Verified decision architecture for explainable enterprise AI.
Lift unstructured documents into structured logic. Scale evaluation across real datasets. Lower outputs into verified rules. Govern model swaps without rewriting a line of business logic. The same engine runs the legal vertical and the commerce vertical.
For peer-reviewed research, IP portfolio, and the founder's bio, see Research.
Six questions every regulated AI deployment has to answer.
Decision Control is the GoPX framework for what makes an AI system defensible in regulated use. Each question maps to a real failure mode of unrestricted LLMs — and to the specific GoPX capability that closes it.
-
Catastrophic forgetting
Fine-tuning a model to fit your domain often erases prior knowledge. The model you ship is not the model you trained — it's a degraded successor that forgot half of what made it useful.
GoPX: GoPX doesn't fine-tune the model — it lifts your domain logic into a separate, inspectable rule layer. The base model stays general; the domain stays editable. Forgetting is not a failure mode because the knowledge was never in the weights.
-
Weak learning
Models pick up shortcuts during training — surface correlations that work on the benchmark but fail when the distribution shifts. The shortcut you can't see is the one that breaks in production.
GoPX: Lifting forces the signal into explicit rules. If the model is winning by exploiting a shortcut, the rule layer makes the shortcut visible — operators see the rule firing, audit it, and replace it with the actual signal.
-
Prompt injection
Adversarial input rewrites the model's instructions. Whether it comes from a malicious user, a poisoned document, or an upstream LLM in an agent chain, prompt injection is the OWASP-LLM #1 risk for production systems.
GoPX: The rule layer is not a prompt. Adversarial text can change what the model says next; it cannot change what the rule layer requires the model to cite. Lowering binds outputs to the rule layer, so injection attempts surface as constraint violations rather than silent rewrites.
-
Execution exposure
When an AI agent takes action — sends a payment, signs a document, books an appointment — the action can't be undone. The window between recommendation and execution is where most operational damage happens.
GoPX: Every action runs against the verified rule set with a recorded reasoning trace. Operators can require dual-control, escalation thresholds, and pre-execution review per action class — declared as logic, not buried in prompt strings.
-
Information sovereignty
When data leaves your perimeter to be processed by a vendor's model, your sovereignty leaves with it. You inherit their data-handling posture, their geography, their breach surface.
GoPX: The rule layer and the audit trail run on infrastructure you choose. The model can be local, hosted, or proxied. The decisions and their evidence stay where you put them — no telemetry hostage.
-
LLM independence
If your AI logic is encoded in prompts and fine-tunes specific to one vendor's model, your AI roadmap inherits their roadmap. Migration costs scale with usage; the lock-in compounds quietly until it's too expensive to leave.
GoPX: The rule layer is model-agnostic by construction. Switch from GPT-5 to Claude 4 to Gemini 3 to Llama-derived models without rewriting business logic. The cost of switching becomes evaluation cost, not rebuild cost.
Four layers, one explainable surface.
Generalized from production deployments. Each layer has a clean interface; each layer is replaceable. No magic, no monolith.
- L1
Protocol interface
Open-protocol surface for incoming requests. Speaks the network's language — payments, catalog queries, compliance asks — and converts them into structured intents the engine can reason over.
- L2
GoPX intelligence engine
Lift / Scale / Lower / Govern stages run here. Documents become logic; logic becomes rules; rules constrain execution. Every artifact is editable and auditable.
- L3
Agentic execution
Agents act against the verified rule set, not against free-form prompts. Each action carries its rule citation and its evidence chain — defensible by construction.
- L4
Conversational interface
Operators interact with the rule layer in plain English: edit constraints, add exceptions, replay decisions, rerun on new data. No DSL to learn, no engineer needed for routine policy changes.
Lift / Scale / Lower / Govern.
The public product verbs map to the research vocabulary here: Lift unstructured language, Scale evaluation across datasets, Lower model output into verified rules, and Govern the application across model swaps.
- Lift
Convert unstructured enterprise language — contracts, policies, regulations, manuals, transcripts — into structured logic. Symbolic pattern lifting transforms neural outputs into editable facts, rules, and relationships. Inspectable in plain English; not embeddings.
- Scale
Evaluate the lifted logic across real datasets. DBSCAN clustering surfaces natural groupings; KRTs (knowledge representation trees) translate cluster behavior into editable rules. Multilingual — proven on Arabic, generalizes across LLM-supported languages. ANCO-HITS propagation handles network-scale signal.
- Lower
Bind every model output strictly to verified rules and indexed evidence. Hallucination drops because the model is no longer free to invent — it's constrained to cite. Peer-reviewed: 85% reduction in hallucination and toxicity. Hard-constraint satisfaction at 100% across 180 production queries on the deployed reference stack.
- Govern
A model-agnostic decision layer that travels with your application. Rules, constraints, and exceptions live here — not in a prompt, not in a fine-tune. Switch from GPT-5 to Claude 4 to Gemini 3 without rewriting a line of logic. The audit trail follows the application, not the model.
Operator-ready numbers. Peer-reviewed foundations.
GoPX is the commercial expression of three decades of research by Dr. Hasan Davulcu (Professor, Arizona State University; Director, CIPS-AI Lab) and his group. Three deployed reference implementations from our Phoenix engineering team validated the stack end-to-end.
- 100%
- hard-constraint satisfaction across 180 production queries
- 113/113
- unit tests passing on the deployed reference stack
- 85%
- peer-reviewed reduction in hallucination and toxicity
- 30+ yrs
- of published research behind the approach
Peer-reviewed in IEEE Transactions on Computational Social Systems: "Beyond the Black Box: Programmable AI and Explainable Text Analysis." — Trivedi, Çetinkaya, Cowan, Newson, Vlahović, Davulcu (forthcoming).
See it on your documents.
A 45-minute session. We run your contracts, policies, or regulations through GoPX live and return structured logic, a decision walkthrough, and an honest read on fit. No pitch deck.
- Run on your documents, not a generic demo
- Structured logic + decision walkthrough
- Honest read on fit — no pitch deck