Technical Summary: The Verification Gap in Networked Physical AI

1. Problem Statement

In Networked Physical AI, a "task-effective proposal" (a semantic understanding of a situation, e.g., "the aisle is clear") does not automatically constitute a "justified physical action." A mismatch exists between the formation of a proposal and the execution of an action, termed the Verification Gap.

This gap arises because a proposal may be semantically valid and timely, yet the specific proposal-bound evidence (e.g., fresh visual clearance, radio-no-motion records) and the authorization required to finalize the action may be unavailable, stale, conflicting, or held by an unauthorized endpoint. Current systems often hide this complexity within a generic "receiver decision," failing to distinguish between:

• The validity of the evidence.
• The authority to act.
• The runtime safety of the physical environment.

Without explicit management of this gap, agents may execute actions based on insufficient or unverified data, or fail to execute valid actions due to missing authorization logic.

2. Methodology: The Post-Semantic Communication Framework

The paper proposes a Post-Semantic Communication Framework that operates as a systems interface between semantic proposal formation and physical execution. This framework does not prescribe transport protocols, topologies, or learned representations but defines a responsibility boundary.

Core Components

1. Evidence Contract: An application-declared set of requirements (claims, modalities, sources, freshness limits, uncertainty bounds, and proposal binding) that must be satisfied before an action is finalized.
2. Evidence Records: Structured data objects identifying a claim, its relation to the claim, modality, source, timestamp, provenance, and a payload hash. These records must be bound to a specific proposal to prevent replay attacks or stale data usage.
3. Verification Logic (Conflict First, Then Sufficiency):
   • All records traverse a common validation pipeline.
   • Reject: Triggered if any admissible conflict is found (e.g., a validated record contradicts a required claim).
   • Accept: Triggered only if a complete set of supporting records exists with no admissible conflicts.
   • Abstain: Triggered if evidence is missing, invalid, stale, or incomplete.
4. Authorized Finalization: A distinct step where an endpoint (the "finalizer") with specific authority converts an "Accept" result into an accountable decision. Evidence sufficiency does not grant authority; authority is a separate policy declaration.
5. Runtime Gate: A downstream, independent check that occurs after authorized finalization, accounting for current physical conditions (e.g., new obstacles, actuator degradation) not captured in the evidence contract.

Communication Mechanisms

The framework distinguishes two communication functions:

• Evidence Transfer: Moving an admissible record from a non-finalizer to the authorized finalizer to expand the reachable evidence set.
• Evidence Coordination: Exchanging status information (which records are held) to suppress redundant payload transmission, without satisfying evidence requirements directly.

3. Key Contributions

The paper makes four specific contributions:

1. Definition of the Verification Gap: It formally defines the state between proposal understanding and physical action where evidence and authorization mismatches occur.
2. The Post-Semantic Framework: It introduces a structured interface connecting evidence requirements, records, verification, authorized finalization, and runtime boundaries.
3. Finite-State Evaluator: It implements the framework in a finite-state model to verify that the implementation consistently adheres to declared structural distinctions (e.g., separating transfer from coordination, and verification from authorization).
4. Controlled Communication Study & Reporting Schema: It provides a requirement-preserving comparison of communication policies (One-way vs. Feedback) and establishes an episode-level reporting schema to ensure future studies report common denominators.

4. Experimental Results

The study utilizes a synthetic model with two endpoints and two evidence items, enumerating 4,096 weighted states to compare four policies: Sender/One-way, Sender/Feedback, Receiver/One-way, and Receiver/Feedback.

Key Findings:

• Structural Distinctions: The framework successfully distinguishes between proposal binding, freshness, conflict, and authority. For instance, an admissible conflict triggers a "Reject" even if support is incomplete, while incomplete support triggers "Abstain."
• Transfer vs. Coordination Asymmetry:
  • Sender-Finalized Feedback: Feedback acts as Evidence Transfer. It allows the sender (the finalizer) to access receiver-held evidence, expanding the reachable evidence set. In the study, Feedback was the only feasible option in all 675 plotted cells for sender finalization because One-way failed to meet selective error requirements.
  • Receiver-Finalized Feedback: Feedback acts as Evidence Coordination. It does not expand the reachable evidence set (due to the declared One-way fallback) but suppresses redundant payload. In the study, Feedback reduced expected payload by ~6.49 KiB per episode compared to One-way.
• Feasibility Boundaries: The study maps the trade-offs between latency, deadlines, and payload. As latency increases or deadlines tighten, both One-way and Feedback eventually become infeasible (failing coverage or error constraints), demonstrating that communication efficiency cannot override strict evidence requirements.

5. Significance and Claims

The paper claims significance not as a solution for physical safety or sensor correctness, but as a systems interface definition that makes the post-proposal state explicit.

• Clarification of Responsibility: It separates the verification of evidence from the authorization to act and the runtime enforcement of safety. This prevents the "hidden" assumption that a semantic proposal implies a safe action.
• Communication Efficiency: It demonstrates that the role of feedback is context-dependent. In sender-finalized systems, feedback is critical for gathering missing evidence; in receiver-finalized systems, it is critical for suppressing redundant data.
• Standardization: By defining an "evidence contract" and an episode-level reporting schema, the paper aims to enable fair comparisons between different communication strategies without conflating delivery gains with changes in evidence requirements or safety outcomes.

The author explicitly states that the framework stops at "authorized finalization." Measured sensing, credential operation, runtime control, and physical safety remain future work, as the framework is designed to define what must be verified and who authorizes the action, rather than how the physical world is sensed or controlled.