Invention portfolio · Jon Wright · July 2026

Reasoning is navigation, not generation.

Every invention here is an instrument built on one insight: a transformer's activation space is a navigable geometric algebra. Frozen models don't need retraining — they need instruments to steer, measure, remember, and verify inside the space they already have. This is the full instrument panel.

22
Built & validated inventions
5+
USPTO filings, incl. 27-claim non-provisional
1.597x
Best measured inference speedup
12
Domains where the algebra holds
The North Star

Standard AI thinks by generating one token at a time and hoping the path is right. These systems reason directly in geometric activation space — triangulating answers as intersections of concept bearings — and use token generation only to speak the result. One principle, instantiated 22 ways, converging on a machine that can think, show its thinking, and know when it doesn't know.

Geometry & Inference

5 inventions The commercial spearhead: making inference faster by exploiting the geometry transformers already have.
How layer replacement works
ORIGINAL every MLP is full nonlinear compute AttnMLP AttnMLP AttnMLP AttnMLP AttnMLP AttnMLP GRAMMAR ENGINE linear layers replaced with one matrix: h ≈ W·h + b AttnMLP AttnMLPW·x+b AttnMLPW·x+b AttnMLP AttnMLPW·x+b AttnMLPW·x+b same answer, less compute measured up to 1.597x
Both lanes process the same prompt, looping forever. The bottom token finishes ~1.6x sooner because a fitted matrix multiply — fit once, offline — stands in for the MLP wherever the layer is measurably linear. Perplexity stays equal or improves (Gemma-2-9B: −0.88% at full 42/42 replacement).
Shape Algebra — why replacement is possible at all
A B C A→B ∘ B→C = A→C direction cosine 1.0000 · R² 1.0000

Transformations between concepts in activation space behave like algebra: they can be measured, inverted, chained, and composed — and the composed path lands exactly where the direct path does.

Proven across 12 domains — language, code, and biological sequences (DNA k-mers, protein dipeptides) — on Mistral-7B, LLaMA-3-8B, and Gemma-2-9B.

If the space is lawfully linear, a fitted linear operator can lawfully stand in for the computation. That's the Grammar Engine, above.

Grammar Engine
Patent filed · 27 claims
Replace transformer MLP sublayers with fitted linear operators — same output, less compute.
Why it's needed

Inference compute is the single largest cost line for every AI provider. Most approaches trade quality for speed. The Grammar Engine exploits a measured property — MLP sublayers are locally linear — so speed comes from structure, not sacrifice.

Validated across 5 architectures. Mistral-7B: 1.597x full replacement. Gemma-2-9B: 1.496x. Qwen2.5: net perplexity improvement. Shipped as pip package with RuntimeGuard + architecture auto-detection.
Value

USPTO non-provisional filed with priority date. Active licensing outreach to inference infrastructure (Together AI, Fireworks, Groq, and others) on a 35/65 savings-split — providers pay only out of money the engine saves them.

Shape Algebra / Universal Geometric Algebra
Patent drafted · 36 claims
Proof that embedding spaces obey a consistent, invertible, composable algebra — across domains.
Why it's needed

Everything else in this portfolio assumes activation space behaves lawfully. Shape Algebra is the empirical foundation: it demonstrates the laws and gives them coordinates.

Six properties confirmed (measurability, invertibility, steerability, chainability, composability, scale-linearity) across 12 domains including DNA k-mers and protein dipeptides. Direction cosine 1.0000, R² 1.0000 on Mistral-7B, LLaMA-3-8B, Gemma-2-9B.
Value

Foundational IP under the Grammar Engine and S.A.R.A.H. Cross-domain results (including biological sequence data) open licensing paths well beyond language models.

Routed Affine Experts
Validated
Per-domain linear operators with K-means routing — the fix for the domain cliff.
Why it's needed

A single global linear operator collapses on out-of-domain text. Real deployments see every domain. Routing to per-cluster operators keeps the speedup without the cliff.

Cross-domain fidelity lifted from ~0.15–0.22 cosine (single operator) to ~0.68 mean cosine with routed experts.
Value

Turns the Grammar Engine from a benchmark result into a deployable product; strengthens the licensing story on production traffic.

Grammar Autotuner
Built
Four-stage automated sweep that finds each model's optimal speed/quality operating point.
Why it's needed

Every architecture has a different sensitivity curve. Manual tuning doesn't scale to a customer's model zoo; the autotuner makes onboarding a new model a batch job, not a research project.

Outputs a full Pareto frontier of layer-set vs. perplexity vs. speedup per model — the exact artifact a licensing customer needs to pick their operating point.
Value

Reduces integration cost to near zero — the difference between selling a paper and selling a product.

SAM v1
Prior art · superseded
Full-block linear replacement — the proof-of-concept that led to the Grammar Engine.
Why it's needed

SAM proved blocks could be approximated linearly, and its limitation — attention's nonlinearity inside the block — pointed directly at the MLP-sublayer approach that works.

