ExtractionReplayFidelity

How much of a run someone could set up again from what the run recorded.

There is no value here that means "run this again and get the same output", and there will not be one. A hosted model can change weights, quantization, routing, safety filtering, and system instructions under a stable model name, and none of that is visible to DICE. Temperature zero narrows the distribution and does not remove batching and floating-point nondeterminism. So the strongest value the model offers is APPROXIMATE, and it is named for what it is.

The field says what the record supports. It is not a promise about the provider, and a host replay policy — how many attempts, against which provider, with what tolerance — stays the host's.

EXPERIMENTAL. The shape may still change while extraction runs (DICE #67) land.

Entries

Link copied to clipboard

The run recorded nothing that would help set it up again. It can be counted and attributed; it cannot be re-run in any recognisable form.

Link copied to clipboard

Identities and fingerprints only. You can tell which prompt template, schema, metamodel and profile version were in play, and compare two runs by those, without being able to reconstruct the input to either.

Link copied to clipboard

Identities, fingerprints, and the requested model configuration. A host can stand up a similar run against the same declared model and settings. The output will differ, sometimes materially, and any comparison between the two is a comparison of two runs rather than a verification of one.

Types

Link copied to clipboard
object Companion

Properties

Link copied to clipboard
Link copied to clipboard

Functions

Link copied to clipboard

Returns the enum constant of this type with the specified name. The string must match exactly an identifier used to declare an enum constant in this type. (Extraneous whitespace characters are not permitted.)

Link copied to clipboard

Returns an array containing the constants of this enum type, in the order they're declared.