Skip to content

Validate real legal corpus admission and benchmark against an independently reviewed oracle #5

Description

@tihotm

Context

The deterministic benchmark harness, external catalog checksum validation, failure paths and self-test model are already implemented and locally exercised. What remains unproven is the factual happy path against an authorized legal corpus and independently reviewed oracle.

Objective

Execute corpus admission and benchmark validation only with authorized legal source material and an independently reviewed oracle. Generated content must not be treated as factual evidence, source of truth or gold truth.

Acceptance criteria

  • Use only an authorized legal corpus.
  • Record provenance and checksums for all admitted artifacts.
  • Use an independently reviewed oracle/gold reference.
  • Run corpus admission and benchmark happy paths with the real artifacts.
  • Record deterministic metadata: catalog version, model id/config, seed, runtime and git commit.
  • Preserve non-compensable correctness/safety gates.
  • Do not weaken thresholds to obtain a pass.
  • Do not fabricate legal facts or synthetic oracle answers to close this issue.

Unblock condition

Authorized corpus plus independently reviewed oracle become available.

Current blocker

This formalizes RJ-BLK-003.

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

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions