ALPHA FORGE 3D map: Gemini, Claude and 4 ML models split the work

The article is written from the perspective of someone running ALPHA FORGE as an individual project, and the stated motivation is that static architecture diagrams fail to show how a system behaves over time. That is why the author turned to Three.js and a real-time 3D simulation: the point is not a prettier diagram but making the sequence — data flowing in, parallel scoring, review, and nightly retraining — visible as a process rather than a snapshot.

The architecture described is layered by trust. Gemini is deliberately placed as an isolated front-line node that never hands raw news text to the downstream reasoning model; the four ML models take structured price data, technical indicators and Gemini's sentiment figures and produce probability scores; Claude then aggregates those numbers into three values — a simulation starting price, a scenario invalidation level and an expected range ceiling — with server-side validation forcing those values into the correct order. Alongside this sit two reliability mechanisms: an append-only prediction ledger in SQLite that freezes the feature snapshot at the moment of prediction and settles outcomes after 1–3 business days and 20–60 business days, and isotonic regression that corrects model confidence toward measured win rates.

How well the design holds up is likely to hinge on the promotion gate rather than the prediction step, since the author's own framing is that nightly continual learning and model replacement are the genuinely hard part. Whether the three conditions — a clear holdout win over the champion, no worsening calibration error, and non-degrading recent paper results — prove sufficient to keep a drifting model out is the open question, and the article presents itself as a personal technical log rather than investment advice or a guarantee of future performance.