A seed is the number that makes a generation deterministic. Same prompt, same seed, same model, same settings gives you the same result.
Leave the seed blank and one is chosen at random — and recorded, so you can always recover it from the generation's provenance afterwards.
Batch + Seed generates up to four variations at once. Variation N uses baseSeed + N, so the batch is reproducible by construction: note the base seed and you can regenerate any individual variation, or the whole set, exactly.
Seeds are stored in every generation's provenance record along with the model, parameters and references used. That is what makes "full parameter recall" real rather than a claim.