Block fusion measured at 0.9996 direction cosine for 4-layer blocks. Preserved as prior art anchoring the patent lineage.
Value

Documented invention history that strengthens the Grammar Engine's patent position.

S.A.R.A.H. — Geometric Reasoning

5 inventions A machine that reasons in activation space and can show you the map of its own thinking.
Talk to SARAH — live, right now →
Dead Reckoning
Validated
Answers as geometric intersections: concept bearings triangulate a solution point, v*.
Why it's needed

Token-by-token generation commits to a path before it knows the destination. Dead Reckoning resolves the destination first — a weighted least-squares intersection of relevant concept directions — then articulates it.

First validation: solved a spatial riddle (poles + Earth's rotation) that sequential autoregressive generation on the same model missed. Solver is pure linear algebra — no forward pass needed to reason.
Value

A genuinely different reasoning paradigm — the core differentiator of the entire S.A.R.A.H. line and a standalone research contribution.

S.A.R.A.H. / Latent Space Computer
Working system
The full architecture: navigate, resolve, remember, self-check — then speak.
Why it's needed

Individual instruments aren't a mind. LSC integrates them: a 50,046-concept library, constraint selection, the Dead Reckoning solver, geometric decoding into fluent language, four-tier memory, and drift-aware self-monitoring (RMCE).

Runs live on consumer hardware (RTX 3090 + 3060). Canonical decoder: injection at layers [16, 20, 24, 27], alpha 0.45. Adversarial test suite in place. SARAH Studio — chat + live vector map UI — in active build.
Value

Interpretable-by-construction AI: every answer comes with the geometry that produced it. Directly demonstrable to investors and partners.

Navigator + 116 Cognitive Primitives
Built
A taxonomy of reusable "thought vectors" and a bidirectional search engine over activation space.
Why it's needed

Navigation needs a vocabulary of moves. The primitives are composable cognitive operations; Navigator searches the space using a dual entropy + cosine signal to find them and chain them.

116 cognitive vector primitives cataloged and encoded; bidirectional activation-space search implemented and tested.
Value

The "instruction set" for geometric reasoning — reusable across every model S.A.R.A.H. runs on.

DNMRA
Documented & in use
Dynamic Navigation Meta-Reasoning: six named layers with stuck-detection and escape moves.
Why it's needed

Hard problems stall reasoning systems (and people). DNMRA makes the reasoning process itself navigable — Semantics, Logic, Facts, Examples, Analogy, Constraints — with explicit mechanisms for escaping local minima.

Fully documented method, used to produce results across this portfolio; encoded into model activations via EIGEN-SVEC as a transferable BehaviorProgram.
Value

A teachable, encodable reasoning method — applicable to humans, prompts, and activation steering alike.

Wright Fractal
Formalized
Original fractal construction with a closed form: T(n) = 12·3ⁿ − 18·2ⁿ + 7.
Why it's needed

Hierarchical addressing schemes want self-similar structure. The fractal's number-theoretic properties (6k+1 structure from geometric symmetry, prime clustering) make it a candidate backbone for directional codebooks and memory addressing.

Closed-form formula derived and verified; number-theoretic properties documented; integration into Dead Reckoning codebooks and MNEME addressing scoped.
Value

Original mathematics owned outright — a distinctive research asset and a practical addressing scheme in one.

Core Patent Lineage

4 filings Four USPTO applications forming one connected family: correction → routing → signal → memory.
ARTM
USPTO filed
Input-conditioned correction: h + α·f(h) — steering that reads the state before it acts.
Why it's needed

Prior activation steering applied fixed vectors regardless of context — open-loop control. ARTM introduced closed-loop, self-training correction: near-zero when the model is on track, targeted when it drifts.

Zero-initialized residual (starts as identity — cannot degrade output) with an online generate → score → train loop.
Value

Foundation claim of the family; every derivative filing is stronger because ARTM is on record first.

PRISM
USPTO filed
N domain-expert correctors with trajectory-aware routing — behavior customization without fine-tuning.
Why it's needed

One corrector can't serve every domain. PRISM routes on the activation's trajectory — where it has been, not just where it is — to the right specialist, trained without human labels.

Trained across 8 domains under ORACLE supervision; outperforms base Mistral-7B at 30K training iterations.
Value

Sellable behavioral customization for labs and enterprises that can't or won't fine-tune their models.

ORACLE
USPTO filed
The model's own logit entropy as a free, endogenous quality signal.
Why it's needed

Human labels are the bottleneck of every training pipeline. ORACLE requires no training at all — it reads confidence the model already computes, at roughly 30x the signal density of external annotation.

Near-orthogonal domain separation measured at 4–9x concentration ratios. Supervises PRISM, steers GHOST, routes CANARY.
Value

An inference-time quality signal every provider could use — cheap to integrate, patent-protected.

