Version-1 seed derivation

Metewand derives independent dataset, implementation, and scheduling seeds from each experiment's authored safe-integer seed. Every derivation returns a DerivedSeedRecord containing both the nonnegative 53-bit value used by workers and the complete SHA-256 digest retained for identity and provenance records.

Transcript

For role dataset, implementation, or scheduling, the digest is:

SHA-256("metewand-seed-v1" || NUL || UTF-8(role) || NUL || canonical_fields)

canonical_fields is RFC 8785 JSON in Metewand's restricted version-1 value domain. It has no whitespace or trailing newline. The three field objects are:

Role	Canonical fields
Dataset	dataset_definition, dataset_parameters, experiment_seed
Implementation	dataset_configuration, experiment_seed, implementation_repetition, problem_configuration
Scheduling	benchmark_name, experiment_name, experiment_seed, scheduling_policy_version

Identity fields contain the complete typed identity string. Dataset parameters are the resolved canonical object after defaults and validation. Integer fields must remain within the safe JSON range. Object properties appear in canonical order in the transcript, independent of their input order.

SEED_DERIVATION_VERSION is 1 and selects the metewand-seed-v1 domain. SCHEDULING_POLICY_VERSION is also 1 and is included as a field of every current scheduling derivation. Changing either contract requires an explicit compatibility review and new golden vectors.

Exposed value

The worker-visible value is the first 53 bits of the derivation digest, interpreted most-significant bit first. Equivalently, implementations read the first eight digest bytes as a big-endian u64 and shift right by 11. This produces a nonnegative integer no greater than 2^53 - 1, so every supported language and JSON implementation can represent it exactly.

The remaining digest bits are not discarded from the record. The full digest distinguishes derivations that happen to share the same exposed value and binds the complete derivation into configuration and logical-specification identities.

Dependency boundaries

Dataset seeds change with the experiment seed, dataset-definition identity, or resolved dataset parameters. Implementation seeds change with the experiment seed, dataset-configuration identity, problem-configuration identity, or zero-based implementation-repetition index. They deliberately omit the implementation identity, which gives every implementation in the same problem and replication block the same value.

Scheduling seeds bind the benchmark and experiment names and the scheduling policy version. They omit the expanded member set, so adding an unrelated candidate does not perturb the scheduling seed. Neither derivation depends on matrix-expansion order.

Rust API and conformance vectors

metewand_core::seed exposes one derivation function and one canonical-field function for each role. The canonical-field functions make the exact hashed bytes inspectable by conformance tools.

fixtures/seeds/v1.json fixes the canonical field bytes, complete digest, and exposed value for all three roles. Rust, Python, and R consume this file directly. The SDK test harnesses reconstruct the transcripts without adding seed derivation to their worker APIs. Future SDKs must use the same vectors. See SDK conformance.