Determinism and reproducibility

The purpose of determinism is not to remove variation. It is to make variation replayable, so that an experiment can be repeated and a bad batch can be reached again.

Seed streams#

NameFrame derives independent deterministic seeds from one master seed. The current build defines 5 streams:

scene
spawn
camera
environment
sensor

They are derived rather than stored, which means two runs with the same master seed and scenario version get the same streams, and changing one dimension does not disturb the others. A capture that varies the weather while holding the population constant is a real experiment rather than two runs that happened to differ.

Seeds are derived per sample as well as per run, so frame 400 of a run is reachable without replaying the 399 before it.

Canonical hashes#

Configuration and manifests are serialised canonically and hashed. The identities worth knowing:

• scenario hash — the public configuration as written;
• generation-plan hash — the plan derived from it;
• CaptureJob hash — what the runtime was actually given;
• configuration snapshot hash — the whole run contract, stored beside the frames and re-verified afterwards;
• manifest hash — per frame, over the placement record.

These are how a delivered dataset can be checked months later against the report that describes it.

What determinism does and does not cover#

This distinction matters more than the guarantee itself.

Covered. Offline label derivation is byte-stable for the same dump: the verifier relabels frames and compares them byte for byte on every capture. Placement decisions, configuration and planning hashes are canonical. Manifests reconcile against the run index.

Not covered. Re-rendering the same scene on different hardware does not guarantee identical pixels. GPUs, drivers and engine builds are not bit-reproducible across machines, and claiming otherwise would be a lie about how rendering works.

So the honest statement is: the decisions are reproducible, the metadata is reproducible, and everything derived from a dump is reproducible. The dump itself is reproducible on the same machine and engine build.