# Draft Claims Analysis — Demo Results

## Document Analyzed
- **Title**: How to Do Great Work — Paul Graham (July 2023)
- **Source**: paulgraham.com/greatwork.html
- **Length**: ~66,500 chars (~11,000 words)

## What Was Built

### `draft claims` CLI command (new)
```bash
draft claims FILE_OR_URL [--model sonnet]
```
Produces an HTML report with:
- **Left column**: Document text with claims highlighted (red border/background)
- **Right margin**: Annotations for each substantive claim — short label + novelty/insight score
- **Top summary**: Highest-scored claims ranked, clickable to scroll to location
- **Filtered section**: Process descriptions / framework labels that were misclassified as "claims"

### Claim quality filter
Claims rated N+I < 5 are demoted to "process descriptions" — shown with dotted underlines, no margin note. Examples filtered:
- "three qualities of work" (N:2, I:2) — framework, not debatable assertion
- "reach frontier of knowledge" (N:1, I:2) — process step
- "hard work is essential" (N:1, I:1) — conventional wisdom
- "undirected thinking solves problems" (N:1, I:2) — description

### Files created
- `draft/claims/report.py` — Claims-focused HTML renderer with margin annotations
- `draft/cli.py` — Added `draft claims` command
- `reports/tmp/draft-demo-polish-20260309.html` — Full report on PG essay

## Results on "How to Do Great Work"

| Metric               | Value |
|----------------------|-------|
| Total segments       | 203   |
| Substantive claims   | 105   |
| Process descriptions | 39    |
| Other roles          | 59    |

### Top-scoring claims (N+I >= 8)
1. **experts should be puzzled** (N:4, I:5) — "This is one of the places where actual expertise differs from the popular im..."
2. **true axis: notice what experts ignore** — The most useful advantage of youth
3. **per-project procrastination most dangerous** — The deadliest procrastination
4. **mathematical elegance as standard** — Mathematical elegance may sound like it's more
5. **originality is a habit of mind** — "creative process" is overly mystifying
6. **curiosity is a kind of originality** — Curiosity and originality are closely related

## Demo Readiness Assessment

**Ready for demo**: Yes, with caveats
- Layout works well — claims highlighted in text, margin annotations aligned
- Claim quality filter successfully separates genuine assertions from frameworks
- Top claims summary is effective for quick overview
- Collision avoidance keeps margin notes from overlapping

**Known limitations**:
- Title defaults to filename stem for local files (pass better title via text content)
- Very dense claim sections can push margin notes far from their source text
- Full 67k essay takes ~5 min for extraction + rating (two LLM passes)
- `_load_source` has a bug for URL inputs (FetchResult unpacking) — works fine for local files