Coppelius's launch graph, derived: every launch of a training run
(Coppelius.TrainingRun: initialize or load, predict, the update the run
issues once per step, the final forward, save) computed from the model's
shape
(Coppelius.Model), the learner's hyperparameters (Coppelius.Learner),
the device images (Coppelius.Build.DeviceImages) and the staging window's
layout (Coppelius.TrainingRun) -- no address, count or constant is
written down that one of those determines.
The graph is the one the legacy generator emitted
(systems/coppelius/reference/oracles/legacy-generated/tools/gen_train.py,
`unified10`), from which the plan was once imported as a recording; it
was re-derived here launch for launch, and the realized tables were
byte-identical to the recording's (the migration, 2026-09-24). Changes
since: `loss.correct.zero` is the target-logit gather, ten loss-record
copies close the updates (docs/observability PRD item 1); and the
workspace is a named layout per phase, certified on the device side, which
moved the tensors the recording aliased while live -- the cross-entropy
backward wrote the logits gradient over the loss rows it was reading, and
two backward phases reused a buffer they still held -- and the output's
and two feed-forward phases' planes with them.
a 32-bit slot of a parameter block, by its word index (byte offset / 4)
56field unrestricted cgSlotValue : NatThe compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.