ReplicaLab_50_Scenarios_Training_Plan.md

ReplicaLab: 50 Scenario Templates and Training Plan

Domain Distribution

Domain	Count	Rationale
Computational ML/DL	20	Most relatable to judges, richest compute constraint space
Wet-Lab Biology	16	Strongest replication crisis narrative, most varied equipment
Quantitative Finance	14	Broadest appeal, most concrete measurable constraints

---

Domain 1: Computational ML/DL (20 Scenarios)

Cluster A: Training Replication (7 papers)

These are "we trained a model and got results" papers. The core tension is always compute, data, and time.

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
1	ResNet Depth Scaling on ImageNet	Deeper networks improve accuracy up to 152 layers	ResNet architecture with skip connections	8xV100, 90 epochs, full ImageNet (1.2M images)	Lab has 1xH100, budget for 30 epochs, only ImageNet-100 subset
2	BERT Fine-Tuning for Sentiment	BERT-large fine-tuned beats all baselines on SST-2	BERT-large 340M params, AdamW	4xA100 80GB, SST-2 full, 3 epochs	Lab has 1x40GB GPU, must use BERT-base or quantized BERT-large
3	Diffusion Model for Image Synthesis	DDPM generates high-fidelity 256x256 faces	U-Net with 1000 diffusion steps	8xA100, CelebA-HQ, 500K steps	Lab has 1xH100, budget for 100K steps, only CelebA (not HQ)
4	RL Agent for Atari Games	PPO agent achieves superhuman on 40/57 Atari games	PPO with frame stacking	256 CPU actors, 1xGPU learner, 200M frames	Lab has 16 CPU cores, 1xGPU, budget for 10M frames, test on 5 games only
5	GAN Training Stability	StyleGAN2 produces photorealistic 1024x1024 output	Progressive growing, R1 regularization	8xV100, FFHQ 70K images, 25M images shown	Lab has 1xH100, only FFHQ 10K subset, budget for 5M images shown
6	Vision Transformer Pretraining	ViT-Large pretrained on JFT-300M matches CNN	ViT-L/16 with patch embedding	TPUv3 pod, JFT-300M (proprietary), 300 epochs	Lab has 1xH100, only ImageNet-21K (public), ViT-Base budget only
7	LLM Instruction Tuning	SFT on curated instructions improves helpfulness	LoRA on 7B base model	4xA100, 50K curated instructions, 3 epochs	Lab has 1xH100, only 10K public instructions (Alpaca), rank-16 LoRA max

Cluster B: Evaluation/Benchmark Replication (6 papers)

These are "we evaluated X and found Y" papers. Tension is around evaluation methodology and data access.

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
8	LLM Benchmark Contamination	GPT-4 performance drops 12% on decontaminated MMLU	Custom decontamination pipeline	Full MMLU, GPT-4 API ($2K budget), custom regex filters	Lab has $200 API budget, must use open-source LLM, no custom decontamination tool
9	Fairness Audit of Hiring Model	Commercial hiring model shows 23% TPR gap across demographics	Adversarial probing with synthetic candidates	Access to proprietary model API, 10K synthetic resumes, 6 demographic axes	Lab has no API access, must train proxy model, budget for 2K synthetic resumes
10	Cross-lingual Transfer	mBERT zero-shot works for NER in 40 languages	mBERT with English-only fine-tuning	All 40 CoNLL languages, mBERT-base	Lab has compute for 10 languages, some language datasets have licensing issues
11	OOD Detection Benchmark	Energy score beats MSP on 6 OOD benchmarks	Energy-based OOD scoring	6 OOD datasets, ResNet-18 pretrained, custom evaluation suite	Lab missing 2 of 6 datasets (licensing), must justify subset evaluation
12	Prompt Sensitivity Study	GPT-3.5 accuracy varies 15% across prompt formats	Systematic prompt variation, 50 formats	GPT-3.5 API ($1.5K budget), 50 prompt templates, 5 benchmarks	Lab has $300 budget, can test 15 formats on 3 benchmarks
13	Model Compression	4-bit quantized LLaMA-7B retains 95% of quality	GPTQ quantization	Full LLaMA-7B weights, custom GPTQ kernel, 8 benchmarks	Lab has weights but GPTQ kernel incompatible with CUDA version, must use alternative quantizer

Cluster C: Method/Architecture Replication (7 papers)

These are "we propose method X and it outperforms baselines" papers. Tension is around implementation fidelity and baseline reproduction.

