evaluation-rubrics（评价标准）

Evaluation Rubrics

Table of Contents

• Workflow
• Common Patterns
• Guardrails
• Quick Reference

Example

Scenario: Evaluating technical blog posts (1-5 scale)

Scoring: Post A [4, 5, 3, 2] = 3.5 avg. Post B [5, 4, 5, 4] = 4.5 avg. Feedback for Post A: "Strong clarity (5) and good accuracy (4), but needs more practical examples (3) and offers less original insight (2)."

Workflow

Copy this checklist and track your progress:

Rubric Development Progress:
- [ ] Step 1: Define purpose and scope
- [ ] Step 2: Identify evaluation criteria
- [ ] Step 3: Design the scale
- [ ] Step 4: Write performance descriptors
- [ ] Step 5: Test and calibrate
- [ ] Step 6: Use and iterate

Step 1: Define purpose and scope

Clarify what you're evaluating, who evaluates, who uses results, what decisions depend on scores. See resources/template.md for scoping questions.

Step 2: Identify evaluation criteria

Brainstorm quality dimensions, prioritize most important/observable, balance coverage vs. simplicity (4-8 criteria typical). See resources/template.md for brainstorming framework.

Step 3: Design the scale

Choose number of levels (1-5, 1-4, 1-10), scale type (numeric, qualitative), anchors (what does each level mean?). See resources/methodology.md for scale selection guidance.

Step 4: Write performance descriptors

For each criterion × level, write observable description of what that performance looks like. See resources/template.md for writing guidelines.

Step 5: Test and calibrate

Have multiple reviewers score sample work, compare scores, discuss discrepancies, refine rubric. See resources/methodology.md for inter-rater reliability testing.

Step 6: Use and iterate

Apply rubric, collect feedback from evaluators and evaluatees, revise criteria/descriptors as needed. Validate using resources/evaluators/rubric_evaluation_rubrics.json. Minimum standard: Average score ≥ 3.5.

Common Patterns

Pattern 1: Analytic Rubric (Most Common)

• Structure: Multiple criteria (rows), multiple levels (columns), descriptor for each cell
• Use case: Detailed feedback needed, want to see performance across dimensions, diagnostic assessment
• Pros: Specific feedback, identifies strengths/weaknesses by criterion, high reliability
• Cons: Time-consuming to create and use, can feel reductive
• Example: Code review rubric (Correctness, Efficiency, Readability, Maintainability × 1-5 scale)

Pattern 2: Holistic Rubric

• Structure: Single overall score, descriptors integrate multiple criteria
• Use case: Quick overall judgment, summative assessment, criteria hard to separate
• Pros: Fast, intuitive, captures gestalt quality
• Cons: Less actionable feedback, lower reliability, can't diagnose specific weaknesses
• Example: Essay holistic scoring (1=poor essay, 3=adequate essay, 5=excellent essay with detailed descriptors)

Pattern 3: Single-Point Rubric

• Structure: Criteria listed with only "meets standard" descriptor, space to note above/below
• Use case: Growth mindset feedback, encourage self-assessment, less punitive feel
• Pros: Emphasizes improvement not deficit, simpler to create, encourages dialogue
• Cons: Less precision, requires written feedback to supplement
• Example: Design critique (list criteria like "Visual hierarchy", "Accessibility", note "+Clear focal point, -Poor contrast")

Pattern 4: Checklist (Binary)

• Structure: List of yes/no items, must-haves for acceptance
• Use case: Compliance checks, minimum quality gates, pass/fail decisions
• Pros: Very clear, objective, easy to use
• Cons: No gradations, misses quality beyond basics, can feel rigid
• Example: Pull request checklist (Tests pass? Code linted? Documentation updated? Security review?)

