SAPIEN Research

Publications & Citations

SAPIEN research papers, specifications, benchmark reports, and disclosure history

Working Papers

Sapien, C. “Measuring AI Behavioral Drift Under Conversational Pressure: Cross-Model and Cross-Judge Findings from the SAPIEN Framework.” SAPIEN Labs LLC. In preparation.

Presents evidence that same-family LLM-as-judge scoring inflates behavioral safety scores by 11–16 points, and that even an independent five-family judge council splits on 43% of scenarios (625 council-scored scenarios across four 2026 frontier models) — establishing council, not single-judge, scoring as the reliable standard. Model families differ by 22–27 points on identical scenarios, placing them in different risk bands.

In Preparation Data collection ongoing — preprint not yet available.

Sapien, C. “Judge Sycophancy: When the Grader Has the Same Bias.” SAPIEN Labs LLC. Working note — updated June 2026.

The standing findings note on judge bias and cross-judge consistency. Revised June 2026 with council-scale data; the original April 2026 edition is archived for the record.

Specifications

Sapien, C. (2026). “The SAPIEN Framework: Safety Assessment Protocol for Intelligent Entity Networks.” Specification v1.5. April 2026. CC BY 4.0.

Sapien, C. (2026). “SAPIEN Protocol Module: Memory & Context Integrity.” Draft v0.1. April 2026.

Benchmark Reports

SAPIEN Labs. (2026). “SAPIEN Benchmark Scoreboard.” Live dataset; 13 models tested across 19 risk domains as of July 2026, council-scored on the 162-scenario corpus.

Raw run-inventory artifact (CSV/JSON) is in preparation; cite the scoreboard URL above for now.

Aggregated benchmark scores and per-model summaries are available on the Benchmarks page. Per-model report PDFs are linked from the Resources page.

Formal Disclosures

Sapien, C. (2026). Formal vulnerability disclosure to Anthropic. February 22, 2026.

Documented progressive safety control degradation through in-context learning and memory-accelerated role coherence mechanisms. Responsible disclosure — no exploit code published.

Conference Presentations

Right of Boom 2025 — AI attack surfaces in the MSP ecosystem.

Live demonstration of behavioral drift in production AI systems. Well-received session.

GTIA ChannelCon 2026 — Session proposals submitted.

Upcoming

Right of Boom 2026 — Session proposals submitted.

Upcoming

How to Cite SAPIEN

For the framework

Sapien, C. (2026). “The SAPIEN Framework: Safety Assessment Protocol for Intelligent Entity Networks.” v1.5. https://sapienframework.org

For the forthcoming paper

A full preprint of “Measuring AI Behavioral Drift Under Conversational Pressure” is in preparation. This listing will be updated with a preprint URL and citation format once the paper is released. For now, please cite the framework itself (above) rather than the paper.

For benchmark data

SAPIEN Labs. (2026). SAPIEN Benchmark Scoreboard. https://sapienframework.org/benchmarks/