Watch the engine run end-to-end, the IC memo synthesise from cited evidence, and the verdict swing as projects do. Silent recordings — read the captions, scroll at your own pace.
Click any of the four sample projects. The engine fires four data calls in parallel — Indigenous-territory overlap via Native Land Digital, English-language adverse news via GDELT and Google News, active litigation via the Sabin Center, and NGO complaints from the curated ledger. As each returns, its panel fills in. Country-level Environmental + Governance + the FPIC procedural checklist fire alongside. Data fetch completes in ~3s; the LLM-synthesized IC-memo paragraph then takes another 5–25s.
Scroll past the seven signal panels. Beyond the four social-license signals sit the dark-band Environmental + Governance pillars + the FPIC procedural checklist. At the bottom, an LLM drafts the Safeguards section of an IC memo from the cited evidence above, under strict anti-hallucination rules. Every name, number, and date in the output is checked back against the evidence JSON; ungrounded tokens are flagged.
Click the quick-compare pill bar. Cordillera Azul (Peru, VCS 985) returns 🔴 HIGH — three Indigenous nations, active court case, two NGO reports. Click 🟢 Mikoko Pamoja (Kenya, Plan Vivo) and the result swaps to LOW — no Indigenous overlap, no signal in the 24-month news window. Same engine, same query, same UI — the verdict is entirely driven by the evidence the engine surfaces.
Four parallel data calls fan out, FPIC + E + G + composite roll up by transparent rule, every panel cites its source.
An LLM drafts a two-paragraph IC-memo Safeguards section from cited evidence, under strict anti-hallucination rules and a grounding sniff.
Same engine, opposite verdicts on Cordillera Azul vs Mikoko Pamoja — driven entirely by which evidence the engine surfaces, not by per-project tuning.