Local-only deterministic training receipt

Multi-token prediction loss lab

This lab audits shifted targets, masks, and denominators on bounded synthetic logits. It does not train a model, reproduce paper results, choose weights, or predict decoding speed.

Declared sequence

Token IDs (2–32 integers) Segment IDs (same length) Ignored target positions Horizons (1–4 values) Head weights (same length)

Head loss ledger

Horizon	Nominal T−h	Ignored	Boundary	Valid denominator	Loss sum	Mean	Weight	Contribution	Accepted / rejected

Multi-token prediction loss lab

Declared sequence

Head loss ledger

Negative fixtures

Exportable receipt