Stage 3: Baseline-Digestion

After Ideation confirms a viable idea and base paper, the Baseline-Digestion stage has Scout digest the baseline paper and its source code into structured reference files. These files become the foundation that downstream agents (Planner, Coder) build upon.

Entering This Stage

What you have:

• A selected idea with baseline paper (ideas/selected.yaml)
• The baseline paper PDF and source code repository

What you don't have yet:

• Structured understanding of the baseline's method, experiments, or code
• Experiment design or implementation plan

Steps

graph TD
    A[1. Paper Reading] --> B[2. Code Scanning]
    B --> C[3. File Generation]
    C --> D[4. Validation Check]
    D --> E[Gate: auto → Advance to Design]

    style A fill:#dbeafe,stroke:#2563eb
    style C fill:#fef3c7,stroke:#d97706
    style E fill:#dcfce7,stroke:#16a34a

1. Paper Reading

Agent: Scout (Gemini)

Scout reads the baseline paper and extracts:

• Core method description (architecture, equations, loss functions)
• Experimental setup (datasets, metrics, hyperparameters)
• Key results and claims
• Ablation findings

2. Code Scanning

Agent: Scout (Gemini)

Scout scans the baseline source code repository and extracts:

• Code structure and directory layout
• Key dependencies and framework versions
• Entry points for training and evaluation
• Configuration files and default hyperparameters

3. File Generation

Agent: Scout (Gemini)

Scout produces exactly 5 structured files in .omc/research/baseline/:

File	Tier	Contents	Consumed By
summary.md	Tier 1 (injected to all agents)	One-page overview of the method, its contribution, and key results	All agents
experiments.md	Tier 2	Datasets used, evaluation metrics, hyperparameter tables, training schedules	Planner
method-details.md	Tier 2	Architecture diagrams, equations, loss formulations, design rationale	Coder
code-notes.md	Tier 2	Code structure, dependencies, entry points, configuration patterns	Coder
reproduction-checklist.md	Tier 2	Numbers to reproduce, known risks, hardware requirements, common pitfalls	Coder

summary.md is Tier 1

summary.md is the only file injected into every agent's context by default. It gives all agents a shared understanding of the baseline without consuming excessive context. The other four files are read on demand by the agents that need them.

4. Validation Check

Agent: Scout (Gemini)

Scout performs a self-check to ensure:

• All 5 files are present and non-empty
• summary.md is concise (under 500 lines)
• Key numbers from the paper appear in experiments.md
• Entry points referenced in code-notes.md exist in the repository

Output Files

.omc/research/baseline/
├── summary.md                  # Tier 1: injected to all agents
├── experiments.md              # Datasets, metrics, hyperparams
├── method-details.md           # Architecture, equations, losses
├── code-notes.md               # Code structure, dependencies, entry points
└── reproduction-checklist.md   # Numbers to match, risks

summary.md

The one-page executive summary. Contains:

• Paper title, authors, venue
• Problem statement (1-2 sentences)
• Core method (1 paragraph)
• Key results (table of main numbers)
• Strengths and limitations (bullet points)

This file is short by design. Every agent receives it in context, so brevity is critical.

experiments.md

Detailed experimental setup:

• Dataset names, sizes, splits, preprocessing
• Evaluation metrics with exact definitions
• Hyperparameter tables (learning rate, batch size, scheduler, etc.)
• Training duration (steps, epochs, wall-clock time)
• Hardware used in the original paper

method-details.md

Technical depth on the method:

• Architecture description with layer-by-layer breakdown
• Key equations (numbered, with variable definitions)
• Loss function components and weighting
• Design rationale and ablation-supported choices

code-notes.md

Practical code understanding:

• Repository structure (key directories and files)
• Dependencies and version requirements
• Training entry point and CLI arguments
• Evaluation entry point and expected outputs
• Configuration file format and important defaults

reproduction-checklist.md

Concrete reproduction targets:

• Exact numbers to match from the paper (with table/figure references)
• Acceptable tolerance ranges
• Known reproduction risks (e.g., "requires specific random seed", "sensitive to LR warmup")
• Hardware-specific considerations
• Estimated reproduction time and compute cost

Gate

Gate Type	Recommended	Behavior
auto	Yes	Pure information extraction; no judgment calls needed
auto-judge	Possible	Judge verifies completeness of extracted files
human	Not recommended	Adds delay with minimal benefit for extraction tasks

Why auto is the default

Baseline-Digestion is pure information extraction — Scout reads a paper and code, then writes structured notes. There are no subjective decisions or creative choices that benefit from human review. The validation check in step 4 catches missing or incomplete output.

Downstream Consumers

The 5 baseline files feed directly into later stages:

Consumer	Files Used	Purpose
Planner (Design stage)	summary.md + experiments.md	Understand what to compare against, what metrics to target
Coder (Implementation stage)	method-details.md + code-notes.md + reproduction-checklist.md	Understand what to implement, how the baseline code works
All agents	summary.md	Shared baseline context

Error Handling

Error	Recovery
Paper PDF unavailable	Scout searches for alternative sources (arXiv, author page)
Source code repository missing	Scout notes this in code-notes.md; Coder implements from paper description alone
Key numbers not found in paper	Scout marks gaps in reproduction-checklist.md with "NOT FOUND" annotations
Output file missing or empty	Scout re-runs the extraction for the missing file

Next Stage

When the gate passes, the pipeline advances to Design with all 5 baseline files available in .omc/research/baseline/.