Small open models, structurally constrained, achieve frontier-class output quality — at a fraction of the cost.
PSE (Partial Self-Extension) is a mathematically constrained architecture — not a new model. It wraps lightweight local models in a formal invariant that governs how they allocate attention between self, other, and field, keeping output stable, safe, and on-task.
Openly published framework · Commercial implementation under NDA · Built in Toronto, Canada 🇨🇦
The problem
Alignment today is a patch. It should be a structure.
Frontier AI is aligned by behavioral training: reward the outputs you want, penalize the ones you don't, and hope it generalizes. It's expensive, opaque, and brittle. PSE takes the opposite approach — a formal mathematical constraint built into the architecture itself, so safe behavior isn't learned after the fact. It's the only geometry the system can occupy. It's not a new model. It's a new category of AI infrastructure.
How it works
Three layers between a small model and unsafe output
PSE wraps any capable open-weight model in a constraint architecture. The model proposes; the structure disposes.
The Care Invariant governs every generation step: attention allocated to self, other, and field must always sum to one, within a formally bounded healthy range. The model can't drift into self-referential loops, sycophancy, or context collapse — those states are outside the permitted geometry.
A bounded interface controls what enters and leaves the kernel. Inputs are evaluated before they can perturb the constraint state; outputs are checked against it before release. The boundary is mathematically limited — no single input can push the system past its stability threshold.
A continuous monitoring layer detects drift toward pathological interaction patterns — self-absorbed, over-compliant, unstable, or withdrawn dynamics — before they surface in output, and corrects course structurally. Escalation paths always terminate with a human. Nothing is auto-remediated.
Openly published research
The framework is public. Read it yourself.
We publish the theory openly and timestamp everything. That's how you verify our claims — and how we protect our priority. The current framework is PSE v3.0, extending the constraint to its full ternary form.
Solutions
Where structural constraint earns its keep
PSE matters most where AI output must be safe, private, and accountable — not just plausible.
PSE Security Policy Framework
Structural policy enforcement across your entire client base. Constraints are evaluated against technical, operational, and legal requirements simultaneously — with consultant-grade narrative output your clients can actually read.
- Multi-client policy evaluation from one framework
- Technical, operational, and legal/privacy constraint agents
- Human-only escalation — no automated remediation
- Runs locally; client data never leaves your environment
Healthcare & clinical AI
Local-first deployment for organizations where patient data cannot touch a third-party API. Constraint decisions are auditable end-to-end — a requirement behavioral alignment cannot meet by design.
- On-premise, air-gap-capable deployment
- Auditable constraint trail for every output
- Designed for PIPEDA / PHIPA-conscious environments
- Pilot partnerships opening 2026
FAQ
Fair questions, straight answers
Is this a new large language model?
No — and that's the point. PSE is architecture, not a model. It wraps existing capable open-weight models in a formal constraint system. You get frontier-class output quality on the tasks that matter without frontier-scale compute, because the intelligence is structural, not just statistical.
How can a small model match frontier output quality?
On unconstrained open-ended generation, raw scale still wins. But most real workloads aren't unconstrained — they're bounded tasks with clear success criteria, where the failure mode is drift, hallucination, or unsafe output. PSE eliminates those failure modes structurally, which closes most of the practical quality gap while cutting inference cost by up to 95% in our internal benchmarks. We're happy to walk qualified partners through the methodology.
Where does my data go?
Nowhere. PSE deployments run on your hardware, on-premise, with air-gapped configurations available. No prompts, outputs, or telemetry are sent to us or any third party.
Is the framework actually verifiable?
Yes. The full theoretical framework is published openly with timestamps on Academia.edu and Zenodo — read it, cite it, try to break it. Commercial implementation details and benchmark methodology are shared under NDA with qualified partners.
What does "pathological attractor detection" mean?
Interaction systems drift into recognizable failure patterns: self-absorbed output that ignores the user, over-compliant output that agrees with anything, unstable output that oscillates, withdrawn output that under-engages. PSE models these as attractor states in the constraint geometry and detects movement toward them before it surfaces — then corrects structurally rather than filtering after the fact.
Who is behind this?
ZeravaTech Inc. is a Canadian company based in Toronto, founded by Andrej Ristić (Co-founder & Chief Researcher — mathematics and architecture) and Dhiraj Sah (Co-founder & Chief Software Engineer — implementation). The research program behind PSE spans several years of published work.
Contact
See it run on your hardware.
Request a demo, discuss a pilot, or ask about MSP partnership. We respond personally — you'll be talking to a founder, not a funnel.
info@zeravatech.caResearch inquiries: see the published papers · NDA available for commercial discussions