Two-Tier Reproduction: Byte-Exact and Within-Tolerance

The breeder record accepts two levels of reproduction — a byte-for-byte match in an identical environment, and a numerically-close match across platforms — and states which each analysis targets.

Two honest definitions of 'same'

'I reproduced it' can mean two different things, and conflating them causes needless disputes. The breeder (Hyperion) record defines both explicitly so a reproducer knows which bar to clear.

Tier 1 — byte-exact

In the pinned container, on a compatible platform, an output file should match the published file bit for bit; its checksum should be identical. This is the strongest form of reproduction and is expected for deterministic analyses run in the reference environment.

Tier 2 — within tolerance

On a different platform or math library, floating-point results can differ in the last digits. Tier 2 accepts a match within a stated numerical tolerance rather than bit-for-bit. The tolerance is published per analysis, so 'close enough' is a defined number, not a judgment call.

def reproduces(observed, expected, tol=None):
    if tol is None:               # Tier 1: byte / exact
        return checksum(observed) == checksum(expected)
    return abs(observed-expected) <= tol*abs(expected)  # Tier 2

A Tier-2 pass with a Tier-1 failure is normal across platforms and does not indicate an error. A Tier-2 failure does, and is a reportable discrepancy.

Separating the tiers prevents a common and pointless argument. Without the distinction, a reproducer on different hardware sees last-digit differences, concludes the record is wrong, and reports a non-existent error; or an author dismisses a genuine discrepancy as mere floating-point noise. Publishing a per-analysis tolerance settles which reading is correct before the dispute starts, because 'close enough' is a number in the deposit rather than a matter of opinion.

This page describes a design-and-simulation study, not a built machine. Construction of the breeder (Hyperion) begins Q2 2027; first-of-a-kind first tritium is targeted near 2030. No hardware net-gain claim is made before FOAK.