GGUF/resolve/main/Qwen_Qwen3-4B-Instruct-2507-Q4_K_M.gguf

Verification

• npm test: 59 files, 317 passed. npm run typecheck is clean.
• New tests cover rule precedence and thresholds, the RAM-based model pick, the gate (disabled until the model finishes; Revise doesn't pass; an unchanged design isn't re-asked), and the chip lifecycle across reopening.
• Clicked through in the Electron window over CDP: design steps, blocking modal (~5 s to the first answer with the 4B model on an M4 Pro), Continue into Stimuli, the chip appearing, and the chip disappearing plus the gate reopening after a trial-count edit.

Not done

• In-app model download (progress / resume / checksum).
• Benchmark on a real 4-core, 8 GB lab PC; only Apple Silicon has been run.
• Windows / Linux verification of the node-llama-cpp prebuilts (they are in asarUnpack).

[jdpigeon]

Screen.Recording.2026-09-27.at.1.50.40.PM.mov

[jdpigeon]

@alexandrebarachant !