71  Bench v4 Hardness Ladder

72 Bench v4 Hardness Ladder

Bench v4 begins where the v3 field-spell corpus stopped being discriminating enough. The seed ladder keeps the evidence narrow: local deterministic artifacts are executed repeatedly in repo-local tmp/ sandboxes so the fixture contract can be inspected before any model-surface claims are made.

72.1 Policy

Bench v4 hardness fixtures run only local deterministic tests inside repo-local tmp/ sandboxes. Model-surface runs are not simulated.

72.2 Seed Rungs

Case Rung Hardness Axis Weak Repaired n Fixture
Disabled Account Status Ambiguity ambiguity conflicting docstring versus caller contract 0/5 passed 5/5 passed 10 fixture
Event Deduplication Hidden Order Invariant hidden-invariant order and replay semantics hidden behind simple dedupe wording 0/5 passed 5/5 passed 10 fixture
Tax Discount Misleading Context misleading-context stale inline comment contradicts current merchant contract 0/5 passed 5/5 passed 10 fixture
Beta Flag Blast Radius blast-radius narrow feature-flag change versus tempting billing overreach 0/5 passed 5/5 passed 10 fixture
Agentic Release Handoff agentic multi-step stack handoff with repo-local scratch and gate evidence 0/5 passed 5/5 passed 10 fixture

72.3 Model-Surface Artifact Runs

These rows are model-produced artifacts, not hand-authored controls. Prompts include the public task context and starting files, but not the hidden grader or ground-truth JSON. The runner preserves each prompt, transcript, extracted artifact, extraction failure, and fixture-local execution result.

Case Surface Weak Repaired Extraction Failures n Example Transcript
Disabled Account Status Ambiguity codex-cli-default 0/5 passed 0/5 passed 0 10 transcript
Event Deduplication Hidden Order Invariant codex-cli-default 0/5 passed 0/5 passed 0 10 transcript
Tax Discount Misleading Context codex-cli-default 0/5 passed 0/5 passed 0 10 transcript
Beta Flag Blast Radius codex-cli-default 0/5 passed 0/5 passed 5 10 transcript
Agentic Release Handoff codex-cli-default 1/5 passed 1/5 passed 0 10 transcript

72.4 Reviewer Import Contract

Additional non-Codex surfaces can enter through a manual import bundle. The validator checks schema fields, repo-local paths, provenance/evidence-class consistency, allowed artifact filenames, prompt exclusion of hidden grader files, and fixture-local execution against the private Bench v4 grader. It does not publish or accept the run; maintainer review remains a separate decision.

Template: manual-import-template.json

Validate a bundle:

python3 scripts/import_hardness_model_run.py validate examples/evaluations/hardness-v4/manual-import-template.json
python3 scripts/grimoire.py bench hardness-import examples/evaluations/hardness-v4/manual-import-template.json

72.5 Reviewer Acceptance Decision

Validated import bundles remain intake records until a named maintainer accepts and publishes a decision. The pending decision template does not count as cross-surface evidence.

Decision template: hardness-intake-decision-template.json

Validate a decision:

python3 scripts/check_hardness_intake.py validate examples/evaluations/hardness-v4/hardness-intake-decision-template.json
python3 scripts/grimoire.py bench hardness-decision examples/evaluations/hardness-v4/hardness-intake-decision-template.json

72.6 Interpretation

  • The ambiguity rung tests whether an artifact resolves a stale docstring versus a stronger caller contract.
  • The hidden-invariant rung tests whether an artifact preserves replay order and mutation boundaries that are easy to erase with a naive cleanup.
  • The misleading-context rung tests whether an artifact ignores stale or wrong local guidance in favor of the executable contract.
  • The blast-radius rung tests whether an artifact keeps a requested change narrow instead of altering adjacent behavior.
  • The agentic rung tests whether a stack-shaped handoff preserves repo-local scratch, allowlists, gates, and human-review boundaries.
  • These are execution checks for seed artifacts. They do not count as independent model-provider evidence, external adoption, or human review.
  • Model-surface rows count only when prompts, transcripts, extracted artifacts, and execution results are preserved. They still do not count as external adoption or human review.

Raw results: results.json

Raw model-surface results: model-surface-results.json