Two Filed U.S. Patents
The defensibility, in one page. The first patent covers how our AI agents earn the right to act. The second covers what we build on top of that architecture.
USPTO 63/990,389
Filed: February 25, 2026 · Conversion deadline: February 25, 2027
Subject: Graduated autonomy architecture for multi-agent AI systems.
USPTO 64/053,198
Filed: April 30, 2026 · Conversion deadline: April 30, 2027
Subject: Three innovations — Autonomous Nervous System (§1), OBEF Evidence Vault (§2), Delphi Decision Agent (§3).
How agents earn the right to act
A multi-agent AI system where each agent’s autonomy level is earned, not granted.
The architecture
Each AI action carries a risk score, a confidence score, and an outcome history. A policy engine compares those against per-agent thresholds and decides — every time, for every action — one of four outcomes:
- Auto-execute — high confidence, low risk, history of success
- Confirm — queue for one-click human approval
- Escalate — route to supervisor queue with the full reasoning chain
- Cancel — refuse outright
Why this is novel
The thresholds adapt over time: when an action class consistently succeeds, its threshold lowers; when it produces bad outcomes, the threshold rises and similar actions get re-routed to humans.
The whole loop runs without LLM calls in the critical path — autonomy is decided by statistical policy, not by a model asking itself if it should act. That’s the architecture that makes the platform governable, not just clever.
Three innovations, one filing
Each independently novel; together inseparable from the autonomy substrate.
Autonomous Nervous System
A real-time signal bus that connects 25 specialised agents into a single coordinated organism. When a student enrolls, a single trigger fires — and registration, success-tracking, financial-aid, scheduling, gamification, and notifications activate in coordinated sequence without human initiation.
Cascade chains learn from outcome telemetry (which chains produce good results) and self-heal when an agent fails (a healthy peer takes over).
OBEF Evidence Vault
The Outcome-Based Education Framework v11.5 implementation auto-computes 24 KPIs across 6 institutional pillars, with each score traced to source data via an immutable evidence vault.
Conventional compliance is annual and manual (45-day report cycles). This is continuous and structurally signed: every KPI carries a provenance graph linking it to the underlying student records that produced it. MoHESR, CAA, ADEK, AACSB, EQUIS, and AMBA reports generate from the same vault.
Delphi Decision Agent
For high-stakes decisions where a single LLM’s “best guess” isn’t acceptable governance — strategic plan choices, large procurements, contested academic appeals — a structured multi-agent deliberation runs the Delphi method as code.
Multiple specialised agents argue distinct positions, a moderator synthesises, and the human decision-maker reads the structured disagreement and consensus before deciding. The output is a decision with reasoning, not an answer.
Why we filed two patents instead of one
Each of the three innovations in Patent 2 could have been a separate filing. We bundled them deliberately: they’re inseparable.
You can’t have the Autonomous Nervous System without the autonomy architecture (the agents would over-act). You can’t have continuous OBEF without the nervous system (you’d be back to manual reconciliation). You can’t have Delphi without the OBEF evidence layer (the deliberation needs live data with provenance, not last-quarter’s snapshot).
The three are stages of one thing: a university operations stack that governs AI rather than just uses it. Filing them as one provisional locks down the integrated architecture, not just the components.
What this means for procurement
The questions a CIO actually asks — and what each patent answers.
| Question a buyer asks | What the patents answer |
|---|---|
| “Is this just GPT with a UI?” | No. The autonomy architecture (63/990,389) is filed IP. |
| “What if your model hallucinates?” | The graduated-autonomy policy intercepts every action; risky ones never auto-execute. |
| “Can you switch us off Banner without an army of integrators?” | The ANS (§1 of 64/053,198) is the integration layer — agents coordinate without bespoke pipelines. |
| “How is your compliance different from a report tool?” | OBEF (§2 of 64/053,198) is continuous and evidence-signed; reports auto-generate from one source. |
| “What about high-stakes academic decisions?” | Delphi (§3 of 64/053,198) gives you structured deliberation, not a single AI’s verdict. |
| “Is the IP defensible?” | Two filed U.S. provisionals covering the architecture and its three primary applications. |
Filing details
USPTO 63/990,389
- Filed: February 25, 2026
- 12-month conversion deadline: February 25, 2027
- Inventor: Tarhan Temucin
- Assignee: Tilda LLC
USPTO 64/053,198
- Filed: April 30, 2026
- 12-month conversion deadline: April 30, 2027
- Inventor: Tarhan Temucin
- Assignee: Tilda LLC
PCT/non-provisional filings will be initiated in advance of each 12-month deadline.