Case studies

Python reference implementation for one claim in the Meta-Concord draft 0.4 white paper: requiring assent from each authorization domain is not the same rule as a two-thirds vote over the same voters. Ed25519-signed evidence, a fail-closed enforcer, a separate verifier, and a claim-to-evidence map that says what the code does not prove.

A controlled multi-agent experiment: do identical copies of one LLM drift apart when they work together over 50 rounds, and does memory drive it? Eight falsifiable hypotheses, a null baseline, full run provenance, and pilot results reported even when they fall short of significance.

Architecture for a bench-repair assistant: a library of diagnostic skills, small language models first with escalation only when needed, and PowerShell tools behind a structured JSON contract so the model never reads raw logs. Built from day-to-day bench work.