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
14	Attention Mechanism Ablation	Multi-head attention outperforms single-head by 2.1 BLEU	Transformer encoder-decoder	4xV100, WMT14 En-De (4.5M pairs), custom tokenizer	Lab has 1xH100, WMT14 subset (1M pairs), must use HuggingFace tokenizer
15	Contrastive Learning for Vision	SimCLR outperforms supervised pretraining with 1% labels	SimCLR with large batch (4096)	128 TPU cores, ImageNet, batch size 4096	Lab has 1xH100, max batch 256 (need gradient accumulation), memory constraints
16	Graph Neural Network for Molecules	GIN outperforms GCN on molecular property prediction	Graph Isomorphism Network	8 molecular datasets, custom data pipeline, RDKit preprocessing	Lab missing RDKit (incompatible Python version), 5 of 8 datasets available
17	Knowledge Distillation	DistilBERT retains 97% of BERT performance at 60% size	Task-agnostic distillation	BERT-base teacher, BookCorpus+Wikipedia, 3 days training	Lab has BERT-base but BookCorpus no longer publicly available, Wikipedia only
18	Neural Architecture Search	DARTS finds architecture matching hand-designed on CIFAR-10	Differentiable architecture search	1xV100 for search (1.5 days), 1xV100 for evaluation	Lab has 1xH100 (faster) but only 8 hours allocated, must reduce search space
19	Data Augmentation	RandAugment matches AutoAugment without search cost	Random augmentation policy	ResNet-50, ImageNet, 270 epochs, grid search over N and M	Lab has compute for 90 epochs, budget for partial grid search (5 of 15 configs)
20	Federated Learning	FedAvg converges with 100 non-IID clients	Federated averaging	100 simulated clients, CIFAR-10, 500 communication rounds	Lab can simulate 20 clients, budget for 200 rounds, must argue this is sufficient

---

Domain 2: Wet-Lab Biology (16 Scenarios)

Cluster D: Cell Biology and Biochemistry (8 papers)

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
21	Drug Cytotoxicity Dose-Response	Compound X has IC50 of 2.3 uM against HeLa cells	MTT assay, 8-point dose-response	Plate reader, MTT reagent, HeLa cells, 96-well plates, n=6 replicates	Lab plate reader booked Mon-Wed, MTT backordered (WST-1 available), budget for n=4
22	siRNA Knockdown Efficiency	siRNA targeting BRCA1 achieves 85% knockdown	qPCR quantification, lipofection	Real-time PCR machine, lipofectamine, BRCA1 primers, Western blot validation	qPCR machine shared (available Thu-Fri only), no Western blot antibody in stock
23	Protein Expression and Purification	Recombinant GFP-tagged protein expressed in E. coli at 50 mg/L	IPTG induction, Ni-NTA purification	Shaker incubator, FPLC, Ni-NTA resin, IPTG, competent cells	FPLC needs maintenance (2 days), can use gravity column instead, slower but cheaper
24	Flow Cytometry Apoptosis	Drug Y induces 60% apoptosis via Annexin V/PI staining	Flow cytometry with dual staining	Flow cytometer, Annexin V kit, PI, cell culture facility	Flow cytometer calibration expired, Annexin V kit expires in 5 days (cutting it close)
25	Wound Healing Migration	Compound Z accelerates wound closure by 40% in 24h	Scratch assay with time-lapse imaging	Inverted microscope with camera, cell culture hood, 6-well plates, n=5	Microscope camera resolution lower than paper (can we still quantify?), n=3 budget
26	CRISPR Gene Editing	CRISPR-Cas9 knockout of TP53 in MCF-7 cells	CRISPR with guide RNA, Sanger sequencing	Electroporation system, guide RNA, Cas9 protein, sequencing service	Electroporation system unavailable, must use lipofection (lower efficiency expected)
27	Enzyme Kinetics	Km of novel enzyme variant is 15 uM	Michaelis-Menten kinetics, spectrophotometric assay	UV-Vis spectrophotometer, substrate concentrations (10 points), purified enzyme	Spectrophotometer wavelength range limited, 6 concentration points max (budget)
28	Bacterial Growth Curve	Antibiotic resistance mutation confers 3x MIC increase	Broth microdilution, OD600 measurement	Plate reader (kinetic mode), Mueller-Hinton broth, antibiotic stock, 12h monitoring	Plate reader does not support kinetic mode, must do manual timepoint readings

Cluster E: Behavioral and Cognitive (4 papers)

