# Research Site

Use when the audience needs to understand or evaluate a source-grounded argument rather than operate a simulated product.

## Required outcome

The reader should be able to identify the central thesis, inspect its evidence, understand the mechanism, and see where the conclusion stops being reliable.

## Recommended information architecture

1. Thesis and audience-relevant takeaway.
2. Evidence basis with provenance and confidence.
3. Mechanism or causal explanation.
4. **Concrete Code Implementation / Architecture (for technical/algorithmic studies):**
   - Provide concrete, production-style, executable PyTorch code architectures (or equivalent system implementations) demonstrating how theoretical mechanics run in code.
   - For multi-paradigm, multi-regime, or comparative algorithms (e.g. sequential generation vs parallel rollouts vs tree search), provide interactive, tabbed code inspectors displaying clean, syntax-highlighted implementations with tensor shapes, KV-cache handling, and evaluation hooks.
5. Competing explanation, limitation, or boundary.
6. Practical implication and invalidation condition.

Use an evidence matrix, comparison, timeline, code inspector, or dependency view only when it matches the information topology. Interaction should help the reader compare evidence, inspect code, or trace a claim; it must not hide the core argument.

## Required evaluation emphasis

- factual and citation fidelity;
- thesis-to-evidence alignment;
- mechanism depth;
- visible uncertainty and counter-evidence;
- information density and reading flow;
- requested interactive effects.

Visual polish cannot compensate for unsupported conclusions.