`reproducibility_mode`

Back to L0 | Browse all axes | Browse all options

Axis reproducibility_mode on sub-layer l0_a (layer l0).

Sub-layer

l0_a

Axis metadata

Default: 'seeded_reproducible'
Sweepable: False
Status: operational
Leaf-config keys: random_seed

Operational status summary

Operational: 2 option(s)
Future: 0 option(s)

Options

`seeded_reproducible` – operational

Fix a deterministic seed and propagate it through every RNG.

The cell-loop reads leaf_config.random_seed (default 0) and applies it to random.seed, numpy.random.seed, torch.manual_seed (when torch is installed), and the PYTHONHASHSEED environment variable for the current process. Each L4 estimator inherits the seed via its params.random_state key (issue #215); per-fit-node random_state overrides the L0 seed when present.

This is the only mode under which macroforecast.replicate(manifest) can verify bit-exact sink hashes. Use it for every study you intend to publish, share, or compare against later.

When to use

Default. Required for any study where bit-exact replication matters (papers, internal benchmarks, regression tests, comparisons across package versions).

When NOT to use

Stochastic exploration where the explicit goal is to characterise the variability of a procedure across seeds. Pick exploratory and re-run; the manifest will record that the seed was unfixed.

References

Stodden, McNutt, Bailey et al. (2016) ‘Enhancing reproducibility for computational methods’, Science 354(6317). (doi:10.1126/science.aah6168)
macroforecast PR #215 – L0 random_seed propagation into L4 random_state.

Related options: exploratory

Examples

Reproducible study (default)

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

Last reviewed 2026-05-04 by macroforecast author.

`exploratory` – operational

Do not fix stochastic seeds; each run draws fresh randomness.

Skips the global RNG seeding that seeded_reproducible performs. Each cell pulls its own randomness from the OS entropy pool; downstream estimators that take an explicit random_state still use whatever the recipe sets per node, but the L0 default of 0 is not propagated.

replicate() cannot guarantee bit-exact sink hashes under this mode – the recipe still re-runs and produces structurally identical manifests, but the numeric forecasts will differ run-over-run.

When to use

Sensitivity studies where you want to measure how much variability the random components introduce. Wrap the run in a sweep over several executions and compare the spread.

When NOT to use

Anything you intend to publish or share. The reviewer will not be able to reproduce your manifest.

References

macroforecast design Part 1, L0 §A: ‘exploratory mode trades reproducibility for unbiased sampling of stochastic variability.’

Related options: seeded_reproducible

Examples

Sensitivity sweep over fresh seeds

0_meta:
  fixed_axes:
    reproducibility_mode: exploratory