Deterministic by Design
Cryptographically Verifiable
Geometric Anomaly Detection
Triaxis AI -- Deterministic Intelligence

AI that can prove every decision it makes. Exactly.

A deterministic substrate for sovereign, auditable AI -- deploy on your own hardware, keep proprietary knowledge inside the enterprise, and reproduce every result exactly.

Triaxis Atum -- Deterministic AI Suite

18 modalities. One kernel. Zero stochastic drift.

A single deterministic primitive collapses into 18 production-grade AI modalities -- 6 that beat stochastic tools outright, 12 with no mainstream ML equivalent.

18 ModalitiesVerifiable InferenceCompositional Programming0.0 Py/Rust Delta1 Kernel Parameter
Triaxis GDI -- Geometric Deterministic Intelligence

Find known and unknown anomalies in your data.

Any data expressed as a vector: sensor telemetry, transactions, network features, or embeddings from another system. Same measurement engine everywhere -- zero configuration between domains.

Known & Unknown AnomaliesZero ConfigurationNo Training LabelsNo GPU14 Benchmark Datasets
The Sheffer Kernel
One primitive. 18 modalities.
Kernel
Neural Networks
Symbolic Regression
Reinforcement Learning
Neuro-Symbolic
Generative Models
Bayesian Optimization
Compositional Programming
Verifiable Inference
Dynamical Inference
Rigorous Enclosure
Exact Expectation
Behavioral Equivalence
Exact Inversion
Distribution Fingerprints
Conformal Prediction
Spectral Engine
Atum Tensor
ASM Shadows

Six compete head-to-head with stochastic AI. Twelve more are kernel-native capabilities with no mainstream machine learning equivalent.

18
Modalities on one kernel
1
Trainable kernel parameter
0.0
Max Py/Rust inference delta
0
Violations in 10M enclosure probes

How this is different from the AI you already know

For most people, “AI” means a chatbot. Triaxis is something else entirely -- infrastructure for decisions that have to hold up under audit, not another app that writes text.

Not an LLM.Provable models (Atum) and geometric anomaly detection (GDI) for production data.
What most people picture when they hear "AI"

Chatbots and large language models

  • --Tools like ChatGPT, Claude, and Copilot -- you type a question, the model writes an answer.
  • --Impressive for conversation, drafting, and search -- but answers can vary, and the reasoning is not something you can replay or prove.
  • --That is not what Triaxis builds. We are not another LLM company.
What high-stakes teams need

Decisions you can stand behind

  • --Run the model again and get the same answer -- not "close enough," but identical.
  • --See which inputs drove the outcome, not a black-box guess after the fact.
  • --Produce evidence an auditor, regulator, or customer can verify independently.
Where Triaxis sits

Deterministic AI plus geometric monitoring

  • --Triaxis Atum is decision-grade AI -- fraud scores, risk models, sensor analytics -- built from a deterministic kernel, not a chat interface.
  • --Triaxis GDI measures vector data for anomalies -- including LLM embeddings -- without requiring you to adopt Atum first.
  • --Use either product alone, or both together when you need provable models and geometric detection on the same streams.

The sections below go deeper into the Sheffer kernel, individual modalities, and measured benchmarks. Start here if Triaxis is new to you.

The gap in mainstream AI

Conventional AI -- the kind behind most chatbots, recommenders, and fraud scores -- is built on randomness and approximation. It can be very accurate, but it was not designed to give you the same answer every time or a proof you can hand to an auditor.

Teams in regulated industries often paper over that gap with shadow models, sampled audits, and manual logs. That works until someone asks: “Show me exactly how this decision was produced.”

What Atum changes

Atum starts from a deterministic mathematical kernel -- one shared primitive instead of millions of opaque parameters. The goal is not just good predictions, but predictions that are repeatable, traceable, and backed by evidence.

Reproducibility, audit trails, and cryptographic verification are properties of how the math works -- not features bolted on after training or checked only by sampling a fraction of decisions.

18 modalities. One kernel.

Six compete head-to-head with stochastic AI. Twelve more are kernel-native capabilities with no mainstream machine learning equivalent.

The 6 -- compete with stochastic AI

Direct successors to established stochastic approaches, benchmarked head-to-head.

Neural Networks
97.78% MNIST accuracy
vs. ReLU 96.89%, GELU 97.03%, Swish 96.69%
Symbolic Regression
43W/31T/26L vs PySR on 100 AI Feynman equations
PySR 1.5.10, matched wall-clock
Reinforcement Learning
7/8 episodes stabilized, mean return -16.50
vs. LQR: 3/8, -24.71 -- 21-parameter policy
Neuro-Symbolic Execution
Heating R2 0.9963, Cooling R2 0.9944, Concrete R2 0.8777
vs. Ridge: 0.9335 / 0.9313 / 0.4342
Generative Models
5/5 UCI datasets beat RealNVP; 1,064/1,064 bit-exact Python/Rust
72,102-79,051 samples/sec at batch 1024 on a two-gen-old GPU
Bayesian Optimization
6W/7L full suite; 11.5x lower regret on constrained Rosenbrock
5-20x faster than scikit-optimize, CPU only

The 12 -- only possible on the Sheffer kernel

Input-space, data-space, and representation capabilities with no mainstream machine learning equivalent.

Weight-space algebra
Compositional Programming

Trained networks as math objects: add, subtract, SHA-256 commit, combine with Boolean logic. No stochastic equivalent.

Max delta ~7.5e-7 on 1,000 probes -- float32 ULP limit
Learn more
Cryptographic audit
Verifiable Inference

Every decision ships with a 128-byte proof. Bit-exact Python/Rust forward pass with Groth16 verification.

