LabGPT (AAA Lab)

Field	Value
Status	🟢 Active
Category	Research
Lead	Reuben Thiessen
Collaborator	Cathy Chase
URL	labgpt.gse.stanford.edu
Platform	Firebase
Started	Feb 2026
Previous iteration	FeedbackGPT (Vercel)

Summary

AI chatbot for Cathy Chase’s AAA Lab research. This is the next iteration of FeedbackGPT — rebuilt from scratch on Firebase with an updated model, custom Stanford domain, and data download capabilities. The original FeedbackGPT was hosted on Vercel; LabGPT replaces it entirely.

History

• Pre-Feb 2026: Original “FeedbackGPT” app hosted on Vercel for Cathy’s lab.
• Feb 2026: Rebuilt as a new app on Firebase. Renamed to LabGPT. Live at labgpt.web.app.
• Feb 27, 2026: Model updated. Custom domain active at labgpt.gse.stanford.edu. Cathy has tested it. Showed her the data download page.

Active Issue: Condition Bleed (Mar 1, 2026)

Cathy is pilot testing two conditions:

• AI as Critic — LabGPT gives feedback on student’s experimental design, but should NOT generate designs itself
• AI as Co-designer — LabGPT collaboratively helps build the experimental design

Problem: In Critic condition, when a student asks a vague follow-up (“how would that look?”), LabGPT responds with a full experimental design — making Critic look like Co-design. The conditions are blurring.

Cathy’s question: Can we add a system prompt constraint that persists across the whole conversation, telling LabGPT to NEVER give an experimental design in Critic mode? Will it hold across the context window?

Reuben’s initial response: We could program it to reinforce the constraint every turn of conversation.

Cathy’s detailed specs (Mar 2 Slack):

Two new versions (keep current as fallback):

Critic condition — system prompt must:

1. Never propose new experimental ideas, alternative designs, or modifications
2. Never suggest specific changes, additions, or improvements
3. No prescriptive recommendations (“you should,” “consider adding,” “it would be better to”)
4. No follow-up questions or offers of additional info at end of responses (user-driven, not AI-driven)

Co-designer condition — system prompt must:

1. Not give explicit feedback on learner’s experimental questions or designs
2. No follow-up questions or offers of additional info at end of responses (same as Critic)

Shared constraint: Both conditions must suppress trailing follow-up prompts like “if you share your goal…” or “if you tell me…” — Cathy wants conversations to be user-driven.

Known issue: Cathy tried adding guardrail language before — it worked initially but drifted after a few turns. Per-turn reinforcement may be needed.

Next steps:

• Build condition-specific system prompts with hard guardrails — Done Mar 3
• Per-turn reinforcement to prevent drift
• Test with vague follow-up scenarios (“how would that look?”) — Done Mar 3
• Deploy as separate versions (keep current as fallback) — Done Mar 3 (URL-based routing)

Updates — Mar 3, 2026

Shipped URL-based condition routing and study-ready UI hardening:

URL-based condition routing:

• /fb → auto-assigns “critic” condition (AI as feedback/critic)
• /cd → auto-assigns “codesign” condition (AI as codesigner)
• No condition picker dialog — participants go straight to chat
• Sidebar only shows chats matching the current URL’s condition

Root path blocked:

• Visiting / shows a styled “check your study link” message (route-guard.js)

UI labels neutralized:

• Condition badges show “FB” / “CD” instead of “Feedback” / “Co-Design” so participants aren’t influenced
• CSV export still uses full critic / codesign names for researchers

System prompts simplified:

• Updated DEFAULT_PROMPTS in functions-labgpt/index.js to match shorter versions from Prompt Lab
• Firestore systemPrompts collection (via Prompt Lab) takes priority; defaults serve as fallback

Other Tasks

• Rename from “FeedbackGPT” → “AAA-Lab-GPT” throughout the app