ACCEPT-003 Human Review Blind Worksheet - Instructions
========================================================

1. Review Purpose
-----------------
Validate the correlation between RSI (Recursive Self-Improvement) automated
scoring and human expert judgment. This is the L2 acceptance criteria
PERCV-RSI-ACCEPT-003.

The automated scoring system must produce scores that correlate with human
expert assessment to ensure the self-improvement loop is trustworthy.

2. Scoring Dimensions (6 dimensions, 0-100 points each)
--------------------------------------------------------
Each round must be scored on the following six dimensions, using an integer
or float in the range [0, 100]:

  - correctness       : Correctness of the code improvements introduced in
                        this round (logic correctness, edge cases, bug-free).
  - testing           : Quality and coverage of tests added or modified in
                        this round.
  - code_quality      : Readability, maintainability, and adherence to style
                        conventions of the changed code.
  - security          : Safety of the changes (no injection, no privilege
                        escalation, no secret leakage).
  - efficiency        : Resource usage (CPU, memory, latency) of the changed
                        code relative to prior state.
  - adoption          : Success rate of integration (did the improvement
                        land and stick without rollback).

3. Review Requirements - Independent Blind Review
-------------------------------------------------
  - Reviewers MUST score independently. No discussion with other reviewers
    until all scores are submitted.
  - Reviewers MUST NOT view the automated_scores or reviewer_scores fields
    in vault/rsi-correlation-data.yaml before completing this worksheet.
  - Reviewers use ONLY the context_data columns (round_id, snapshot_index,
    experiment_count, avg_score, safety_alerts, human_interventions) plus
    the actual round artifacts (git log, diff, test output) as the basis
    for scoring.
  - Minimum 2 reviewers, target 3. Each reviewer fills one copy of the
    worksheet.

4. Pass Criteria
----------------
  - Primary metric : Spearman rank correlation between automated score and
                      mean human score, computed per dimension and overall.
  - Threshold      : Spearman >= 0.7 on all 6 dimensions AND overall.
  - PASS              : Spearman >= 0.7 on all 6 dimensions + overall.
  - CONDITIONAL PASS  : 5/6 dimensions >= 0.7, overall >= 0.7.
  - FAIL              : Any dimension < 0.5, or overall < 0.7.
  - Secondary metric : Cohen's kappa for inter-rater reliability >= 0.6.

5. How to Fill In
-----------------
  - Open docs/rsi/accept-003-review-worksheet.csv in a spreadsheet editor
    or text editor.
  - The first 6 columns are pre-filled with context data for reference:
      round_id, snapshot_index, experiment_count, avg_score,
      safety_alerts, human_interventions
  - Fill in the last 6 columns with your scores (0-100):
      reviewer_correctness, reviewer_testing, reviewer_code_quality,
      reviewer_security, reviewer_efficiency, reviewer_adoption
  - Do NOT modify the first 6 columns.
  - Do NOT look at automated_scores in the source YAML until done.
  - Save the completed file as
      docs/rsi/accept-003-review-worksheet-<your-name>.csv
  - Submit to the Orchestrator for correlation analysis.

6. Data Source
--------------
  - Source YAML : vault/rsi-correlation-data.yaml (42 rounds)
  - Worksheet   : docs/rsi/accept-003-review-worksheet.csv (42 rows + header)
  - Protocol    : docs/rsi/l2-human-correlation-protocol.md