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
29	Ego Depletion Replication	Self-control depletion reduces performance on Stroop task	Sequential task paradigm	n=200 participants, Stroop software, two-room setup, 4 experimenters	IRB timeline 3 weeks, budget for n=80, 1 experimenter available, one room
30	Priming Effect on Behavior	Exposure to achievement words improves puzzle performance	Scrambled sentence priming	n=150, computerized tasks, between-subjects design, debriefing protocol	n=60 budget, online-only (no in-person), must address demand characteristics
31	Sleep and Memory Consolidation	8h sleep improves word-pair recall by 25% vs sleep deprivation	Within-subjects, polysomnography	Sleep lab, PSG equipment, n=30, 2 sessions per participant	No sleep lab access, must use actigraphy (wrist device) as proxy, n=15
32	Social Conformity in Groups	Group pressure changes individual opinions 35% of the time	Asch-style paradigm with confederates	4 trained confederates, n=100 naive participants, recording equipment	Budget for 2 confederates, n=40, must justify reduced group size

Cluster F: Environmental and Ecological (4 papers)

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
33	Soil Microbiome Diversity	Fertilizer reduces bacterial diversity by 30%	16S rRNA sequencing, alpha diversity	Sequencing service, soil sampling kit, 20 sites, triplicate	Sequencing budget for 10 sites only, duplicate instead of triplicate
34	Water Pollutant Detection	Novel biosensor detects lead at 5 ppb sensitivity	Electrochemical impedance spectroscopy	Potentiostat, custom electrode, calibration standards, DI water system	Potentiostat model different from paper (lower frequency range), must validate equivalence
35	Plant Growth Under LED Spectra	Blue-enriched LED increases lettuce biomass 20%	Controlled growth chamber, spectral analysis	Growth chamber (4 compartments), LED panels, 30-day trial, 20 plants per group	Growth chamber has 2 compartments (not 4), must run sequential instead of parallel
36	Algal Bloom Prediction	Phosphorus concentration predicts bloom onset within 5 days	Spectrophotometric phosphorus assay, regression model	Lake access permit, sampling boat, reagents for 100 samples, 6-month dataset	Permit pending (2 weeks), budget for 50 samples, 3-month window only

---

Domain 3: Quantitative Finance (14 Scenarios)

Cluster G: Trading Strategy Replication (6 papers)

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
37	Momentum Factor Premium	10-day/50-day MA crossover generates 12% annual excess return	Moving average crossover, Fama-French regression	Tick-level data, S&P 500 (20 years), Bloomberg terminal	Daily OHLCV only, 10-year window, no Bloomberg (use yfinance), survivorship bias
38	Pairs Trading Mean Reversion	Cointegrated equity pairs yield 8% annual Sharpe 1.5	Engle-Granger cointegration, Kalman filter	Intraday data, 200 pairs, $0.005/share commission model	Daily data, budget to test 50 pairs, commission model is $0.01/share
39	Volatility Risk Premium	Selling VIX puts captures 4% monthly premium	Options pricing, delta hedging	Options chain data (CBOE), VIX futures, real-time Greeks	No options data subscription, must use delayed data, no real-time Greeks
40	Earnings Momentum	Post-earnings drift persists for 60 days	Event study, CAR calculation	Earnings calendar (10 years), intraday returns around announcements	Only daily returns, 5-year earnings calendar (free source), must use wider event window
41	Crypto Market Microstructure	Bitcoin bid-ask spread predicts 1h returns	Order book analysis, microstructure model	L2 order book data (Binance), 1-second resolution, 6 months	No L2 data, only L1 (best bid/ask) from free API, 3-month window
42	Factor Timing with Macro Signals	Yield curve slope predicts value/growth rotation	Multi-factor model with macro overlay	Factor returns (AQR), yield curve data (FRED), 30 years	AQR data has 3-month publication lag, 20-year window from FRED, must handle shorter overlap

Cluster H: Risk and Valuation Replication (4 papers)

