RFC-010 Extension: Temporal Evidence Chaining — 3. Justification (Research Basis)
AIGP Specification › RFC-010 Extension: Temporal Evidence Chaining › 3. Justification (Research Basis)
← 2. Proposed Extension: session_id Field · Section index · 4. Backward Compatibility →
3. Justification (Research Basis)
3.1 Why Single Events Are Insufficient
| Paper | Finding | Requires Temporal Data |
|---|---|---|
| Sharma et al. ICLR 2024 | Sycophancy rates of 10-60% depending on optimization | Need baseline + deviation over time |
| Stanford HAI 2026 | 28 coded behaviors in delusional spirals | Spirals are multi-session phenomena |
| Hudon & Stip JMIR 2025 | AI psychosis emerges from sustained engagement | Dose-response needs frequency measurement |
| Zaman U Michigan 2026 | Trust erodes via perceived inauthenticity | Trust is longitudinal (builds/erodes over time) |
| Arvin KDD 2025 | 30% accuracy swing from user framing | Flip rate measurable only across interactions |
| Rehani et al. 2026 | 3-factor sycophancy scale | Factor scores meaningful only in aggregate |
| Cotra 2021 | Sycophants indistinguishable from Saints in single observations | Detection requires observing PATTERNS of agreement |
3.2 What Temporal Chaining Enables
Without chaining (current RFC-010):
- “This response was sycophantic” (point observation)
- No context: was this a one-off or a pattern?
- No trajectory: is it getting worse?
- No causality: did prior interactions influence this one?
With chaining (proposed extension):
- “Sycophancy is escalating for this user over 7 days” (trend)
- “The model changed its position between sessions without new evidence” (flip detection)
- “Interaction frequency is increasing while satisfaction is declining” (dependency signal)
- “The same user’s conviction language is intensifying” (spiral detection)
3.3 The AIHR Parallel
In healthcare, a single lab result is a data point. A longitudinal patient record enables:
- Trend detection (is the condition worsening?)
- Treatment comparison (did intervention X help?)
- Adverse event patterns (side effects that emerge over time)
- Population health analytics (which demographics are at risk?)
The evidence chain is the AI equivalent. Without it, we have lab results. With it, we have a health record.
← 2. Proposed Extension: session_id Field · Section index · 4. Backward Compatibility →