Publish five-benchmark routing tables

[SivaKesava1]

Summary

• publish six routing tables from the 2026-10-03-five-benchmark training release
• incorporate accepted Aider Polyglot, SWE-bench Verified, SWE-bench Pro, SWT-Bench Verified, and Terminal-Bench 2 evidence
• retain schema 5 and all six cost/cost-speed profiles

Training provenance

• 79,227 observations
• 21 measured label cells; 210 published cells
• capability slope: 0.07691434935428793
• Terminal-Bench 2 contributes one qualifying analyzed cell from 10 tasks in its 73-task fitted common cohort

Replay versus previous tables

Across 1,844 evaluated tasks:

• economy: +2.33 passes (+0.127 pp), lower credits and runtime
• balanced: -2.67 passes (-0.145 pp), nearly unchanged credits, faster runtime
• robust: -0.33 passes (-0.018 pp), lower credits and runtime

Validation

• final table hashes are byte-identical to the reviewed candidate
• publication validated all six tables
• python -m pytest tests/test_routing.py tests/test_package.py: 49 passed, 3 deselected