Consumer World Model

下一行為概率在支援的候選類別中合計為100%，尚未經校準。購買窗口另行預測7/30/90天內至少購買一次的累計概率，不與行為類別一起正規化。行為分布未按靜默時長調整；公共先驗參考尚未通過工作區驗證。

需求分位數表達整體不確定性，不是個人購買區間，覆蓋率尚未經校準。缺失資料保持不可用；在工作區評測完成前，不宣稱精度提升。

# Bearer API key; predictions:read
GET /api/v2/cwm/capabilities
GET /api/v2/cwm/members/{id}/next-behaviors

# metrics:read additionally enables purchaseWindows in next-behaviors
# predictions:read AND metrics:read for standalone trend/windows
GET /api/v2/cwm/trends?metric=gmv&horizon_days=30
GET /api/v2/cwm/members/{id}/purchase-windows

# Read-only MCP equivalents
cwm_get_forecast_capabilities
cwm_predict_next_behaviors
cwm_forecast_aggregate_trend
cwm_predict_purchase_windows

Purchase probabilities use 0–1 units and cumulative, monotonic windows. Until trained and validated, the response contains probabilities: {"7":null,"30":null,"90":null} and calibrated: false. Existing CWM 0–100 mass scores keep their meaning. Aggregate forecasts return one point and q10..q90 per complete UTC day; missing daily buckets are not filled with zero. Category names require catalog_gmv or units_sold; draft input metrics are disclosed. Tenant and member authorization remain server-side. Readiness distinguishes service_unavailable, awaiting_training, awaiting_validation and ready. TimesFM 2.5 weights use Apache 2.0.

Next-behavior responses use predictionKind: next_event and unit: probability_0_1. One response includes distribution and purchaseWindows; individual expert results are internal. Purchase fusion is invoked automatically when metrics access permits. A validated tenant head serves cumulative 7/30/90-day probabilities; missing permission, training, validation or inputs leaves all windows null with an explicit reason. Available behavior predictions remain readable. These windows are not summed with the next-event distribution. The brand model's rows include all 48 standard events: 26 are consumer behaviors, 39 enter the sequence vocabulary, and 23 is the prediction-target ceiling. Only covered candidates supported by the active, evaluated generation receive probabilities; all other rows have probability: null and an explicit availability. Legacy input-frequency coverage is weaker corpus-presence evidence and cannot certify provenance-masked tag targets. New snapshots use positive targets from optimizer training windows, excluding validation members and masked rows. The final distribution includes rank, entropyBits, retainedModelMass, coverageBasis, modelVersion and calibrated: false. An available, evaluated brand model is used directly without requesting alternate expert outputs. Otherwise, valid public-prior distributions are equally averaged internally into one degraded pv / fav / cart / buy / rmcart / buy_disc reference. aggregation distinguishes single_tenant_model from uniform_valid_experts; vocabulary identifies the candidate key space. The two vocabularies are never blended. The uniform reference is not a learned multiclass gate and is not tenant-validated; public class-training coverage is unverified. Rounded zero mass does not prove absent class training. This endpoint needs predictions:read; purchaseWindows additionally needs metrics:read or *. Without metrics permission, the purchase component reports metrics_scope_required and makes no metric or fusion call. Status is ok only when both earned components serve, degraded for partial/reference results, and unavailable when neither serves. Store access and live processing eligibility are rechecked before personal external processing; insufficient evidence remains unavailable. One request shares its live CWM fingerprint internally across the behavior and purchase heads.