Compact conversation context by calibrated judgment

01 · Role in the system

What Jev does here

On the LoCoMo benchmark - ten very long two-speaker conversations with question-answer annotations - every conversation chunk is scored by Jev for relevance to the question being asked, and deterministic Python selects the chunks that fit a token budget. The HTTP adapter posts the native TypeSafe request shape (state plus typed questions) through OpenRouter's decisions endpoint, records actual cost from the response's usage field rather than estimating, and treats 429s as normal operating condition with bounded backoff - with concurrency measured live and capped at sixteen. The repository publishes the full comparison matrix, figures over 486 questions, and a probe showing that answering from minimal evidence scores below the best method, so the analysis can be replayed offline without a key.

02 · Control boundary

Where Jev sits

Per-chunk relevance judgments over a conversation; deterministic budget-aware selection in Python; benchmark harness with replayable artifacts and measured cost accounting.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• A benchmark analysis rather than a shipped tool: the deliverable is the comparison matrix and findings, run through OpenRouter rather than the native endpoint.
• Measured on LoCoMo's ten two-speaker conversations; other conversation shapes are untested here.
• Benchmark figures are the author's own committed runs.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• pdrpinto ↗Community · github · public · checked 2026-09-24