For state-spaces/mamba-130m-hf revision 1e76775f…, the compiler selected 96 projection-weight tensors and partitioned them into ideal 128 × 128 SVD/MZI blocks.
This note identifies the local result artefacts and the limits of the current evidence. It is not a photonic-hardware result.
Evidence boundary / Current
What this result proves—and what it does not.
Stages 01–03 are retained local software evidence only: selected projection-weight tensors from a retained Mamba-130M checkpoint were factorised into ideal MZI blocks and checked in a local software reconstruction report. The proposed physical bench remains beyond an uncrossed evidence gap.
Local software evidence only: 69.43% is partial checkpoint coverage, and block_ok=True is a software threshold flag—not a hardware test or pass. Reproducibility is not closed; the proposed fibre-loop bench has not been built and has no physical data. No photonic-hardware, optical-execution, loop-retention, or measured energy, latency, loss, throughput, TOPS/W, thermal, cost or system-performance result exists. Detailed recorded results, artefact identifiers, reproducibility status and the next evidence gate follow below.
Recorded result
Selected checkpoint elements
89,653,248 / 129,135,360
Partial checkpoint coverage
69.43%
Projection tensors
96
Ideal MZI blocks
5,760
Ideal mesh targets
11,520
Mean block relative Frobenius error
1.264228 × 10−6
Maximum block relative Frobenius error
1.397348 × 10−6
All 5,760 rows in the local block-reconstruction report record block_ok=True against the configured software reconstruction threshold. This is not a hardware test or pass.
Artefact identifiers
These hashes identify the retained local files. The files are not presented here as a public reproduction package.
The result is locally auditable but not yet independently reproducible from a clean repository clone. The primary reports, factors and compiler scripts are not in a committed immutable run package, and the run lacks a recorded code commit, dependency lock, seed, environment manifest and aggregate factor manifest.
Until that package is published, the defensible claim is limited to a locally verified software reconstruction result. No optical hardware, loop, energy, latency, loss, TOPS/W or thermal performance was measured.
Next evidence gate
After reproducibility closes, build and commission a proposed fibre-loop bench, then measure coherent-state lifetime, round-trip insertion loss, phase drift and noise, and end-to-end wall power. No physical photonic bench or chip has been built.