Retrocausal Influence and Agentic Consensus: Theoretically Grounded Interpretability for Long-Horizon Autonomous Space Systems

Srikanth Vemula

Keywords: Retrocausal interpretability, Multi-agent verification, Space robotics, Human–AI teaming, Long-horizon autonomy

Abstract:

This work proposes Retrocausal Proximity-Guided Attention (RPGA), a trajectory-level attribution scheme that back-traces long-horizon decisions in autonomous space systems to sparse, counterfactually salient subsets of past inputs. RPGA defines time-weighted retrocausal influence scores using admissible, physics-respecting edits and selects a small “retroset” of decisive timesteps. Here, “retrocausal” is operational and model-relative: influence is defined via counterfactual edits within a fixed policy and edit-oracle, rather than full causal identification of the external environment. We further introduce Explanation Markets, a multi-agent protocol
(Explainer–Skeptic–Auditor) that aggregates rationales under proper scoring rules, yielding consensus explanations and calibrated risk bands via conformal prediction. Under standard Lipschitz, mixing, and smoothness assumptions, we prove conditional faithfulness guarantees for RPGA and no-regret consensus plus coverage guarantees for the market. Synthetic but mission-inspired experiments in lunar-style rover navigation, spacecraft systems health, and mission-ops dialog show improved error interception and reduced contradiction rates compared to attribution and LLM baselines; these should be viewed as early evidence on stylized tasks
rather than exhaustive validation.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 3