RFC-010 Extension: Temporal Evidence Chaining — 1. Problem Statement
AIGP Specification › RFC-010 Extension: Temporal Evidence Chaining › 1. Problem Statement
Section index · 2. Proposed Extension: session_id Field →
RFC-010 Extension: Temporal Evidence Chaining
PRIVATE AND PROPRIETARY — NOT A PUBLIC RFC. Owned by Kanjani AI Research & Causum. See NOTICE.md.
Status: DRAFT
Authors: Kanjani AI Research & Causum
Date: 2026-06-07
1. Problem Statement
RFC-010 defines the AIGP protocol for recording individual AI invocations (REQUEST, RECORD, TRACE). Each invocation is treated as an isolated event — a single point in time with no relationship to prior or subsequent invocations.
However, the phenomena that determine whether AI systems are safe, trustworthy, and aligned are temporal:
- Sycophancy escalation emerges across multiple interactions (Sharma et al., ICLR 2024)
- Delusional spirals develop over sessions (Stanford HAI 2026; Hudon & Stip, JMIR 2025)
- Trust erosion occurs as users perceive inauthenticity over time (Zaman, U Michigan 2026)
- Model drift is only detectable by comparing current behavior to historical baseline
- Dose-response relationships require measuring frequency of exposure (AI psychosis research)
A single RECORD cannot detect these. The protocol must support events across space and time.