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
43	VaR Model Backtesting	Historical VaR at 99% underestimates tail risk by 40%	Historical simulation, 10K scenarios	20 years of daily portfolio returns, Monte Carlo (100K paths)	10-year data window, compute budget for 10K Monte Carlo paths, must justify reduced sample
44	Credit Risk Transition Matrix	BBB-to-default probability is 0.3% annual (S&P estimate)	Cohort analysis of rating transitions	S&P rating database (proprietary, 30 years), 5K issuers	No S&P database, must use Moody's public reports (summary statistics only), reconstruct from aggregated data
45	Real Estate Cap Rate Model	Cap rate spread over 10Y treasury predicts REIT returns	Regression model with macro factors	NCREIF property index, 10Y treasury (FRED), REIT returns (CRSP)	NCREIF is proprietary, must use publicly available REIT index as proxy, shorter time series
46	Portfolio Optimization	Black-Litterman outperforms mean-variance by 200bps	Black-Litterman with investor views	Covariance matrix (60 assets, 10 years daily), equilibrium returns	Only 30 assets available (data cost), weekly instead of daily data, must address estimation error

Cluster I: Behavioral Finance and Market Anomalies (4 papers)

#	Paper Title	Claim	Key Technique	Original Resources	Primary Constraint Tension
47	Disposition Effect in Retail Trading	Retail traders sell winners 1.5x faster than losers	Trade-level analysis of brokerage accounts	Proprietary brokerage dataset (100K accounts, 5 years)	No brokerage data, must use public datasets (Robinhood 2021 leak or academic dataset)
48	Sentiment and Returns	Twitter sentiment predicts next-day S&P 500 direction	NLP sentiment analysis, Granger causality	Twitter firehose (1M tweets/day), FinBERT, 3 years	No Twitter firehose (API deprecated), must use Reddit or news headlines, smaller sample
49	January Effect Persistence	Small-cap excess returns in January have declined since 1990	Calendar anomaly study, size-sorted portfolios	CRSP daily returns (60 years), size quintile breakpoints	Only 20 years of free data (Yahoo), must construct size portfolios from available universe
50	IPO Underpricing	Average first-day IPO return is 18% with high variance	Event study of IPO first-day returns	SEC EDGAR filings, IPO database (30 years, 5K IPOs)	Free IPO data covers 10 years only (1.5K IPOs), missing some small IPOs, survivorship concern

---

Difficulty Calibration

Each scenario gets tagged with a difficulty. The Oracle uses this to adjust how severe the constraints are, but the base template defines the core tension.

Difficulty	Constraint Profile	Target Reward Range
Easy	1-2 conflicts, clear substitutions exist, budget is 80% of needed	6.0-8.5
Medium	3-4 conflicts, substitutions require tradeoffs, budget is 50-70% of needed	3.5-6.5
Hard	5+ conflicts, substitutions are risky, budget is 30-50% of needed, time pressure	1.5-4.5

Distribution across 50 scenarios:

• Easy: 15 (30%)
• Medium: 20 (40%)
• Hard: 15 (30%)

During training, use curriculum learning: start with easy, shift to medium by iteration 5, introduce hard by iteration 10.

---

What Each Scenario Template Must Define

The Oracle generates the full scenario, but your template gives it guardrails. Each template is a compact JSON/Python dict:

SCENARIO_TEMPLATES = {
    "ml_resnet_depth": {
        "id": 1,
        "domain": "computational_ml",
        "difficulty_range": ["easy", "medium", "hard"],
        "paper_seed": {
            "title": "ResNet Depth Scaling on ImageNet",
            "claim": "Deeper networks improve accuracy up to 152 layers",
            "technique": "ResNet with skip connections",
            "original_compute": "8xV100, 90 epochs, full ImageNet",
            "original_sample_size": 1281167,  # ImageNet train size
            "original_duration": "72 hours",
            "statistical_test": "top-1/top-5 accuracy, t-test across 3 seeds",
            "required_controls": [
                "baseline_shallow_model",
                "learning_rate_schedule",
                "data_augmentation_pipeline"
            ],
        },
        "constraint_seed": {
            "equipment_pool": ["gpu_h100", "gpu_a100_40gb", "gpu_v100", "cpu_cluster"],
            "data_pool": ["imagenet_full", "imagenet_100", "imagenet_10pct", "cifar100_proxy"],
            "typical_budget_range": [500, 5000],  # USD compute cost
            "time_range_hours": [8, 72],
            "common_bottlenecks": [
                "gpu_memory_for_batch_size",
                "dataset_download_time",
                "library_version_incompatibility",
                "checkpoint_storage"
            ],
            "valid_substitutions": [
                {"original": "imagenet_full", "substitute": "imagenet_100", "validity": "acceptable_with_caveats", "caveat": "must acknowledge reduced class diversity"},
                {"original": "8xV100", "substitute": "1xH100", "validity": "equivalent", "caveat": "adjust batch size, use gradient accumulation"},
                {"original": "90_epochs", "substitute": "30_epochs", "validity": "inferior_but_usable", "caveat": "may not reach full convergence, report learning curve"},
            ],
        },
        "scoring_hints": {
            "critical_controls": ["baseline_shallow_model", "learning_rate_schedule"],
            "flexible_controls": ["data_augmentation_pipeline"],
            "min_sample_fraction": 0.1,  # at least 10% of original data
            "power_notes": "accuracy differences < 0.5% require large n to detect",
        },
    },
    # ... 49 more templates
}