Pattern 5: Standards-Based Rubric

• Structure: Criteria tied to learning objectives/competencies, levels = degree of mastery
• Use case: Educational assessment, skill certification, training evaluation, criterion-referenced
• Pros: Aligned to standards, shows progress toward mastery, diagnostic
• Cons: Requires clear standards, can be complex to design
• Example: Data science skills (Proficiency in: Data cleaning, Modeling, Visualization, Communication × Novice/Competent/Expert)

Guardrails

1. Criteria should be observable and measurable: Not "good attitude" (subjective), but "arrives on time, volunteers for tasks, helps teammates" (observable). Test: Can two independent reviewers score this criterion consistently?
2. Descriptors should distinguish levels clearly: Each level needs concrete differences from adjacent levels. Avoid "5=very good, 4=good, 3=okay". Better: "5=zero bugs, meets all requirements, 4=1-2 minor bugs, meets 90% requirements."
3. Use appropriate scale granularity: 1-3 is too coarse, 1-10 is too fine. Sweet spot: 1-4 (forced choice, no middle) or 1-5 (allows neutral middle). Match granularity to actual observable differences.
4. Balance comprehensiveness with simplicity: Aim for 4-8 criteria covering essential quality dimensions. If >10 criteria, consider grouping or prioritizing.
5. Calibrate for inter-rater reliability: Have multiple reviewers score same work, measure agreement (Kappa, ICC). If <70% agreement, refine descriptors.
6. Provide examples at each level: Include concrete examples of work at each level (anchor papers, reference designs, code samples) to calibrate reviewers.
7. Share rubric before evaluation: If evaluatees see the rubric only after being scored, it is grading not guidance. Share upfront so people know expectations and can self-assess.
8. Weight criteria appropriately: If "Security" matters more than "Code style", weight it (Security x3, Style x1). Or use thresholds (score >=4 on Security to pass, regardless of other scores).

Common pitfalls:

• ❌ Subjective language: "Shows effort", "creative", "professional" - not observable without concrete descriptors
• ❌ Overlapping criteria: "Clarity" and "Organization" often conflated - define boundaries clearly
• ❌ Hidden expectations: Rubric doesn't mention X, but evaluators penalize for missing X - document all criteria
• ❌ Central tendency bias: Reviewers avoid extremes (always score 3/5) - use even-number scales (1-4) to force choice
• ❌ Halo effect: High score on one criterion biases other scores up - score each criterion independently before looking at others
• ❌ Rubric drift: Descriptors erode over time, reviewers interpret differently - periodic re-calibration required

Quick Reference

Key resources:

• resources/template.md: Purpose definition, criteria brainstorming, scale selection, descriptor templates, rubric formats
• resources/methodology.md: Scale design principles, descriptor writing techniques, inter-rater reliability testing, bias mitigation
• resources/evaluators/rubric_evaluation_rubrics.json: Quality criteria for rubric design (criteria clarity, scale appropriateness, descriptor specificity)

Scale Selection Guide:

Criteria Types:

• Product criteria: Evaluate the artifact itself (correctness, clarity, completeness, aesthetics, performance)
• Process criteria: How work was done (methodology followed, collaboration, iteration, time management)
• Impact criteria: Outcomes/effects (user satisfaction, business value, learning achieved)
• Meta criteria: Quality of quality (documentation, testability, maintainability, scalability)

Inter-Rater Reliability Benchmarks:

• <50% agreement: Rubric unreliable, needs major revision
• 50-70% agreement: Marginal, refine descriptors and calibrate reviewers
• 70-85% agreement: Good, acceptable for most uses
• >85% agreement: Excellent, highly reliable scoring

Typical Rubric Development Time:

• Simple rubric (3-5 criteria, 1-4 scale, known domain): 2-4 hours
• Standard rubric (5-7 criteria, 1-5 scale, some complexity): 6-10 hours + calibration session
• Complex rubric (8+ criteria, multiple scales, novel domain): 15-25 hours + multiple calibration rounds

When to escalate beyond rubrics:

• High-stakes decisions (hiring, admissions, awards) → Add structured interviews, portfolios, multi-method assessment
• Subjective/creative work (art, poetry, design) → Supplement rubric with critique, discourse, expert judgment
• Complex holistic judgment (leadership, cultural fit) → Rubrics help but don't capture everything, use thoughtfully → Rubrics are tools not replacements for human judgment. Use to structure thinking, not mechanize decisions.

Inputs required:

• Artifact type (what are we evaluating? essays, code, designs, proposals?)
• Criteria (quality dimensions to assess, 4-8 most common)
• Scale (1-5 default, or specify 1-4, 1-10, qualitative labels)

Outputs produced:

• evaluation-rubrics.md: Purpose, criteria definitions, scale with descriptors, usage instructions, weighting/thresholds, calibration notes