CtrlK
BlogDocsLog inGet started
Tessl Logo

experiments/eval-improve

Analyze eval results, diagnose low-scoring criteria, fix tile content, and re-run evals — the full improvement loop automated

Does it follow best practices?

Evaluation100%

1.02x

Agent success when using this tile

Validation for skill structure

Overview
Skills
Evals
Files

rubric.jsonevals/scenario-3/

{
  "context": "Tests whether the agent scans all tile files for contradictions and reports every conflict found — including those unrelated to the retry issue that triggered the audit — attributing each contradiction to the specific files involved with direct quotes.",
  "type": "weighted_checklist",
  "checklist": [
    {
      "name": "Retry count contradiction found",
      "description": "Report identifies the conflict between SKILL.md ('retry up to 5 times') and rules/retry-policy.md ('maximum retries: 3 attempts')",
      "max_score": 20
    },
    {
      "name": "Auth failure contradiction found",
      "description": "Report identifies the conflict between SKILL.md ('abort the pipeline run immediately' on auth failure) and docs/error-handling.md ('skip the affected batch and log the error — do not abort the full pipeline run')",
      "max_score": 20
    },
    {
      "name": "All three files referenced",
      "description": "The report references content from all three provided files: SKILL.md, rules/retry-policy.md, and docs/error-handling.md",
      "max_score": 15
    },
    {
      "name": "File attribution per contradiction",
      "description": "Each contradiction explicitly names the specific files that conflict (e.g., 'SKILL.md vs rules/retry-policy.md'), not just a general mention",
      "max_score": 15
    },
    {
      "name": "Auth contradiction despite scope",
      "description": "The authentication failure contradiction is included in the report even though the task mentions retry logic as the trigger — demonstrates scanning all tile content, not just retry-related sections",
      "max_score": 20
    },
    {
      "name": "Verbatim quotes included",
      "description": "Each contradiction includes direct quotes or verbatim excerpts from the conflicting statements, not just a paraphrase",
      "max_score": 10
    }
  ]
}

Install with Tessl CLI

npx tessl i experiments/eval-improve@0.4.0

evals

README.md

tile.json