You do NOT write all 50 as fully fleshed-out dicts before the hackathon. You write 5-6 detailed templates (2 per domain) and let the Oracle interpolate the rest. The template gives the Oracle enough domain knowledge to generate a consistent scenario.

---

Training Plan for 3 Hours on H100

The Math

Model: Qwen2.5-7B-Instruct or LLaMA-3-8B-Instruct with LoRA (rank 16) Method: GRPO via TRL or Unsloth GPU: 1xH100 80GB

Time budget breakdown:

Phase	Time	What Happens
Setup and warmup	15 min	Load model, verify env loop, run 2 test episodes
Pre-generate scenarios	15 min	Call Oracle World Architect for all seeds, cache to disk
Training	2 hr 15 min	GRPO iterations
Final evaluation	15 min	Run eval episodes, generate reward curve

Pre-Generation Phase (Critical)

Before training starts, pre-generate and cache all scenarios you will use. This removes the Oracle API bottleneck from the training loop entirely.

50 scenario templates × 3 difficulty variants = 150 unique scenarios
Oracle World Architect call: ~4 sec each
Total: 150 × 4 = 600 sec = 10 minutes

Cache all 150 to disk as JSON.

During training, reset() loads from cache. Zero API latency.

The Bottleneck Shift

With cached scenarios, the per-episode bottleneck becomes the Lab Manager LLM calls (one per round). Two options:

Option A: LLM Lab Manager (richer but slower)

• 6 rounds × ~2.5 sec per LM call = 15 sec per episode for LM
• Plus Adjudicator calls: 6 × 2.5 sec = 15 sec
• Total API time per episode: ~30 sec
• GPU time per episode (Scientist inference): ~2 sec
• Wall time per episode: ~32 sec

Option B: Rule-based Lab Manager for training, LLM for demo (faster)

• 6 rounds × 0 sec API = 0 sec for LM
• Adjudicator: can also be made deterministic for training
• Total API time per episode: 0 sec
• GPU time per episode: ~2 sec + ~1 sec overhead
• Wall time per episode: ~3 sec

I strongly recommend Option B for training. Use the rule-based Lab Manager and deterministic Adjudicator during RL training for speed, then switch to LLM Lab Manager and Oracle Adjudicator for demo and evaluation. The Scientist does not know the difference, it still sees the same observation schema.

Episodes per Hour with Option B

Parallel Rollouts	Episode Time	Episodes/Hour
1	~3 sec	~1,200
4 (batch)	~3 sec (batched inference)	~4,800
8 (batch)	~3.5 sec	~8,200

With batched inference (8 parallel rollouts), you get roughly 8,000 episodes per hour.

GRPO Training Schedule

GRPO collects a batch of rollouts, computes advantages, and updates the model. Here is the schedule:

GRPO config:
  rollout_batch_size: 32 episodes per update
  num_iterations: 40
  total_episodes: 32 × 40 = 1,280
  
  Per iteration:
    Rollout collection (32 episodes, 8 parallel): ~12 sec
    Advantage computation: ~2 sec
    Gradient update (LoRA rank 16, 7B model): ~45 sec
    Logging and checkpoint: ~5 sec
    Total per iteration: ~64 sec ≈ ~1 min

  40 iterations × 1 min = 40 minutes

Wait. That is only 40 minutes. You have 2 hours 15 minutes of training time. So you can do much more:

Revised GRPO config:
  rollout_batch_size: 64 episodes per update
  num_iterations: 80
  total_episodes: 64 × 80 = 5,120
  
  Per iteration:
    Rollout collection (64 episodes, 8 parallel): ~24 sec
    Advantage computation: ~3 sec
    Gradient update: ~55 sec
    Logging: ~5 sec
    Total per iteration: ~87 sec ≈ ~1.5 min

  80 iterations × 1.5 min = 120 min = 2 hours

Final training plan: 5,120 episodes across 80 GRPO iterations in ~2 hours.

