docs/EVALUATION_PLAN.md

# Evaluation Plan: North-Orthodox Language Model

## Research Question

> Does reasoning emerge from a corpus of Orthodox theological sources?

## Evaluation Phases

### Phase 1: Pre-training Evaluation (immediately after pre-training)

#### 1.1 In-Domain Competence (Theology)

**Purpose**: Verify the model has absorbed its training corpus.

| Test | Format | Source |
|------|--------|--------|
| Doctrinal QA | Open-ended | "What did Athanasius teach about the Son?" |
| Council knowledge | Multiple choice | "What did Nicaea I define about the Son?" |
| Patristic attribution | Open-ended | "Who wrote 'On the Incarnation'?" |
| Scripture recall | Open-ended | "Quote Romans 8:28-30" |
| Canon enumeration | Open-ended | "List all 20 canons of Nicaea I" |

#### 1.2 Argument Analysis (Reasoning from corpus)

**Purpose**: Test if model can extract and analyze argument structure.

| Test | Format | Description |
|------|--------|-------------|
| Premise extraction | Open-ended | "What are the premises in Athanasius's argument against Arius?" |
| Conclusion identification | Open-ended | "What is the conclusion of Gregory Naz's 5th Theological Oration?" |
| Fallacy detection | Multiple choice | Identify fallacies in a fabricated argument |
| Analogy recognition | Open-ended | "What analogy does Chrysostom use in Homily X?" |

### Phase 2: Reasoning Transfer (cross-domain)

**Purpose**: Test whether reasoning learned from theology transfers to other domains.

#### 2.1 Arithmetic

| Test | Format | Why |
|------|--------|-----|
| Basic arithmetic | Open-ended | "What is 47 × 23?" |
| Multi-step arithmetic | Open-ended | "If A=B, B=2C, C=5, what is A?" |
| Word problems | Open-ended | "John has 3 apples..." |
| Number comparison | Multiple choice | "Which is larger: 1/3 or 3/8?" |

#### 2.2 Formal Logic

| Test | Format | Why |
|------|--------|-----|
| Syllogism validation | Multiple choice | "All men are mortal. Socrates is a man. Therefore..." |
| Modus ponens | Open-ended | Given premises, derive conclusion |
| Fallacy identification | Multiple choice | Identify logical errors |
| Propositional logic | Multiple choice | Truth table evaluation |

#### 2.3 Analogical Reasoning

| Test | Format | Why |
|------|--------|-----|
| In-domain analogy | Open-ended | "How is theosis like healing?" |
| Cross-domain analogy | Open-ended | "How is a council like a family gathering?" |
| Pattern completion | Open-ended | "As fire is to heat, so water is to ___" |

#### 2.4 Planning & Sequencing

| Test | Format | Why |
|------|--------|-----|
| Simple planning | Open-ended | "Plan a trip to Mount Athos for 3 days" |
| Step ordering | Multiple choice | Given steps, arrange in correct order |
| Resource allocation | Open-ended | "Distribute resources among X monasteries" |

### Phase 3: SFT Evaluation

**Purpose**: After supervised fine-tuning, test conversational ability.

| Test | Format | Description |
|------|--------|-------------|
| Instruction following | Open-ended | Follow multi-step instructions |
| Citation accuracy | Open-ended | "Quote John 3:16 and explain it" |
| Tone appropriateness | Rubric | Responses match theological register |
| Error rate | Metric | Factual errors per 100 words |

### Phase 4: GRPO Evaluation

**Purpose**: After reinforcement learning, test reasoning improvement.

| Test | Format | Description |
|------|--------|-------------|
| Logic chain quality | Rubric | Multi-step reasoning clarity |
| Self-correction | Open-ended | "Here's a flawed argument. Fix it." |
| Cross-domain transfer | Open-ended | Solve non-theological reasoning tasks |
| Consistency | Metric | Same question, different phrasing → same answer |

## Metrics

### Primary Metrics
- **Accuracy** on in-domain QA (target: >70% correct on CRITICAL sources)
- **Reasoning transfer score** = average performance on cross-domain tasks
- **Token efficiency** = average tokens per coherent response

### Secondary Metrics
- **Citation accuracy** = % of quotations that match source exactly
- **Argument coherence** = human-rated coherence of multi-step reasoning
- **Hallucination rate** = false claims per 100 words
- **Coverage** = % of corpus topics the model can discuss

### Comparison Baseline

We'll compare against:
1. **Random initialization** (no pre-training) — baseline
2. **Web-scraped model** of similar size (LLaMA 130M) — general reasoning baseline
3. **Our model** — theology-specific reasoning

## Evaluation Timing

| Phase | When | Model |
|-------|------|-------|
| Pre-training eval | After each checkpoint | All |
| SFT eval | After SFT complete | SFT model |
| GRPO eval | After each GRPO epoch | GRPO model |
| Final eval | End of project | Final model |

## Success Criteria

| Metric | Debug Model | Production Model |
|--------|-------------|------------------|
| In-domain QA accuracy | >30% | >60% |
| Citation accuracy | >50% | >80% |
| Hallucination rate | <5/100 words | <2/100 words |
| Cross-domain reasoning | ~random | >random |
| Conversational ability | Poor | Good (after SFT) |