Skip to content

Latest commit

 

History

History
49 lines (37 loc) · 1.92 KB

File metadata and controls

49 lines (37 loc) · 1.92 KB

Routing and cache live calibration

Field Value
Evidence level Live calibration
Date 30 July 2026
Repository / commit ripgrep 14.1.1 / 4649aa9700619f94cf9c66876e9549d83420e16c
Task Locate the implementation of --glob-case-insensitive
Route locate.implementation
Models Main gpt-5.6-sol; worker gpt-5.6-luna; medium reasoning
Codex / tier 0.144.0 / default
Pricing digest b3:eca7f25b4a34dcf0f177601be1675bb97c16f24347d998dc9a2899e10b73cee9
Provider calls Three main observations; one logical worker in the miss arm
Automatic retries None

Protocol

The frozen natural task and quality oracle were identical across frontier direct, routed miss, and proof-backed hit arms. The miss could use at most one worker; the hit had to remain cache-only.

Result

All final responses completed and passed the quality oracle.

Arm Main cost Worker cost Total cost Workers Main discovery
MainOnly 3.267775 0 3.267775 credits 0 3
NeedleMiss 0.349300 0.486690 0.835990 credits 1 0
NeedleHit 0.357425 0 0.357425 credits 0 0

The miss published one validated implementation artifact. The hit reused the same artifact through an authoritative CoverageHit, with validation and sufficiency certificates, positive observed reuse economics, zero stale or contradicted evidence, and worker_avoided=true.

Observed reductions for this sample:

  • routing: 74.41%;
  • cache relative to miss: 57.24%;
  • end-to-end relative to direct: 89.06%.

The source remained clean. Focused-test execution was not required by the task.

Limits and non-claims

This validates the current live route → worker → validation → publication → zero-worker reuse path for one pinned location task. It does not establish a confidence interval, multi-task behavior, other routes, or general savings.