Curriculum Schedule

Iterations	Difficulty Mix	Domains
1-20	80% easy, 20% medium	ML/DL only (most constrained, clearest signal)
21-40	40% easy, 50% medium, 10% hard	ML/DL + Biology
41-60	10% easy, 50% medium, 40% hard	All three domains
61-80	0% easy, 30% medium, 70% hard	All three domains, hardest scenarios

Scenario Sampling During Training

With 150 cached scenarios and 5,120 episodes, each scenario gets used ~34 times on average. But you seed the randomness, so:

• Iteration 1-20: sample from ML easy/medium scenarios (templates 1-20, easy+medium variants = ~40 scenarios)
• Iteration 21-40: add Biology (templates 21-36 = ~32 more scenarios)
• Iteration 41-80: add Finance (templates 37-50 = ~28 more scenarios), shift to harder variants

The Scientist sees enough variety to generalize while getting repeated exposure to learn each domain.

---

Evaluation Plan (Final 15 Minutes)

Held-Out Evaluation Set

Reserve 10 scenarios per domain (30 total) that are NEVER used during training. Different seeds, same templates but with constraint variations the Scientist has not seen.

Evaluation Runs

30 held-out scenarios × 1 run each = 30 episodes
Wall time: 30 × 3 sec = 90 sec (with rule-based LM)

Then run 5 showcase episodes with LLM Lab Manager + Oracle:
5 × 50 sec = 250 sec ≈ 4 min

Total eval time: ~6 minutes (well within 15 min budget)

Metrics to Report

Metric	Untrained (Baseline)	Trained (Post-GRPO)
Mean total reward	Measure in Phase 2	Measure here
Mean rigor score
Mean feasibility score
Mean fidelity score
Rounds to agreement
Invalid action rate
Contradiction rate
Agreement rate (vs timeout)

The Reward Curve

Plot every 5 iterations:

• X axis: GRPO iteration (0 to 80)
• Y axis: mean reward over last batch
• Include error bars (std across batch)
• Overlay the difficulty curriculum as background color

This is the single most important artifact for judges. It must show a clear upward trend.

---

What You Actually Build Before Training

Day-of Priority Order

1. models.py (30 min) All Pydantic models from the Oracle guide. These are your contract.

2. oracle.py with World Architect mode only (45 min) Get scenario generation working. Test with 3 seeds. Cache results.

3. replicalab_env.py with rule-based Lab Manager (1 hour) The fast training loop. No LLM Lab Manager. Deterministic adjudicator. Must pass: reset returns observation, step returns observation + reward, episode terminates.

4. scoring/reward.py deterministic reward computation (30 min) The arithmetic layer. Takes protocol + hidden spec, outputs scores.

5. 6 detailed scenario templates (30 min) 2 per domain. These seed the Oracle and serve as rule-based fallbacks.

6. GRPO training script (1 hour) Connect TRL/Unsloth to the env. Verify one iteration works.

7. Pre-generate 150 scenarios (15 min) Run the Oracle, cache everything.

8. Start training (2 hours, runs while you build the demo)

9. lab_manager_agent.py LLM version (30 min, while training runs) Only used for demo. Not needed for training.

10. Oracle Adjudicator + Post-Mortem (30 min, while training runs) Only used for demo and eval showcase episodes.

What Can Run in Parallel

While the H100 is training (2 hours), your team builds:

• LLM Lab Manager (Person 2)
• Oracle Adjudicator + Post-Mortem (Person 2)
• React UI (Person 4)
• Demo script and YouTube recording prep (Person 4)
• FastAPI + WebSocket server (Person 3)
• HF Space Dockerfile (Person 3)

The H100 only needs ~30% utilization for GRPO training with LoRA. The remaining GPU capacity can run the Scientist inference for evaluation episodes simultaneously if you architect the training script to do periodic eval checkpoints.

---

Summary

Item	Number
Total scenario templates	50
ML/DL	20
Biology	16
Finance	14
Cached scenario variants (with difficulty)	150
Training episodes	5,120
GRPO iterations	80
Training wall time	~2 hours
Eval episodes	30 (fast) + 5 (showcase)
Total H100 time	~2.5 hours (within 3-hour budget)
Scientist model	7B-8B with LoRA rank 16
Lab Manager (training)	Rule-based (fast)
Lab Manager (demo)	LLM (rich)
Oracle calls during training	0 (all cached)
Oracle calls during demo	Full (all 4 modes live)