GRADING-RUBRIC.md

HaleES Public Grading Rubric (Dual-Layer)

This document defines the public dual-layer grading model used in the HaleES Architecture Specification.

The rubric combines:

1. Gradient scoring (0–100) for evaluative detail
2. Binary decision (0|1) for enforceable pass/fail gating

---

1) Categories

Each output is graded across five required categories:

1. accuracy
   • Correctness of content versus contract requirements and inputs.
2. efficiency
   • Resource-consciousness and avoidance of unnecessary complexity.
3. constraint_adherence
   • Compliance with explicit contract constraints and boundaries.
4. quality
   • Clarity, structure, usability, and professional completeness.
5. timeliness
   • Alignment to required response window or operational tempo.

All five categories must be present for a valid grading result.

---

2) Scoring Range

• Each category score: 0 to 100
• global_score: 0 to 100
• confidence: 0.0 to 1.0 (tracked independently)

Interpretation guideline:

• 90–100: excellent
• 85–89: acceptable pass band
• 70–84: near-pass but requires revision
• below 70: insufficient

---

3) Global Score Calculation

Default public calculation uses equal weighting across the five categories:

[ \text{global_score} = \text{round}\left(\frac{accuracy + efficiency + constraint_adherence + quality + timeliness}{5}\right) ]

Rounding method should be consistent per implementation (recommended: nearest integer).

---

4) Threshold and Binary Decision

Decision threshold:

• binary_decision = 1 if global_score >= 85
• binary_decision = 0 if global_score < 85

This threshold is the normative public baseline.

---

5) Confidence

confidence is recorded separately from pass/fail.

• Confidence does not override the threshold rule.
• A high confidence score cannot convert a failing global score into pass.
• A low confidence score does not automatically fail a passing output, but may trigger human review depending on policy.

---

6) Pass/Fail Logic

The grading model is intentionally dual-layer:

• 0–100 evaluates the degree of quality and compliance.
• 0|1 decides acceptance status.

Operational rule:

• No decision exists without scoring.
• No scoring matters without a decision.

This creates both explainability (why) and enforceability (what outcome).

---

7) Iteration Rules

When binary_decision = 0, system behavior should be:

1. Append specific feedback by category.
2. Re-run execution against the same contract (or a versioned amendment if approved).
3. Re-grade and re-evaluate decision.
4. Stop once pass is achieved or max iterations reached.

Default max iterations: 5.

If max iterations are exhausted without passing, mark result as fail and escalate per local policy.

---

8) Required Sample Values

The following sample is included as a normative example:

• accuracy: 92
• efficiency: 88
• constraint_adherence: 96
• quality: 90
• timeliness: 94
• global_score: 92
• confidence: 0.91
• binary_decision: 1
• result: PASS

---

9) Sample JSON Grading Result

{
  "rubric_version": "halees-dual-layer-v1",
  "category_scores": {
    "accuracy": 92,
    "efficiency": 88,
    "constraint_adherence": 96,
    "quality": 90,
    "timeliness": 94
  },
  "global_score": 92,
  "threshold": 85,
  "binary_decision": 1,
  "confidence": 0.91,
  "result": "PASS",
  "iteration": 2,
  "feedback": []
}

---

10) Sample Markdown Grading Result

### Grading Result
- rubric_version: halees-dual-layer-v1
- accuracy: 92
- efficiency: 88
- constraint_adherence: 96
- quality: 90
- timeliness: 94
- global_score: 92
- threshold: 85
- binary_decision: 1
- confidence: 0.91
- result: PASS
- iteration: 2

---

11) Implementation Notes (Public Boundary)

This rubric specifies public grading semantics and reporting structure only.

It does not disclose closed-source implementation details such as internal model routing logic, proprietary grading engines, hosted enforcement internals, or production runtime infrastructure.