Skip to content

Make Codex and Claude Code subscriptions go further: a local model router that gives routine agent work to economical models and keeps your strongest model for the hard parts.

2.6 KB
Raw Blame

Execution plans and exact request bodies

studies.json indexes all nine campaigns, their planned counts, response records, dataset association, expected model identities and archive paths. backends.json retains the evaluated Jev/Qwen versions and relevant transport conditions without naming the hosted decision access service.

Each campaign directory contains study.json, jobs.jsonl.gz and requests.jsonl.gz. Plans preserve the original request order, identifiers, repeats and exact semantic request bodies. Some native templates are shared by both model versions; their jobs and backend table supply the model identity. Other archives store complete per-job bodies. Neither format exposes labels to the model. Gzip reduces repeated prompt text; no custom archive format is needed.

Campaign metadata Exact requests Job plan
Engineering matched Requests Jobs
Independent engineering Requests Jobs
Mathematics matched Requests Jobs
Direct expression Requests Jobs
Direct probes Requests Jobs
Development V3 Requests Jobs
Abandoned engineering Requests Jobs
Completed format development Requests Jobs
Stopped format development Requests Jobs

The two format-development ledgers recorded results without separate start events. Their metadata keeps unobserved plan entries distinct from observed starts. Abandoned, diagnostic and completed campaigns are never pooled into one accuracy. The reproduction guide describes offline parsing and evaluation; response archives retain the recorded outputs.