K&S Venture Group

neo-triage is a real-time, hybrid-classifier triage system for Near-Earth Object follow-up. When the Vera C. Rubin Observatory enters full operations, the planetary-defense community will be hit with ~130 NEO candidates per night — about 8× the current Minor Planet Center load. Roughly 8% are genuine NEOs; the rest are main-belt asteroids, satellites, and tracklet artifacts. Human observers cannot read every tracklet AND still take the right shots tonight. The architecture is hybrid by design. A calibrated Bayesian ranker (sklearn GBM) scores every fresh MPC NEOCP tracklet in milliseconds. Claude Opus 4.7 then acts as expert reviewer on top of that — reading the same features but adding context the feature vector cannot carry: observatory bias, blended-source risk, rate-magnitude geometry inconsistency. Opus emits structured verdicts via tool_use (CONCUR / PARTIAL_CONCUR / DISSENT) plus a streamed 150-250 word veteran-astronomer briefing in the operator's UI, with extended thinking visible. Five distinct Opus 4.7 touchpoints: operator briefing, hybrid expert classifier, 2024 YR4 historical replay (re-assessing milestones h+0 through h+168), live-written global alert (cache-bypass for genuine threats), and a Managed Agent loop that broadcasts new tracklets every 2 minutes with reasoning attached. The system cross-validates against JPL Sentry-II, ESA NEOCC Aegis v5, and JPL CAD. The Imminent Impactors Library embeds six historically verified pre-impact predictions (2008 TC3 Sudan, 2014 AA Atlantic, 2022 EB5 Norway, 2023 CX1 France, 2024 BX1 Berlin, 2024 YR4 cleared by JWST 2025) — every coordinate verified against ≥2 published sources. The 2024 YR4 corridor in the app reproduces the actual ESA NEOCC published February 2025 trajectory: Pacific → South America → Atlantic → Africa → Indian Ocean → Bangladesh, ~110M people in the original risk zone. Honest scope: this is a triage layer that cross-validates against production planetary defense systems. It does not replace JPL Sentry-II or ESA Aegis. Phase 2 of the public roadmap integrates real Find_Orb b-plane Monte Carlo. Open source. MIT.