Skip to content
LogoGeekjourney

2026-W42 —

As of October 12: revisiting JEV, Dream RSI, and harness engineering, and recalibrating the past month's project status.

Conversations: Sources degraded

This week’s focus

Only verifiable public evidence is recorded. An absence of records does not mean an absence of effort.

Exploring

Revisited

Building

Output

Sharing

No public evidence

Living

No public evidence

Exploring

Revisited

  • Recurring question

    JEV: compare accuracy, calibration, latency, and cost on the same tasks.

    Currently focused on small decisions in classification and routing, including the division of work with general LLMs, error boundaries, and practical integration value.

    JEV · Evaluation · Harness

  • Recurring question

    Dream RSI / RRSI: ground policy iteration in real feedback.

    Continuing research on replay, policy improvement, and generalization, with video-rsi as a bounded experiment. Multiple generation rounds and model judging need calibration through real use.

    Dream RSI · RRSI · Pi

  • Insight & judgment

    Evaluate agents by whether complete tasks finish well at a reasonable cost.

    Consider task success, cost, caching, steps, recovery on long-running work, and verification together, while studying improvements across models and harnesses.

    Harness · Cost per Task · Verifier

    • Research focus · 2026-10-12

Building

Output

Last updated: 2026-10-12