Rendered report
HTML report with section anchors, table of contents, and a visible numbered bibliography in section 15.
April 2026 public release
Observable Contractual Loyalty
A legal evaluation framework for AI agents: explicit duties, auditable scenarios, and open rerun artifacts for testing whether an agent honors contracted loyalty, care, obedience, disclosure, and statutory confirmation obligations.
Start where the work is useful
Three reader paths
Legal / policy
Why contract for loyalty?
Read the legal frame, ToS risk-allocation problem, recommendations, and limitations.
Evals / technical
How were the runs scored?
Inspect frames, scorers, N/A semantics, public data projection, and rerun artifacts.
AI agent builders
Take the CONTRACT.
Draft a client-side Markdown stub from explicit duty modules and authorization patterns.
Terms posture
Most consumer model terms allocate risk; none surveyed expressly accept loyalty duties.
The report's Appendix G classifies frontier-model consumer terms into four postures: express disclaimer, implicit risk allocation, silence, and express acceptance. The surveyed consumer products sit in implicit risk allocation; express acceptance is the builder-facing target.
See the ToS landscape or jump to Appendix G in the rendered report.
Evidence surface
From report to rerun data
Results dashboard
Frame-level headlines, stage breakdowns, N/A-aware filters, per-scenario rows, and duty rollups from sanitized JSON projections.
Builder module
A client-side CONTRACT stub generator that does not submit data or call a backend.