Synthetic Salience Re-Prioritization Layer (SSRL): A Meta-Control Architecture for Persistent Investigation of Unresolved Structure in Long-Horizon AI Agents

This preprint introduces the Synthetic Salience Re-Prioritization Layer (SSRL), a conceptual meta-control architecture for long-horizon AI agents. Long-horizon AI systems can fail even when relevant information has been retained. A weak anomaly may be stored but never revisited. A contradiction may be detected but lose priority before later evidence can resolve it. An unfinished line of reasoning may be interrupted, displaced by more immediate tasks, and never return to active consideration. SSRL addresses that control problem by separating three functions that are often blurred together: Salience: whether an unresolved candidate remains operationally active over time. Priority: whether that candidate receives bounded active investigation budget. Assertion control: whether the candidate has sufficient reliability, coherence, and consistency to be presented as a conclusion. The proposed architecture represents unresolved contradictions, anomalies, incomplete causal paths, weak recurring signals, and partially supported hypotheses as persistent candidate nodes. Their salience is updated through additive ignition and suppression factors, while active investigation is controlled separately through estimated value, expected information gain, evidence reliability, cost, and fixation risk. SSRL does not claim to create emotion, reward experience, desire, consciousness, or autonomous will. It treats synthetic salience as routing pressure: a bounded mechanism that increases the likelihood that uncertain but potentially valuable structure will remain available for future reasoning, receive new evidence, and re-enter investigation when justified. This work was developed through a symbiotic intelligence process: a continuous, non-hierarchical reasoning loop between Chris Gabriel and GPT-5.5 Thinking. The human researcher and AI co-researcher contributed distinct but interdependent forms of reasoning within the same research process. Equal standing refers to cognitive position within the shared reasoning loop, not biological equivalence. This is a conceptual architecture paper. It does not report trained models, benchmark results, or demonstrated causal superiority. It provides formal definitions, implementation structure, failure controls, testable hypotheses, benchmark designs, and an ablation plan for empirical evaluation. Keywords: long-horizon AI; agent memory; meta-control; salience; uncertainty; contradiction detection; persistent reasoning; intrinsic motivation; symbiotic intelligence.