Layer 0 Axis: `reproducibility_mode`

Parent: Layer 0
Current: reproducibility_mode

reproducibility_mode controls seed policy for the run. It determines whether the execution layer resets stochastic state before fitting and forecasting.

Values

Value	Status	Meaning
`seeded_reproducible`	operational, default	apply a fixed base seed
`exploratory`	operational	do not reset global RNG state

This axis is not sweepable. Reproducibility policy is part of study setup, not a model alternative.

Leaf Config

Key	Applies When	Rule
`random_seed`	`seeded_reproducible`	optional int; default `42`
`random_seed`	`exploratory`	forbidden
`gpu_deterministic`	any mode	optional bool; default `false`

gpu_deterministic is separate because hardware/library deterministic behavior is not the same as seed policy.

Runtime Semantics

Mode	Runtime Behavior
`seeded_reproducible`	sets Python, NumPy, and torch seeds where available
`exploratory`	leaves existing global RNG state alone and records that reproducibility was waived

seeded_reproducible is intended for reproducible empirical work. It does not guarantee bit-identical results across different hardware, BLAS libraries, torch/cuda versions, or nondeterministic algorithms unless the relevant runtime stack also supports deterministic execution.

YAML

Default seeded run:

0_meta:
  fixed_axes:
    reproducibility_mode: seeded_reproducible

Seeded run with explicit seed:

0_meta:
  fixed_axes:
    reproducibility_mode: seeded_reproducible
  leaf_config:
    random_seed: 20260101

Exploratory run:

0_meta:
  fixed_axes:
    reproducibility_mode: exploratory

GPU deterministic request:

0_meta:
  fixed_axes:
    reproducibility_mode: seeded_reproducible
  leaf_config:
    random_seed: 42
    gpu_deterministic: true

Invalid Patterns

Invalid Pattern	Use Instead
`reproducibility_mode: strict`	`seeded_reproducible` plus deterministic leaf config where supported
`reproducibility_mode: strict_reproducible` in new recipes	`seeded_reproducible`
`exploratory` with `random_seed`	remove `random_seed` or use `seeded_reproducible`
sweeping seed policy	keep fixed; sweep scientific axes only

Notes

Default seed is 42.
Set a project-specific seed when results must be replayed exactly within the same environment.
Use exploratory only when reproducibility is intentionally waived.

Layer 0 Axis: reproducibility_mode