jev-linkmap · www.bles-software.com
The internal link map, rebuilt in 6 seconds
Jev read 566 pages and answered 8,460 link questions for $0.273. Claude Opus 5, same queue, same rubric, same clock, had finished 0 pages when Jev was done. Priced on 24 random pages, Claude Opus 5 costs $0.118 a page against $0.0005: 244 times more, about $66.58 for the full pass.
The race
Both judges pull from one queue, read the same page copy, the same 15 candidate targets and the same anchor phrases, and answer by the same rubric. The clock stops the moment Jev finishes, and only pages the frontier model completed by then count. Calls already in flight are allowed to land so the chart shows when they did; no new page starts after the clock.
Jev ran 32 requests at a time, Claude Opus 5 ran 8 at a time.
Jev
- Pages finished by the clock
- 566
- Pages finished in all
- 566
- Link decisions
- 8,460
- Cost
- $0.273
- Cost per page
- $0.0005
- Model time per page
- 0.32s
- Links placed
- 679
Claude Opus 5
- Pages finished by the clock
- 0
- Pages finished in all
- 8
- Link decisions
- 120
- Cost
- $0.215
- Cost per page
- $0.027
- Model time per page
- 9.91s
- Links placed
- 11
The 8 pages Claude Opus 5 landed are the short pages at the top of the queue, so its per-page cost here runs low. The price used above comes from 24 random pages judged with this same rubric: $0.118 and 9.8s of model time a page.
Two ways to run it
Jev is System 1: fast, cheap and literal, so the question it is asked is the product. You can run it alone with a rubric that is already trained, or put a deep model behind it as System 2 to train the rubric on your own site first.
System 1Jev alone
- Models you need
- Jev
- This site
- 6.0s, $0.273
- Links placed
- 679
- Rubric
- v3, already trained
One key, no other account. Use it when your site looks like the one the rubric was trained on, or when seconds and cents matter more than the last few links.
python3 -m jev_linkmap.site https://your-site.com/sitemap.xml --mode system1System 1 + System 2Jev + a deep model
- Models you need
- Jev + Claude or Codex
- Referee links Jev finds
- 45% to 65%
- Same anchor as referee
- 71% to 88%
- One-time training cost
- $15.51
A deep model referees small blocks of pages, reads where it and Jev disagreed, and rewrites Jev's rubric. Then Jev maps the whole site alone with the rubric that scored best. You pay for the deep model once, on a few dozen pages, never on the whole site.
python3 -m jev_linkmap.site https://your-site.com/sitemap.xml --mode system1+system2
# System 2 on your ChatGPT plan, through the Codex CLI
LINKMAP_DEEP=codex python3 -m jev_linkmap.site https://your-site.com/sitemap.xml --mode system1+system2What System 2 adds
After each block of pages System 2 rewrites the rubric: the question wording, up to six one-line lessons, the two thresholds. Every rubric version is then scored on one held-out block that no version was written from. Jev never sees the deep model, only the new rubric.
| Rubric | Decisions | Same call as referee | Jev links the referee confirms | Referee links Jev found | Same anchor | Jev placed | Referee placed |
|---|---|---|---|---|---|---|---|
| v1 (System 1, written by hand) | 360 | 81% | 83% | 45% | 71% | 58 | 106 |
| v2 (System 1 + System 2) | 360 | 82% | 84% | 55% | 70% | 75 | 115 |
| v3 (System 1 + System 2) | 360 | 84% | 83% | 65% | 88% | 92 | 117 |
The link map
Jev placed 679 links. Before anything went on the live site an editor pass (claude-opus-5, $2.07) read only those links, twenty to a call, and kept the 287 below on 216 pages. It can cut a link, it can never add one. The highlighted words are already in the page, so placing a link means wrapping them, not rewriting anything.
| Anchor in its sentence | Links to | Confidence |
|---|