Hello
Ask a panel of models, not just one. A judge fuses their answers — and the Minority Report shows exactly where they disagreed.
Ask a panel of models, not just one. A judge fuses their answers — and the Minority Report shows exactly where they disagreed.
Config over code: the panel and judge are tunable here, not baked into the app.
One OpenRouter slug per line. These run in parallel for Fusion. Slugs drift — a wrong one returns a 400 that names it, surfaced in the chat.
Fuses the panel's answers and writes the structured dissent. This is the asset you own — point it at your own fine-tune later.
Runs the panel (and single models) with OpenRouter's web plugin so answers cite real, current sources, shown under each answer. Adds per-search cost. The native "OpenRouter Fusion" option already searches on its own.
Runs a cheap 2-model scout pass first; if they substantially agree, returns a merged answer and skips the full panel + judge. Disagreement escalates to the full deliberation, reusing the scout answers.
Higher = escalate more often (more thorough, costlier). Lower = trust the scouts more (cheaper). 50–95.
Assign roles (proposer, devil's advocate, fact-checker) and run a rebuttal round where each model sees the others and revises — surfacing who changed their mind. Stronger dissent, ~2× cost. Applies when the full panel runs.
When the server has a ledger configured, an identical prompt + panel returns instantly from cache (no models re-run) and an anonymized corpus of where models disagree accumulates. The prompt is stored only as a hash. No-ops if the server hasn't enabled it.
The cheap model that scores scout agreement and writes the fast-path answer.
Leave blank when the server holds OPENROUTER_KEY (gated mode) — the proxy ignores any client key then. Set it only for bring-your-own-key use. Stored in this browser's localStorage; sent as a Bearer header to /api/proxy. Never hardcoded.