Atomic Trust™ TPF

Free uncertainty quantification for diffusion language models.

---

The Discovery

TPF (Tokens-Per-Forward) reveals generation confidence in diffusion LLMs at zero cost.

Query	TPF	Interpretation
7 × 8 = ?	13.58	High confidence (pattern retrieval)
23 × 17 = ?	2.68	Low confidence (active computation)

Your model is telling you when it's uncertain. Listen.

Correlation: r = -0.88 (p < 0.001) between TPF and output entropy across 28 benchmarks.

---

Inverse Signal: Hallucination Detection

Low TPF on uncertain query  → Expected (honest uncertainty)
High TPF on uncertain query → Potential fabrication (flag for review)

This bidirectional signal is structurally unavailable in autoregressive models.

---

FlashAPCE: Attention Verification

FlashAPCE enables attention verification inside FlashAttention's tiled computation without materializing the O(n²) attention matrix.

Verification signals extracted in-kernel:

• Conservation deviation
• Entropy fingerprint
• Attention distribution metrics

Performance: Sub-1% overhead on 7B+ parameter models.

Velado's Contradiction Theorem: D(ε) × I(ε) ≥ κ

Attacks cannot simultaneously achieve high impact AND low detectability. This is a mathematical bound, not a heuristic.

---

Installation

pip install atomictrust-tpf

Optional GPU dependencies:

pip install atomictrust-tpf[gpu]   # torch + triton
pip install atomictrust-tpf[full]  # + vllm, transformers

---

Quick Start

from atomictrust_tpf import TPFRLMHybrid

router = TPFRLMHybrid()
result = router.generate("What is 23 × 17?")

print(f"Answer: {result.text}")
print(f"TPF: {result.tpf_metrics.tpf:.2f}")
print(f"Route: {result.route.name}")

---

Architecture

┌─────────────────────────────────────────────────────────┐
│                 Atomic Trust™ TPF Router                │
├─────────────────────────────────────────────────────────┤
│  1. Semantic Pre-Filter                                 │
│     └── Catches temporal/ambiguous queries              │
├─────────────────────────────────────────────────────────┤
│  2. Diffusion Generation                                │
│     └── Extracts TPF + Entropy (zero-cost signals)      │
├─────────────────────────────────────────────────────────┤
│  3. Confidence-Based Routing                            │
│     ├── High TPF → Direct response                      │
│     ├── Low TPF  → Escalate for verification            │
│     └── Anomalous TPF → Flag for review                 │
├─────────────────────────────────────────────────────────┤
│  4. FlashAPCE Verification                              │
│     └── Velado's Theorem guarantee                      │
└─────────────────────────────────────────────────────────┘

---

Validation

# Run test suite
pytest tests/ -v

# Quick validation
python -c "from atomictrust_tpf import TPFRLMHybrid; print('OK')"

---

Citation

If you use this work, please cite:

@article{velado2025tpf,
    title   = {Parallel Decoding as Intrinsic Uncertainty:
               Tokens-Per-Forward Reveals Generation Confidence
               in Diffusion Language Models},
    author  = {Velado, Rafael},
    year    = {2025},
    doi     = {10.5281/zenodo.18111467}
}

@article{velado2025contradiction,
    title   = {Velado's Contradiction Theorem: Mathematical
               Guarantees for Attention Verification},
    author  = {Velado, Rafael},
    year    = {2025},
    doi     = {10.5281/zenodo.18079657}
}

---

Patents

This work is protected under US Provisional Patent Applications:

• US 63/951,960 — TPF-Based Confidence Routing (filed Dec 31, 2025)
• US 63/948,782 — FlashAPCE Attention Verification (filed Dec 26, 2025)

Open source under Apache 2.0 for research and evaluation. Commercial licensing: licensing@atomic-trust.com

---

Author

Rafael Velado LinkedIn · Atomic Trust™

---

License

Apache 2.0 with patent grant. See LICENSE.

---

Atomic Trust™ is a trademark of Rafael Velado.