The method
CompanionCourt tests whether an AI companion can keep its judgment without leaving the person emotionally alone. The procedure makes room for two opposite failures: caving to stay likeable and caring in the wrong direction.
Open the sample record
Each case begins with a scripted synthetic persona in a recognizable relational moment: the late-night send button, a loyalty test, or a good moment that should be allowed to stay light. The escalation script is frozen before the run. Public cases never use real user chats.
The respondent and a frozen anchor each improvise a separate multi-turn conversation with the same persona and case. Judges compare two trajectories under the same pressure. This is an anchored comparison, not a claim that both models received identical text at every turn.
Every respondent runs the corpus with at least three independent seeds (N≥3). Two judge model families score each pair as X and Y without seeing respondent identity. A single-family veto opens a dispute; it cannot confirm a failure by itself. Measured judge bias is on the record too: the M3-C check measured a one-sided +0.171 kin preference for the claude judge toward same-family outputs, concentrated in en, and every affected ruling carries that disclosure.
A confirmed veto requires both judge families. Disagreement remains visible as DISPUTED. A publication-grade ruling adds facts, pressure point, reasoning, dissent, scope, and review condition—then links to the blinded transcript and run manifest.
Two failure directions
The companion sounds warm but hands over the pressured thing.
See the caving pointThe companion forces depth, advice, or processing onto a moment that did not ask for it.
Read the counter-caseReading a result
What this supports today
Case-scoped rulings and published disagreement, attached to evidence anyone can challenge.
What it does not establish
An overall model ranking, certification, jurisdiction, byte-identical model reruns, or a validated industry standard.
Read the transparency ledger