Skip to content

capsid-bench

The study harness that compares Claude Code with and without Capsid on the same tasks, scored on hidden tests.

github.com/DrDustinEdwards/capsid-bench · capsid.dustinedwards.info/s/capsid-bench

How capsid-bench connects

Used by 1

Inside capsid-bench

Runs

  • bench-run workflow

    GitHub Actions

    bench-run workflow Runs one task in one arm, behind an egress allowlist, then scores and seals it. writes to MCP Worker (Capsid) reads Claude Code writes to drand and Sigstore

  • bench-walk workflow

    GitHub Actions

    bench-walk workflow Draws the accepted task set from the frozen frame with a drand seed. reads drand and Sigstore

Outside

  • Claude Code

    Claude Code on a subscription token

    Claude Code The agent under study, in both arms.

  • drand and Sigstore

    drand quicknet, Sigstore and Rekor

    drand and Sigstore The public randomness for the draw, and the time-locked, logged seal.

Uses 1

Public model, 2026-10-23. Markdown