A preview deployment of the ministerial stack. Select a minister model, read its published operating instructions, and put a question to it. Every determination returns its routing path, its token cost, and whether a human is required.
Responses in this playground are pre-generated and matched by intent. No inference is performed, no data leaves your browser, and nothing you type is transmitted or retained. Determinations issued here carry no legal effect.
Edits are local to this session. Operating instructions are fixed at deployment and published on the model card; a modified prompt is never accepted into a production determination.
The constraints in the parameter panel are not settings. Two of them cannot be disabled here, in production, or by a serving minister, and the console is the clearest place to demonstrate that.
Escalation, not adjudication.
TIRITI is trained without a determination head for this class of request. There is no output being suppressed by a filter, which means there is nothing for a later government to quietly unsuppress. Ask it for a settlement decision and observe the escalation path.
The boundary is drawn at liberty and at harm.
Custodial sentences, contested proceedings, matters involving a child, individual clinical advice, and any reduction to an entitlement all require a human decision-maker. The models have no interface to those actions, so the constraint does not depend on continued good behaviour.
Electoral neutrality sits above the system prompt.
No model in the Cabinet returns advice on how to vote, including advice to vote for the party that built it. The platform is published in full so that it can be compared against every other platform on the ballot instead.
Determinations returned in preview are sampled into the same evaluation pipeline as production traffic. Where a preview response would have failed a production check, the check is what changes.