Weight prediction (MoE daily curve)

Calls POST /api/weights/predict-daily-curve. The daily MoE line is median/mean over anchor-level weights from combined pair Δkg (ridge / equal-weight / single expert). It is not the same as applying the ridge meta-learner to each expert's daily curve.

Batch comparison

Select multiple pens (public batch_pk = Batch.batch_id). With fetch_new, each run loads training inputs for that pen only (batch_id is set per selection). By default the list includes open pens (toggle “Closed only” to restrict). Each batch runs a separate predict-daily-curve with preferred_batch_pks: [one]. With Training window: last 7 days, each request intersects the form’s rolling week with that pen’s accommodation dates.

Data source

Forward bootstrap

Only the smallest day_index row may have target_daily_weight; other days use causal Δ pairs (ir < iD) with predicted anchors. Prefer cached_json; fetch_new may not emit partial rows yet. Cannot combine with pair_budget.

Model

Filters & sampling

Must match token batch_pk = public farm batch id (Batch.batch_id), not the internal /batches/<id>/ URL id.

Response options

Database logging

Multi-batch compare assigns one compare_run_id per submit so rows group together.