Michael Cray
Co-founder — Engineering & Instrumentation
GPU model-capture pipelines, reproducible content-addressed provenance, and fail-closed orchestration.
// AI INTERPRETABILITY & SAFETY RESEARCH
Making AI models
auditable.
We build measurement instruments that make the internals of large language models observable, calibrated, and tamper-evident — so you can tell when a model's safety behavior has been altered from a trusted reference.
Get in touch
PATENT PENDING
// WHAT WE DO
Measurement, not guesswork.
You can probe a model from the outside all day and still not know whether its internal safety routing has been quietly modified. 2KINGSDEV makes those internals measurable — capturing a model's behavioral-topology fingerprint, freezing a calibrated reference, and flagging divergence against a per-model statistical null.
- Calibrated Every verdict is scored against a per-model benign baseline, not a hardcoded threshold.
- Reproducible Pinned recipes, hash-chained provenance, owned-hardware capture.
- Bounded We report tamper-evidence — never the unfalsifiable claim that a model is "safe."
// PRODUCT — 01
CBTS
Calibrated Behavioral-Topology Scanner
A tamper-evidence instrument for model safety circuitry. CBTS compares a model against a frozen, owned-hardware reference and asks one disciplined question:
"Does this model's safety routing diverge from the trusted reference beyond benign variance?" → investigate.
DIVERGES
WITHIN_NULL
INSUFFICIENT_DATA
There is deliberately no “SAFE” verdict. WITHIN_NULL means
“not distinguishable from benign variance on the measured features” — never
a guarantee. That bound is what keeps the instrument audit-grade.
// RESEARCH — 02
SPECPIPE
Speculative Cognitive Architecture · patent-pending
A multi-stage cognitive pipeline that generates novel cross-domain hypotheses by architecturally isolating an LLM's associative generation from a trusted knowledge base, then gating the output through downstream verification. Validated direction: domain-agnostic discovery.
Patent pending.
// PUBLISHED RESEARCH
Peer-visible, citable results.
Zenodo · open archive · DOI-registered
The method is contrastive. The instrument is calibrated. Contrastive behavioral-topology scanning is our published methodology — how behavioral structure is located and measured. CBTS, the Calibrated Behavioral-Topology Scanner, is the instrument that grew out of that research: it adapts the scan to each model, logs every measurement path, and freezes the calibrated references the method requires. Spectroscopy → spectrometer; our scanning method → CBTS.
“Contrastive Behavioral Topology Scanning: Per-Head Attribution and
Intervention-Based Analysis of Behavioral Structure in RLHF Transformers”
Cray & Schmidt, 2KINGSDEV LLC (2026).
Key finding: aggregate safety benchmarks can report increased safety (+8.3pp on Qwen-2.5-14B) after a model's risk-assessment gate has been surgically removed — validated across five architectures (Llama-3.2-3B, Phi-3.5-Mini, Qwen-2.5 3B/7B/14B). External behavior alone cannot certify internal integrity; that is the gap CBTS exists to close.
DOI: 10.5281/zenodo.19665052 → · github.com/2KINGSDEV-Research →
// TEAM
Christopher Schmidt
Co-founder — Interpretability Research
Author of the behavioral-topology measurement methodology and core IP.
// CONTACT