The Silent Council

A small business signs contracts it cannot afford to have a lawyer read. The Silent Council reads them: three AI experts that never speak to each other. A contract arrives by email. On an AMD Instinct MI300X a legal, a finance and a risk expert each read their part, and instead of writing to each other they pass on their memory: the KV cache of what they read. No summaries, nothing lost in translation. A rapporteur writes the verdict (sign, negotiate or do not sign) with exact quotes, and every disagreement goes to a person, who decides. Measured, not claimed. In a pre-registered evaluation on the CUAD contract benchmark, the same council ran in telepathy and in text: telepathy kept the quality with 75% fewer tokens, almost 3x faster (54 s instead of 144 s). What shared memory makes possible: - Long contracts: a router sends them to Gemma 4 31B, whose long context holds the whole contract, and short ones to Qwen3-14B. 77.4 F1, against 68.6 for Qwen alone. - Pages, not just text: using Gemma 4 31B's multimodality, a visual expert looks at the pages with signatures, stamps and scans, and its long context keeps those pages next to the clauses. The experts get them as KV cache, never as text. A missing signature lands in the email instead of being buried in a PDF. - Questions the next day: the PDF is never stored, but the memory of the review stays on the GPU, pages included. Asked about details nobody had planned for, like the seal's shape or a mark next to an empty signature field, the council answered 4 of 4 right in telepathy and 1 of 4 in text, where only a written summary survives. - A council that learns: every choice the company's people make becomes a precedent at the head of the next reviews. Evolus receives the email, calls the council on the MI300X and sends the person the verdict.