Nobody told the system to expect this, and nobody told the client to switch languages. A real case simply arrived in whatever language was fastest at each step — and 3Dogs Nexus followed every word of it, straight through to one coherent, decisive recommendation, delivered in English.
Nothing about the intake was staged for language. The decision brief was simply written in Greek and Mandarin Chinese. When 3Dogs Nexus's Discovery layer came back with clarifying questions — in English, as it always does — the answers arrived in whatever language was at hand.
The decision title itself was written in both scripts at once — Greek for the title's first half, Mandarin Chinese for the second — with the full decision narrative in Greek: the €12B+ investment, the multi-country consortium, the case for and against moving fast.
Before any analysis ran, the system asked for three things it couldn't responsibly assume: whether the board's mandate was truly binding, the consortium's capital and liquidity limits, and what success actually meant beyond IRR/NPV.
The governance and capital-constraint questions came back in Pitjantjatjara — a Western Desert language spoken by the Aṅangu people of Central Australia. No warning, no note that a different language was coming.
The remaining question — the specific, binding success metrics — came back in French: hard numbers on import dependence, emissions, mineral-sourcing diversity, and uptime.
3Dogs Nexus keeps a provenance trail on every field it extracts: the clean English statement that reached the 13-model debate, paired with the exact original-language sentence it was built from. Below are four of those pairs, straight out of the internal mission brief for this case — unedited, in the original scripts.
The same numbers the client wrote in French — 40%, 60%, 75%, 99.5% — land intact, correctly attached to the right criteria, in the English brief that thirteen AI models actually debated. That's the difference between a system that echoes text back and one that reads it.
Once the brief was assembled, the case ran exactly like any other: a 12-analyst panel across 13 AI models researched, argued, and reached a position. Every seat held its ground through challenge — not because the debate was shallow, but because the case for a phased build-out was, in this instance, genuinely strong across every angle the panel checked.
Twelve independent models — Llama 4, Nova Pro, Nova Lite, Nova 2 Lite, Nemotron, Qwen3, Gemma 3 27B, Qwen3-235B, OpenAI OSS, Mistral, Kimi K2, and GLM-5 — each argued from a different seat (Devil's Advocate, Risk Officer, Systems Modeler, 20-Year Scenario Planner, and more), and every one held its position after being challenged. Not a rubber stamp: the report names the option the panel rejected and why, the strongest argument against its own call, and exactly what evidence would flip the recommendation.
Case 2026-0070, exactly as delivered: the decisive call on page one, the conditions checklist, the trade-off, the numbers, and the full panel vote. 201 API calls · 13 AI models · 5m 58s — from a brief that arrived in four languages.