Skip to content

Report shared-start and seed-spread context in validate_implementations.py #359

Description

@neuromechanist

Found in epic #324 Phase 15 (#351).

The harness fits pamica and the reference from independent random starts at one seed. So its log-likelihood gap is one draw from the start-to-start spread:

  • the reference's own final LL has a standard deviation of 2.6e-4 over 8 seeds at the harness settings;
  • the harness gap moved from 2.9e-5 to 2.7e-4 when the epic changed the initial draw.

The update-rule comparison is the shared-start one: the reference loaded with pamica's initial state gives an LL gap of 1.6e-6 and a correlation of 0.99999993.

Proposal: have validate_implementations.py also report a shared-start comparison through native_oracle.run_seeded_reference, the reference's own seed spread, or both, next to its independent-start figure, so the output carries its own context.

Scripts and records: .context/issue-351/harness_gap.py, raw/harness_gap_head.json.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions