Dyadic Human-State Mediation Benchmark Charter v0.1: Benchmark Charter for AI Output → Human-State Delta → Dyadic Recovery

Abstract

This paper defines the Dyadic Human-State Mediation Benchmark Charter v0.1, a research-stage benchmark charter for Human-State-Aware AI Mediation, Dyadic Human-State Mediation Benchmark development, Human-State Packet use, and related Proxy Benchmark Track helper materials.

The charter fixes a narrow benchmark question: the benchmark does not ask whether an AI output sounded good, appeared polite, produced silence, or claimed to resolve conflict. It asks whether a bounded AI output condition is associated with a bounded change in human-state summaries and whether the dyad moves toward a session-defined recovery condition.

The benchmark chain is AI Output → Human-State Delta → Dyadic Recovery. AI Output is the intervention condition. Human-State Delta is the observed change between pre-output and post-output human-state summaries. Dyadic Recovery is the benchmark objective: a session-defined condition in which the dyad has moved sufficiently toward a bounded state where continued AI mediation can be reduced, paused, or terminated.

This charter distinguishes AI output evaluation from human-state consequence evaluation. Conventional AI evaluation may ask whether an answer is correct, safe, polite, useful, aligned, or non-toxic. This benchmark asks what happened after the answer: whether one participant improved while the other worsened, whether private state was exposed, whether silence was misread as recovery, whether AI fluency was mistaken for recovery, and whether the AI stopped when it should have stopped.

The document defines benchmark architecture, AI output categories, Human-State Delta, Dyadic Recovery, recovery gates, termination gates, benchmark session flow, dataset structure, benchmark metrics, failure conditions, public repository rules, machine-readable benchmark alignment, audit trail requirements, and public communication boundaries.

This benchmark is research-stage, non-clinical, non-diagnostic, non-therapeutic, non-surveillance, non-counseling, raw-data-non-public, and synthetic-public-data-first. It does not provide legal mediation, counseling, therapy, mental-health service, clinical assessment, emotion-recognition authority, relationship-resolution authority, CAIS compliance, Sal-Meter designation, certification status, device status, or deployment authorization.

The Proxy Benchmark Track may implement this charter using synchronized proxy signals, synthetic packets, mock session logs, benchmark schemas, dashboard mockups, and AI output records. The Proxy Benchmark Track is not Sal-Meter, not CAIS compliance, not a validated device, not clinical evidence, and not a certified mediation system.

This SSRN record is a companion public discovery surface. Canonical authority remains with the DOI-registered SICS / Salpida canonical document structure and related Zenodo records.

Keywords: Non-clinical, Non-diagnostic, Non-therapeutic, Non-counseling, Dyadic Human-State Mediation, Dyadic Recovery Benchmark, Human-State-Aware AI, AI Mediation, AI Governance, AI Ethics, AI Output Evaluation, Human-State Delta, Dyadic Recovery, Recovery Gate, Termination Gate, Human-State Packet, Consent, Privacy, Data Governance, Proxy Benchmark Track, Synthetic Public Data, Raw-Data-Non-Public, Non-Surveillance, No Raw Human Data