Phase 0 Validation - Project Charter

Phase 0 Validation - Project Charter

1. Purpose

### 1.1 Statement
Must provide empirical validation of core Anticipatory Transformer components on small scale (2M parameter model) to prove:
1. Dual-pathway architecture achieves frequency separation (fast: syntax/local, slow: semantics/global)
2. Orthogonality penalty prevents mode collapse while maintaining specialization
3. Trajectory-aware attention with additive bias is numerically stable and improves context efficiency
4. Commitment targets (counterfactual stability) correlate with generation quality
5. Kernel slice-based context selection outperforms fixed-window baselines

### 1.2 Falsifiability
If any core component fails validation (e.g., pathways collapse to identical representations, commitment targets are uncorrelated with quality), the architecture requires fundamental revision before full-scale implementation.

2. Non-goals

### 2.1 Must not attempt full-scale training
Phase 0 is validation only. Full training requires Phase 1+ with proper infrastructure.

### 2.2 Must not optimize for production metrics
Validation targets correctness and feasibility, not state-of-the-art performance.

### 2.3 Must not implement all features
Only core architectural components required for validation. Advanced features (regime detection, slice cross-attention) are Phase 1+.

3. Success Criteria

### 3.1 Frequency Separation Validation
Fast pathway shows dominant energy in 5-50 Hz band (syntax), slow pathway in 0.5-5 Hz band (semantics). Measured via FFT on hidden state time series.

Validation: Plot frequency spectra, verify <10

### 3.2 Orthogonality Enforcement
Cross-covariance Frobenius norm ||cov(h_fast, h_slow)||_F decreases during training while task performance maintains or improves.

Validation: Training curve shows L_ortho decreasing, perplexity not increasing.

### 3.3 Commitment Correlation
Counterfactual stability scores correlate with human judgments of response quality (r > 0.6) on small evaluation set (n=100).

Validation: Compute Pearson correlation between commitment scores and quality ratings.

### 3.4 Context Efficiency
Slice-based context selection achieves equivalent perplexity with 30

Validation: Measure perplexity at multiple context budgets, demonstrate compression advantage.

### 3.5 Numerical Stability
Trajectory attention with additive bias maintains stable gradients (no NaN/Inf) across 10K training steps.

Validation: Monitor gradient norms, verify no numerical errors in logs.

4. Direction Constraints

### 4.1 Must remain compatible with kernel slice interface
All context selection must use SliceExport primitives as defined in core/cc-graph-kernel/src/types/slice.rs.

### 4.2 Must use standard PyTorch patterns
No custom CUDA kernels in Phase 0. FlashAttention via official xformers library is acceptable.

### 4.3 Should target single-GPU training
Validation experiments should complete on 1x A100 (40GB) within 48 hours.

### 4.4 Should preserve reproducibility
All experiments require fixed seeds, deterministic ops, and logged hyperparameters.

5. Commitment Level

Draft. Will be promoted to Locked upon successful validation of all 5 success criteria.

6. Traceability

### 6.1 Source signals
- docs/architecture/23-ANTICIPATORY_TRANSFORMER.md (Section VI.1: Phase 0 validation plan)
- User request for continuation after architecture revision
- core/cc-graph-kernel/src/types/slice.rs (kernel interface contract)

7. Change History

- 2026-01-04 v0.1: Initial draft for Phase 0 validation experiments