0.0 max delta Python/Rust -- 2.7ms verify -- ~250 proofs/sec
Learn more
Dynamical systems
Compositional Dynamical Inference

35-parameter formulas for complex systems with a zero-knowledge audit package. Not a Neural ODE approximation.

NRMSE 0.048 at N=64 -- Neural ODE never breaks 0.05 within N=512
Learn more
Input-space bounds
Rigorous Enclosure Engine

Guaranteed output bounds over input boxes for committed Sheffer models -- theorems, not Monte Carlo samples.

0 violations in 1e7 probes -- byte-deterministic certificates
Deterministic UQ
Exact Expectation Engine

Exact means, variances, and Sobol indices for shallow Sheffer models. Monte Carlo replaced by closed-form quadrature.

2056-3349x speedup vs MC -- Sobol agrees with SALib within tolerance
Function identity
Behavioral Equivalence

Canonical normal form where equal bytes mean equal functions -- court-grade model forensics and deduplication.

200k rewrite variants, 0 mismatches -- 100k ULP-tamper pairs, 0 collisions
Certified recourse
Exact Inversion

Run committed Sheffer kernels backwards: exact preimages, interval clamp regions, and provably minimal counterfactuals.

Round-trip worst 3.04e-10 on 1e6 probes -- ReLU preimage is NP-hard
Data-space drift
Distribution Fingerprints

Deterministic drift detection on any tabular data via committed kernel mean embeddings -- not limited to Sheffer models.

Power 0.367 vs RBF-MMD 0.316 -- byte-identical across rebuilds
Calibrated intervals
Conformal Prediction

Distribution-free prediction intervals around any deterministic Atum point predictor -- no posterior sampling.

Coverage within +/- 1% at three levels -- committed calibrator artifact
Ordered signals
Deterministic Spectral Engine

Hash-committed spectral features for time-series and waveform data -- cyber, clinical, and industrial signal paths.

ECG AFib test AUC 0.8637 -- byte-stable bundle hashes across runs
High-dimensional regression
Atum Tensor

Sheffer-native tensor train with rank-cliff hardening and hash commitment -- mixed-alpha CompProg composition.

14/16 task cells at R2 >= 0.99 -- 16/16 clear R2 >= 0.85
Deterministic distillation
ASM (Algebraic Shadow Models)

Distill any stochastic model into a committed CPU shadow with an empirical disagreement bound over a declared envelope.

98.21% true-label retention at 40.11% parameter count
Triaxis GDI -- Geometric Deterministic Intelligence

Geometric anomaly detection. With or without Atum.

GDI stands for Geometric Deterministic Intelligence. It finds anomalies in data by measuring distances and angles in mathematical space -- known rare events and novel patterns alike -- not by training another opaque model.

Geometric Deterministic Intelligence finds anomalies in data through deterministic geometric measurement -- known rare events and novel patterns alike. No labeled training data, no domain-specific models, and no GPU required.

GDI is a standalone product. Most teams can deploy it on existing data streams without adopting Atum. The two products are siblings -- built to run together when you want provable models and continuous geometric measurement in one stack.

GDI at a glance
Portfolio-measured on commodity CPU
14
Public benchmark datasets
156s
Back-to-back CPU run (same code)
97.6%
Best F1 (ESC-50)
Zero Config
No retuning between datasets
59,515/sec
Peak throughput (Household Power)
Measurement, not modeling

GDI computes distances and angles in the space your data already occupies. No millions of fitted parameters, no stochastic training runs.

The explanation is the detection

Every flag reports which features drove the score and by how much -- verifiable facts about the data, not a separate interpretability layer.

Any vector, zero configuration

Raw telemetry, tabular features, or embeddings from any model. Same deterministic engine routes to the right geometric measurement automatically.

Three ways teams deploy GDI

GDI on your data

Detect anomalies in telemetry, transactions, sensors, or embeddings -- standalone, without Atum.

GDI, then Atum

Find where deterministic, auditable inference matters most, then bring in the Sheffer kernel for those workloads.

Atum, then GDI

Deploy deterministic models and keep geometric measurement on the same streams for drift and novel deviations.

Built for regulated industries

Technical properties teams in these sectors often need -- reproducibility, explainability, and geometric anomaly detection.

Framework names below are context only. We do not claim certification, clearance, or formal validation against any regulatory standard.

Financial Services
Context: SR 11-7 / OCC model risk

Teams often need reproducible model runs and audit trails. Atum targets byte-exact inference and algebraic accountability -- we have not validated against SR 11-7 or similar rules.

Healthcare
Context: FDA SaMD / clinical AI

Regulated health AI teams care about traceable decisions and evidence packages. Our verifiable inference path is built for that class of need -- not a cleared or certified medical device.

Defense
Context: CDAO / mission assurance

Mission systems favor deterministic, explainable behavior. The kernel is designed for traceability and proof -- formal accreditation against DoD frameworks is the customer's process, not ours.

Aerospace & manufacturing
Context: DO-178C / industrial QA

Safety-critical software teams reference assurance levels like DO-178C. GDI and Atum provide measurement and reproducibility properties those programs often ask for -- without claiming compliance.

The numbers, exactly as measured

Two products, two evidence bases. Atum metrics come from the Sheffer kernel portfolio; GDI metrics come from the public multi-domain benchmark program.

Triaxis Atum

Deterministic AI suite -- Sheffer kernel
18
Modalities on one kernel
1
Trainable kernel parameter
0.0
Max Py/Rust inference delta
0
Violations in 10M enclosure probes

Triaxis GDI

Geometric Deterministic Intelligence
14
Public benchmark datasets
156s
Back-to-back CPU run (same code)
97.6%
Best F1 (ESC-50)
Zero Config
No retuning between datasets
59,515/sec
Peak throughput (Household Power)

Start with the pilot. One model risk portfolio. 90 days.

Measured against your own baseline costs.