docs/pitch_outline.md

Three-Minute Pitch + Two-Minute Q&A Outline (DOC 10)

Pitch Structure (3 minutes)

1. The Problem (30 seconds)

"Over 70% of landmark studies fail to replicate. The gap isn't bad science -- it's that real-world constraints force compromises that nobody planned for. Budgets shrink, equipment breaks, timelines slip. The protocol that worked in Theory A fails under Constraint B."

2. Our Solution (30 seconds)

"ReplicaLab is an OpenEnv environment where an AI Scientist learns to negotiate realistic replication plans. A Lab Manager enforces real constraints -- GPU budgets, scheduling conflicts, equipment limits. A deterministic Judge scores every plan. Through RL, the Scientist gets measurably better at navigating tradeoffs."

3. Live Demo (60 seconds)

• Show HF Space or local frontend
• Start an ML Benchmark episode (seed 42, medium difficulty)
• Point out the Scientist's proposal and Lab Manager's feasibility report
• Show the Judge scoring: rigor, feasibility, fidelity breakdown
• Toggle to training results: before vs after comparison

4. Technical Architecture (30 seconds)

"Three scenario families -- math, ML, finance -- each with deterministic seed-based generation. The reward formula is multiplicative: 10 x rigor x feasibility x fidelity. Every dimension must score well. The entire judge is deterministic -- same seed, same actions, same score. No LLM-as-judge variance."

5. Results (20 seconds)

"After RL training: 67% higher reward, 32% fewer negotiation rounds, invalid actions drop from 15% to 4%, agreement rate jumps from 50% to 80%."

6. Close (10 seconds)

"ReplicaLab. An OpenEnv world where agents learn to negotiate science."

---

Anticipated Q&A Topics

Question	Talking Points
Why deterministic scoring?	Noisy rewards make RL unstable. Deterministic judge = reproducible training. Optional Oracle layer adds richness without corrupting the reward signal.
How does difficulty scaling work?	Mechanical constraint tightening: budgets shrink, resources go out of stock, scheduling conflicts appear. Same outer contract at every difficulty.
What model do you train?	Qwen3-8B with GRPO via Unsloth/TRL. 4B fallback for faster iteration.
How many scenarios?	3 domain families x 3 difficulties x infinite seeds. Each seed produces a unique but deterministic scenario.
Why not LLM-as-judge?	Variance. Two runs of the same episode would get different scores. We need a stable reward signal for RL. The optional Oracle post-mortem adds natural language analysis without replacing the score.
What's the Lab Manager?	Hybrid: deterministic feasibility checker (ground truth) + optional model narration. Checker output is always the source of truth.
Fallback if UI breaks?	/web endpoint serves a self-contained HTML interface with no build step.