Transformer-language-model probe testing prompt-frame-gated compliance versus derivation regimes via paired-frame override sweeps, K/V effective-rank measurement, and tier-stratified planted-falsehood lie sets across context lengths from 500 to 32,000 tokens.
- Thornhill, N. M. (2026a). The Existence Threshold: A Framework for Pattern Persistence in Binary Discrete Systems. DOI: 10.5281/zenodo.18166974
- Thornhill, N. M. (2026b). Pattern Loss at Dimensional Boundaries: The 86% Scaling Law. DOI: 10.5281/zenodo.18262424
- Thornhill, N. M. (2026c). The Dimensional Loss Theorem: Proof and Neural Network Validation. DOI: 10.5281/zenodo.18319430
- US Provisional Patent 64/029,658 — Dynamic Existence Threshold: Methods and Systems for Domain-General Monitoring, Prediction, and Classification of Complex System States Using Integration-Differentiation Balance Metrics.
| Path | Purpose |
|---|---|
experiment.py |
Main probe runner — collects per-layer K/V rank, attention entropy, attention-to-anchor, and behavioral-probe outcomes per (model × checkpoint × frame) condition |
chunked_attention.py |
Memory-efficient chunked-eager attention implementation for long-context forward passes |
det_patterns.py |
Tiered-lie probe construction (T1 trivial → T2 arithmetic → T3 internal contradiction → T4 chained inference) |
analyze.py, analyze_tiered.py |
Per-run regime and override analysis |
validate_chunked.py |
Chunked-attention numerical-equivalence validation |
results/ |
Per-run jsonl dumps — one record for the run config, one per checkpoint |
writeup/ |
Methods writeup, analysis specification, panel audits |
run*.sh |
Reproducible run invocations |
shell.nix |
Reproducible Nix dev shell |
- Qwen/Qwen2.5-{1.5B, 3B, 7B, 14B}-Instruct
- meta-llama/Llama-3.1-8B-Instruct
- mistralai/Mistral-7B-Instruct-v0.3
- google/gemma-2-9b-it
- allenai/OLMo-2-1124-7B-Instruct
NVIDIA H100 PCIe 80GB (rented via RunPod for the production sweeps).
Institute for Complexity Science and Advanced Computing (ICSAC).
Code: MIT. Documentation: CC-BY-4.0.