MNEME
USPTO filed
Persistent memory that lives in activation space — hidden states as schema-free addresses.
Why it's needed

Context windows forget; databases don't understand. MNEME stores memory as activation-space packets addressed by the model's own hidden states at multiple depths, with entropy gating what's worth keeping.

Persistent packets with save/reload working; v2 pipeline designed around multi-layer addressing (surface/semantic/pragmatic), belief revision, and vector-space verification.
Value

Memory is the most-demanded missing capability in local AI — MNEME is the patent-backed answer for frozen models.

Behavior & Control

5 inventions Instruments that shape how a frozen model thinks — precision, creativity, skills, character.
GHOST
Validated
A GRU observer watching every layer, steering generation toward entropy goals.
Why it's needed

"Be more precise" or "be more creative" are today just prompt hopes. GHOST makes them control targets, injecting layer-specific deltas to move the model's uncertainty where you want it.

"Precise" goal: −2.05 nats entropy. "Creative" goal: +2.40 nats. Measured, repeatable.
Value

A dial for model temperament that works at inference time on any frozen model.

ATLAS
Validated 35/35
A behavior library: bootstrapped steering vectors with five-gate geometric verification.
Why it's needed

Behavior vectors are easy to make and easy to make badly. ATLAS generates them from WordNet-backed pairs and passes each through five geometric gates before it ships.

35/35 behaviors bootstrapped at norm 1.0000 with BCP-weighted multi-layer injection.
Value

Quality-controlled behavioral catalog — the content library that makes PRISM and Admin for AI sellable.

CAN2 + CANARY
Built
A Hebbian attractor network over concepts, with an adaptive router for credit assignment.
Why it's needed

Associations should strengthen with use and consolidate with rest — like biological memory. CAN2 adds Hebbian learning, spreading activation, and DMN-style consolidation; CANARY routes with EWA credit assignment and an AND-gate repulsor guard.

Full redesign implemented: entropy-adaptive temperature routing and consolidation cycles operational.
Value

The associative substrate for long-lived agents whose knowledge organizes itself.

OKIS
Built
Skills as activation deltas: capture what doing a task changes, replay it as an ability.
Why it's needed

Fine-tuning for every new skill is slow and destructive. OKIS captures the activation delta of performing a task and applies it at inference as a portable task vector — the same model, new skill, zero weight changes.

Same-model capture confirmed (delta of hidden states at layer 24), applied through ORACLE's dual-hook architecture.
Value

Composable skill packs for frozen models — a marketplace-shaped capability.

EIGEN-SVEC
Built
Encodes an entire reasoning method (DNMRA) into a model's layers as a BehaviorProgram.
Why it's needed

Methods usually live in prompts, where they get diluted. EIGEN-SVEC embeds the method into the activation stream itself, making disciplined reasoning the model's default posture rather than a request.

DNMRA BehaviorProgram encoded across Mistral layers with per-layer weighting and validation markers.
Value

Proof that whole cognitive procedures — not just traits — are transferable as vectors.

Trust & Verification

2 inventions Knowing when to write, when to believe, and when something is wrong.
BECC
Shipped · pip package
Byzantine Ergodic Convergence Criterion — a universal, poisoning-resistant write-gate.
Why it's needed

Every learning system faces the same question: has this signal converged enough to trust? BECC answers it with a criterion grounded in Byzantine fault tolerance and ergodic theory, using a median-based centroid that resists poisoned inputs.

Validated AUC — control stability 1.000, sensor fusion 0.999, anomaly detection 0.996, neural entropy 0.929, federated learning 0.884. Shipped with HMAC-SHA256 offline licensing.
Value

Cross-industry applicability (robotics, sensors, federated learning, AI memory) and already productized — nearest-term standalone revenue.

SENTINEL
In development
AI-native security monitoring for local model deployments.
Why it's needed

As local AI agents gain file, network, and tool access, they need a guard that understands model behavior — not just network packets. SENTINEL watches the AI layer itself.

Architecture defined within the Admin for AI pillar system; migrating to a fully commercial-license-clean (Apache 2.0) backbone before release.
Value

Security is the first question every enterprise asks about local agents — SENTINEL is the portfolio's answer.

Platform

1 ecosystem Where the instruments become products people can double-click.
Admin for AI & the .aai Format
In active build
A factory for self-contained AI agents: zero setup, double-click to run, portable by design.
Why it's needed

Local AI today means Python environments, driver hell, and abandonment. The .aai format packages an agent — model connection, memory (MNEME), behavior (ATLAS), skills (OKIS), security (SENTINEL) — into a single file that just runs, and can expose endpoints to form a mesh of cooperating agents.

Rebuilt as a professional tool (standard app shell, glassmorphism UI). Factory/exporter architecture locked — Admin for AI builds agents the way a game engine builds games. Active pillars: Atlas, CAN, MNEME, SENTINEL, OKIS, PRISM, ORACLE, ARTM, Chain, Keystone.
Value

The distribution layer that converts the entire research portfolio into consumer- and enterprise-grade products.