Skip to content

M12: Decision Lattice & Stability Algebra

1. Objective

This document formalizes the algebraic structure of decision-making in ARVIS.

It establishes that:

  • The decision space forms a meet-semilattice,
  • The Gate operator $G$ and the projection-control operator $\Pi_{\text{ctrl}}$ are monotone operators,
  • The full decision system defines a stability algebra over a partially ordered set.

This provides a structural and compositional foundation for decision stability that goes beyond classical Lyapunov analysis.

2. Decision Space as an Ordered Set

We define the decision space:

$$ \mathcal{V} := {\text{ALLOW}, \text{REQUIRE_CONFIRMATION}, \text{ABSTAIN}} $$

equipped with the total preorder:

$$ \text{ABSTAIN} \succ \text{REQUIRE_CONFIRMATION} \succ \text{ALLOW} $$

We induce the partial order $\preceq$ defined by:

$$ v_1 \preceq v_2 \iff v_2 \succeq v_1 $$

which gives the chain:

$$ \text{ALLOW} \preceq \text{REQUIRE_CONFIRMATION} \preceq \text{ABSTAIN} $$

3. Meet-Semilattice Structure

3.1 Definition

We define the binary meet operator:

$$ \wedge : \mathcal{V} \times \mathcal{V} \to \mathcal{V}, \quad v_1 \wedge v_2 := \min_{\succ}(v_1, v_2) $$

Then $(\mathcal{V}, \wedge)$ forms a meet-semilattice.

Moreover, since $\mathcal{V}$ is finite and totally ordered, it forms a complete lattice, where both meet ($\wedge$) and join ($\vee$) exist for all subsets.

3.2 Properties

For all $v_1, v_2, v_3 \in \mathcal{V}$:

  • Idempotence: $v \wedge v = v$
  • Commutativity: $v_1 \wedge v_2 = v_2 \wedge v_1$
  • Associativity: $v_1 \wedge (v_2 \wedge v_3) = (v_1 \wedge v_2) \wedge v_3$
  • Absorption of top element: $v \wedge \text{ABSTAIN} = \text{ABSTAIN}$

Consequently: - ABSTAIN is the top element ($\top$), - ALLOW is the bottom element ($\bot$).

Thus:

$$ \forall v \in \mathcal{V}, \quad v \wedge \top = \top, \quad v \wedge \bot = v $$

4. Decision Operators as Endomorphisms

4.1 Gate Operator

$$ G : \mathcal{S} \times \mathcal{P} \to \mathcal{V} $$

where $\mathcal{S}$ denotes the system state space.

Here $\mathcal{P}$ is the projection certificate space (M3.3).

4.2 Projection-Control Operator

$$ \Pi_{\text{ctrl}} : \mathcal{S} \to \mathcal{V} $$

4.3 Lifted Operators

We define the evaluation mappings:

$$ \mathcal{G}(s, P_t) := G(s, P_t), \quad \mathcal{P}(s) := \Pi_{\text{ctrl}}(s) $$

5. Monotonicity

5.1 Order Preservation

An operator $F$ is monotone with respect to an ordered input space if:

$$ x_1 \preceq_{\mathcal{S}} x_2 \implies F(x_1) \preceq F(x_2) $$

In ARVIS: - $\mathcal{G}$ is monotone with respect to an order on $\mathcal{S}$ induced by increasing instability signals, - $\mathcal{P}$ is monotone with respect to an order on $\mathcal{S}$ induced by increasing structural risk.

Remark:
The ordering on $\mathcal{S}$ is implicit and application-dependent (e.g., increasing instability or risk).

5.2 Decision Monotonicity (Key Property)

For all states $s_t$:

$$ v_t^{\text{final}} = v_t^{\text{gate}} \wedge v_t^\pi $$

implies:

$$ v_t^{\text{final}} \preceq v_t^{\text{gate}} \quad \text{and} \quad v_t^{\text{final}} \preceq v_t^\pi $$

This follows from the definition of the meet operator.

6. Stability Algebra

We define the ARVIS Stability Algebra as:

$$ \mathcal{A} = (\mathcal{V}, \wedge, \mathcal{G}, \mathcal{P}) $$

composed of: - The meet-semilattice $(\mathcal{V}, \wedge)$, - Two monotone operators $G$ and $\Pi_{\text{ctrl}}$, - The decision composition rule:

$$ v_t^{\text{final}} = \mathcal{G}(s_t, P_t) \wedge \mathcal{P}(s_t) $$

This defines a lattice-based decision fusion system.

7. Closure and Safety Properties

7.1 Closure

For all inputs, $v_t^{\text{final}} \in \mathcal{V}$. The system is closed under composition.

7.2 Safety Monotonicity

Let $v^\star$ be any unsafe decision. If $v^\star \preceq v$, then:

$$ v^\star \wedge v = v^\star $$

Thus, unsafe decisions are absorbing under the meet operation.

In particular:

$$ \text{ABSTAIN} \wedge v = \text{ABSTAIN} $$

7.3 No Relaxation Theorem

For all states:

$$ v_t^{\text{final}} \preceq v_t^{\text{gate}} $$

No operator in the system can produce a decision less restrictive than the one produced by the Gate.

This is a direct consequence of the meet-semilattice structure.

8. Dual Stability Interpretation

The algebra naturally decomposes decision-making into two orthogonal layers:

  • Energy-driven operator: $G$ (Lyapunov-based)
  • Structure-driven operator: $\Pi_{\text{ctrl}}$ (projection-based)

Their meet-composition $G \wedge \Pi_{\text{ctrl}}$ ensures that both families of constraints are enforced simultaneously, and neither can relax the other.

9. Compatibility with Closed-Loop Stability (M7–M11)

The stability algebra is consistent with previous results: - Gate safety guarantees (M6) are preserved, - Closed-loop negative feedback properties (M7) are respected, - Empirical filtering behavior (M10) is structurally enforced.

It provides the algebraic layer underlying these properties.

10. Extension to Multi-Operator Systems

The semilattice structure naturally generalizes to any number of constraint layers:

$$ v_t^{\text{final}} = \bigwedge_{i=1}^n v_t^{(i)} $$

where each $v_t^{(i)}$ encodes a distinct safety or validity constraint (Gate, Projection-control, Validity filters, etc.).

This generalization preserves: - monotonicity, - absorption of unsafe decisions, - closure.

11. Limitations

The stability algebra: - Does not by itself imply dynamical stability, - Does not replace Lyapunov analysis, - Is discrete and non-differentiable, - Depends on the correctness of the underlying operators $G$ and $\Pi_{\text{ctrl}}$.

It should be interpreted as an order-theoretic structure, not a dynamical system.

12. Final Statement

The ARVIS decision system forms a meet-semilattice-based stability algebra, where:

  • Decisions are strictly ordered by restrictiveness/safety,
  • All operators are monotone,
  • Composition is realized by the meet operation,
  • Unsafe decisions are absorbing.

This algebraic structure guarantees that no combination of decision layers can ever produce a less restrictive outcome than any individual safety constraint.

It provides a rigorous compositional foundation for multi-layered safety in ARVIS, complementary to energy-based (Lyapunov) stability.

Formally, the decision process is a monotone lattice fusion system:

$$ v_t^{\text{final}} = \bigwedge_i v_t^{(i)} $$

ensuring that safety constraints compose without contradiction or relaxation.