Grand Diomande Research · Full HTML Reader

Anticipatory Transformer - Phase 0 Validation

Empirical validation of core architectural components for the Anticipatory Transformer as defined in [docs/architecture/23-ANTICIPATORY_TRANSFORMER.md](../../docs/architecture/23-ANTICIPATORY_TRANSFORMER.md).

Agents That Account for Themselves research note experiment writeup candidate score 24 .md

Full Public Reader

Anticipatory Transformer - Phase 0 Validation

Empirical validation of core architectural components for the Anticipatory Transformer as defined in [docs/architecture/23-ANTICIPATORY_TRANSFORMER.md](../../docs/architecture/23-ANTICIPATORY_TRANSFORMER.md).

Overview

Phase 0 validates 5 core hypotheses on a 2M parameter mini-transformer:

1. Frequency Separation: Dual pathways naturally specialize (fast: syntax 5-50 Hz, slow: semantics 0.5-5 Hz)
2. Orthogonality: Cross-covariance penalty prevents mode collapse while maintaining specialization
3. Trajectory Attention: Additive bias mechanism is numerically stable and improves context efficiency
4. Commitment Targets: Counterfactual stability correlates with generation quality (r > 0.6)
5. Slice-Based Context: Priority-queue selection achieves 30

Directory Structure

cc-anticipatory-transformer/
├── phase0/
│   ├── models/              # Model implementations
│   │   ├── dual_pathway.py  # Fast/slow pathway architecture
│   │   ├── attention.py     # Trajectory-aware attention with additive bias
│   │   ├── losses.py        # Orthogonality penalty and commitment targets
│   │   └── config.py        # Model configuration
│   ├── experiments/         # Validation scripts
│   │   ├── exp1_frequency_separation.py
│   │   ├── exp2_orthogonality.py
│   │   ├── exp3_trajectory_attention.py
│   │   ├── exp4_commitment_correlation.py
│   │   └── exp5_context_efficiency.py
│   ├── data/               # Datasets for validation
│   └── results/            # Experiment outputs, plots, logs
└── README.md

Requirements

  • Python 3.10+
  • PyTorch 2.1+ with CUDA 12.1
  • 1x A100 (40GB) or equivalent GPU
  • ~100GB disk space for checkpoints and data

See [requirements.txt](requirements.txt) for full dependency list.

Quick Start

bash
# Install dependencies
pip install -r requirements.txt

# Run frequency separation validation
python phase0/experiments/exp1_frequency_separation.py

# Run full validation suite
bash phase0/run_all_validations.sh

Validation Criteria

Each experiment must meet specific success criteria as defined in [23-ANTICIPATORY_TRANSFORMER_PHASE0_CHARTER.md](../../docs/architecture/23-ANTICIPATORY_TRANSFORMER_PHASE0_CHARTER.md):

ExperimentSuccess CriterionMeasurement
Frequency SeparationFast: 5-50 Hz, Slow: 0.5-5 Hz, <10
OrthogonalityL_ortho ↓ during training, perplexity stableTraining curves
Trajectory AttentionNo NaN/Inf gradients over 10K stepsGradient monitoring
Commitment Correlationr > 0.6 with quality ratings (n=100)Pearson correlation
Context Efficiency30

Current Status

Phase: Setup
Progress: 0/5 validations complete

See [../../docs/architecture/23-ANTICIPATORY_TRANSFORMER_PHASE0_CHARTER.md](../../docs/architecture/23-ANTICIPATORY_TRANSFORMER_PHASE0_CHARTER.md) for full project charter.

Citation

If you use this work, please cite:

@techreport{anticipatory_transformer_2026,
  title={Anticipatory Transformer: Motion Intelligence for Language Models},
  author={Comp-Core Architecture Team},
  year={2026},
  institution={Comp-Core Project}
}

Promotion Decision

Attach run IDs, datasets, metrics, and reproduction commands.

Source Anchor

Comp-Core/core/_recovered/cc-anticipatory-transformer/README.md

Detected Structure

Method · Evaluation · Code Anchors · Architecture