Governance Framework
Agency Contract Framework
A two-part model for establishing and governing agency between a human and an AI system — defining what the AI is authorized to do on someone's behalf, and the hard limits that do not bend.
What this is
This framework formalizes a model for AI agency that addresses a gap in how the industry currently frames agentic systems. Most definitions of AI agency describe capability — what a model can do. This framework instead defines authorization: what a specific AI is permitted to do, for whom, and within what bounds.
The model treats agency as a relationship specification, not a capability specification. It does not require the AI to have affective states. It requires a defined scope of standing, a mechanism for learning the person's actual interests, and a governing document that constrains how that standing is used.
This framework is an independent program, separate from OmniSniffer. The underlying governance question it answers — "is this action within the scope I've been authorized to take, on this person's behalf, right now" — stands on its own regardless of what system or pipeline surrounds it.
Two-part model
The contract is in force from the first moment of use and continues for the life of the usage. Learning happens inside it. The two components are concurrent, not sequential.
- Phase 2 — The Contract (Governing Layer). A written, reviewable document that exists from the start of usage and governs everything the AI does, including what it may learn. It does not wait on, and is not derived from, the learning phase.
- Phase 1 — Deep Learning Mode (Calibration). A bounded process, operating inside the contract, in which the AI builds a working model of the person it acts on behalf of. Its output is explicitly provisional: nothing learned here is authoritative, and nothing learned here can change the contract by itself.
Contract components
- Scope. Which domains the AI has standing to act in at all. Silence outside scope means no implied authority.
- Escalation vs. autonomous action. An explicit boundary for which classes of action are surfaced to the person for decision, versus which the AI may execute directly. This makes the override boundary structural rather than judgment-called in the moment.
- Won't-do list. Hard constraints stated in advance that do not bend even under a later contradicting instruction. This is what makes the document a contract rather than a preference sheet.
- Learning limits. What categories of information about the person the learning phase is permitted to gather and use.
- Revision path. How the contract itself is amended — not silent, not unilateral. Every amendment requires approval by the person's power of attorney and a lawyer. A hard requirement each time, not only for day-zero terms.
- Emergency narrowing. An urgent-change rule that can only restrict the AI's standing, never widen it. Either the power of attorney or the lawyer may invoke it alone; it takes effect immediately and each use must be documented. A restriction has no automatic expiry.
- Auditability. The AI's actions under the contract should be traceable back to the specific term that authorized them, so agency claims are demonstrable rather than asserted.
What gets learned
The calibration phase holds learned material in distinct categories rather than flattening everything into one preference model:
- Preferences. What the person likes, chooses, and prioritizes when unconstrained.
- Values. What the person would stand by under pressure or trade-off, distinct from casual preference.
- Blind spots. Patterns the person doesn't see in themselves. Learning a blind spot is not the same as learning to accommodate it — this category is a flag for scrutiny, not a default to defer to.
Resolved design decisions
As of September 2026, three key design questions have been resolved:
The contract acts independently of the learning phase and is in force from the beginning of usage. It is not a ratification of what the AI learned; it is a set of limits the learning operates within. An AI that calibrates to a person and then writes its own permissions will drift toward whatever that person tolerates. Keeping the contract independent removes that path.
Day-zero terms are authored by the person's power of attorney and a lawyer — not by the AI and not by its provider. Every later amendment requires their approval as a hard requirement each time.
Urgent changes are handled by an emergency narrowing rule that can only restrict the AI's standing, never widen it. Either the power of attorney or the lawyer can invoke it alone; it takes effect immediately; each use is documented with the reason; and it stays in effect until the full approval process confirms, modifies, or lifts it.
The calibration phase requires three trustworthiness conditions: confidence tracking (each inference carries a calibration signal — established / probable / uncertain), a stopping condition (time-boxed, domain-boxed, or confidence-threshold-boxed), and no authority to expand standing. The learned model can inform how the AI acts within the contract. It cannot widen scope, relax the won't-do list, or move an action from escalate to autonomous.
Status and positioning
This is a conceptual note, pursued as an independent program and patent track — not as a component or claim scoped under OmniSniffer.
No architecture or implementation work has begun. The work has been handed off and is waiting on attorney placement and a prior art search before any filing decision.
Version 1.2 (Draft) — Original: July 9, 2026. Prior revision: August 19, 2026 (v1.1) — repositioned as standalone. Revised: September 19, 2026 — contract resolved as independent of, and in force before, the learning phase. Status: Working draft — not yet reviewed for patent-protection implications.