Method

What these numbers are

Effort

Changed lines are additions plus deletions on authored sample pull requests. A theme share is that theme's lines divided by 1,000. It is not time, difficulty, or productivity. The sample groups were written by hand.

Signals

Demand counts are request records. Engineering counts are merged pull requests with the same primary label. A gap flag means at least 5 demand records and related PRs under 20 percent of that demand count. Six requests and one PR is a flag. It is not proof the work was ignored or finished.

Budget

Calls equal weeks times the sum of people times sessions per person per week. Cost uses uncached standard input and output rates from the 29 September 2026 official-source seed. Cache, batch, and regional discounts are not applied. A missing applicable price makes the total partial. Jev always prices classification when its row applies. Writer choice is the lowest estimated session cost among reviewed candidates for that tier, with a stable id tie-break. That is not a quality score.

How an answer is assembled

1. Source records
2. Bounded candidate selection in code
3. Jev judgments when a private run is configured
4. Code aggregation
5. Cited answer

The sample desks do not call Jev. A private run would still keep aggregation in code.

What is not here

No Artificial Analysis scores, speed tables, or raw catalog. No Open Art rankings. No live GitHub import until a named repository and read-only token are configured. No model is called to render this sample.

Source notes live in research/2026-09-29-source-verification.md.