---
name: claim-evidence-chain
description: Argumentation template for constructing paper sections. Each claim must be paired with explicit evidence and an explicit narrative function. Loaded by outline-architect and draft-writer.
---

# Claim-Evidence Chain

This skill encodes the argumentative discipline used to build sections. The unit of paper structure is the **claim**, not the paragraph or section.

## The Three-Part Unit

Every claim in your outline / draft has THREE parts:

```
CLAIM:    The proposition you assert.
EVIDENCE: The specific data / citation / equation that supports it.
FUNCTION: Why this claim is HERE, in THIS section, in THIS order.
```

A claim with no FUNCTION is decoration — delete it.
A claim with no EVIDENCE is hand-waving — find evidence or delete it.
A claim with no CLAIM (just data) is a result without an argument — make the assertion explicit.

## Claim Quality Checklist

For each claim, verify:

- [ ] **Specific**: Not "our method works well" but "our method achieves AUC X on task Y"
- [ ] **Falsifiable**: A reader could in principle disagree with it
- [ ] **Bounded**: The claim's scope is clear (don't overclaim)
- [ ] **Non-redundant**: Not stated elsewhere in the paper

## Evidence Types and What They Can Support

| Evidence Type | Can Support | Cannot Support |
|---------------|-------------|----------------|
| Single experiment | "X happened in this setting" | "X happens in general" |
| Ablation | "Component Y contributes to performance" | "Y is necessary in all settings" |
| Citation | "Prior work observed X" | "X is true" |
| Theoretical proof | "Under assumptions A, X holds" | "X holds in practice" |
| Visualization | Existence claims, qualitative trends | Quantitative magnitude |
| Statistical test | "Difference is unlikely under null" | "Effect is large/important" |

Use this table to check whether the EVIDENCE actually supports the CLAIM. Mismatches are the #1 source of reviewer attacks.

## Function Categories

A claim's FUNCTION in the narrative is one of:

1. **Setup** — establishes a definition or fact later claims will use
2. **Motivation** — argues for why the work matters
3. **Result** — reports a finding from the work itself
4. **Interpretation** — offers a reading of a result
5. **Comparison** — places the work relative to existing work
6. **Limitation** — bounds the scope of a result
7. **Implication** — extends from the result to broader consequences

Every claim should fit cleanly into one. A claim that doesn't fit is probably either misplaced or split-able.

## Section-Level Patterns

### IMRaD-style sections
- **Introduction**: Setup → Motivation → Comparison → Result-preview
- **Method**: Setup → (Setup → Setup → ...) — methods are usually pure setup chains
- **Results**: Result → Interpretation → Result → Interpretation → ...
- **Discussion**: Implication → Limitation → Implication → ...

### Benchmark-paper sections
- **Tasks/Metrics**: Setup chain
- **Baselines**: Setup → Comparison
- **Analysis**: Result → Interpretation chains, often with **load-bearing claims**

## Load-Bearing Claims

Mark claims that the entire paper depends on with `[LOAD-BEARING]`. Example:

> "Our active-learning strategy outperforms random sampling on out-of-distribution tasks. [LOAD-BEARING]"

Load-bearing claims must:
- Be supported by the strongest evidence in the paper
- Be defensible against the most hostile reviewer reading
- Survive the Devil's Advocate Attack pass from peer-reviewer

A typical paper has 2-4 load-bearing claims. More than 6 means the argument is overloaded.

## Anti-Patterns

- **Claim laundering**: Citing your own earlier section as evidence for a claim, in a way that creates circular support
- **Bait-and-switch**: Introduction promises X, paper actually shows X' (weaker version)
- **Evidence padding**: Stacking weak evidence to feel like strong evidence — three weak signals don't equal one strong signal
- **Conclusion creep**: Discussion section claims things stronger than the Results actually showed