Migrating to v0.25.2

Crystallize v0.25.2 tightens reproducibility and injection rules. Use this guide to adapt existing experiments.

Strict Parameter Injection (Breaking Change)

• Old behaviour: Missing parameters in step signatures were silently filled with None or ignored.
• New behaviour: Missing required parameters raise TypeError at runtime.

Fix: Ensure every required parameter is supplied by config.yaml or has a default in Python.

# Before: crashes if 'factor' is not in the context/config
@pipeline_step()
def scale(data, ctx: FrozenContext, factor: float) -> list[float]:
    return [x * factor for x in data]

# After: provide a default or add 'factor' to config.yaml treatments
@pipeline_step()
def scale(data, ctx: FrozenContext, factor: float = 1.0) -> list[float]:
    return [x * factor for x in data]

In YAML, add missing params under treatments or the datasource configuration:

treatments:
  tuned:
    factor: 1.2

Deterministic Seeding

Seeding now uses (master_seed + replicate_id * 31337) % 2**32 instead of Python’s hash(). Seeds are stable across OSs and between serial/process execution.

• Impact: Numerical results may differ from v0.25.1 if you relied on the old hash-based seeds.
• Action: Rerun key experiments to refresh baselines; no code changes required.

CLI Updates

• Threaded execution is flagged as non-reproducible when SeedPlugin is active; use executor_type="process" for deterministic runs.
• The CLI includes an LLM Data tab with XML summaries for downstream analysis.