Evals Implement

evals-implement

What this skill does

Generates the complete executable evaluation implementation following EDD Principle VIII (Close Production Loop) from the published goldset, with automated unit testing to verify evaluator correctness.

Output:

1. Grader/Metric Implementation - Python evaluators for each goldset criterion with binary pass/fail
   • PromptFoo: Python grader functions with JSON output in evals/{system}/graders/
   • DeepEval: Custom metric classes inheriting from BaseMetric
2. Evaluator Unit Tests - Automated tests (evals/{system}/tests/test_check_*.py) that run the goldset pass/fail examples against the generated graders to ensure the evaluator itself is accurate
3. Evaluation Configuration - Complete config file (config.js or config.py) with Tier 1 + Tier 2 evaluation structure
4. Auto-handoff to /evals-validate to run validation

Key EDD Principles Applied:

• Principle VIII: Close Production Loop - Failure type gates route to appropriate actions
• Principle II: Binary Pass/Fail - Ensure graders return strictly 1.0 (pass) or 0.0 (fail)
• Principle IX: Test Data as Code - Unit test generated code against dataset examples

When to use

• After /evals-clarify: Convert accepted goldset criteria into executable code
• Regenerating configs: Re-build evaluator suite after adding new goldset criteria
• Adding unit tests: Hardening the evaluator itself against regression or bugs

When NOT to use

• Goldset not published: Run /evals-clarify to generate goldset.json first
• Running evaluations: Use /evals-validate to run the suite against application outputs

Process

User Input

textCopy code
$ARGUMENTS

• --system SYSTEM — Override active evaluation framework (promptfoo or deepeval)
• --no-tests — Skip automated unit test generation for graders (not recommended)

Execution Steps

Phase 1: Trace-to-Grader Synthesis (Automated Eval Engineering)

• Reads evals/{system}/goldset.json.
• Maps rich evidence fields from the goldset criteria into grader logic (Trace-to-Grader Synthesis):
  • Uses pass_condition and fail_condition as the grader's core rubric.
  • Extracts pass/fail examples to act as raw data anchors and few-shot classification anchors inside the grader logic.
  • Injects Root Cause Analysis and axial_coding notes as contextual prompt guidelines or regex patterns to catch exact failure manifestations.
• For PromptFoo: Generates Python grader functions (evals/{system}/graders/check_*.py) containing specialized, dynamic LLM-judge templates or regex checks compiled from these goldset inputs.
• For DeepEval: Generates Custom Metric classes inheriting from BaseMetric compiled from these goldset inputs.
• All graders conform strictly to the binary pass/fail standard (returning only 1.0 or 0.0, with zero Likert scale leakage).

Phase 2: Unit Test Generation

• Generates matching unit tests (evals/{system}/tests/test_check_*.py) for each grader.
• Unit tests verify the grader correctly identifies the goldset's training pass and fail examples.

Phase 2b: Closed-Loop Grader Self-Tuning

• Executes generated unit tests (pytest evals/{system}/tests/) to verify evaluator accuracy.
• Grader Calibration Loop:
  1. Inspects test results to detect any misclassifications (false positives/negatives) on the training cases.
  2. If any test fails, triggers a feedback edit step that parses the failure reasons and automatically adjusts the grader's internal prompt rubric, regex stubs, or score thresholds.
  3. Re-runs pytest to check accuracy.
  4. Repeats for up to 3 iterations (the hard circuit-breaker limit).
• Holdout Locking: Ensure the holdout validation set (holdout.json) remains completely isolated and is never loaded or exposed to the self-tuning loop (to prevent overfitting).
• Failure Escalation: If the grader does not converge to 100% training accuracy within 3 iterations, the loop halts, surfaces the failing test case details, and raises an error rather than passing silently.

Phase 3: Config Generation

• Generates the unified framework configuration file (config.js or config.py).
• Configures separate Tier 1 (fast checks, <30s, deterministic) and Tier 2 (semantic checks, <5min, LLM-judge) pipelines.

Phase 4: Auto-Handoff

Trigger /evals-validate to run validation.

Verification

• evals/{system}/graders/ contains Python grader scripts for each criterion compiled dynamically from goldset pass/fail examples and root-cause analyses
• evals/{system}/tests/ contains matching unit test files
• Framework config (config.js or config.py) successfully generated
• Grader calibration self-tuning loop ran and converged to 100% training accuracy within the 3-iteration cap (or raised explicit non-convergence errors)
• Holdout dataset protection confirmed (validation holdout.json remained completely isolated and untouched during tuning)
• All grader unit tests pass locally (pytest evals/{system}/tests/)
• Handover summary lists generated graders, self-tuning iterations, and test results