Online Resource 1
Exact softmax sampling from residual quantum overlaps
Quantum Information Processing

Author: Vikram Lex
Affiliation: KarLex AI, Newark, DE, USA
Email: vikram@karlexai.com
ORCID: https://orcid.org/0009-0009-8364-244X

Source code, frozen pretrained-model query/key/value traces, model and dataset metadata, dependency versions, complete experiment records including resource failures, editable manuscript materials, and reproduction instructions.

Extract the complete archive into an empty directory. The original code/ and paper/ directory layout is preserved. Start with paper/residual_attention/REPRODUCE.md and code/certified_attention/README.md. Both editable manuscript sources and numerical inputs are included, so the generated figures and tables can be reinserted by compiling those original sources. The separately supplied manuscript_source.zip contains the same submitted main text in the flat layout used by the journal's submission system.

The reproduction instructions use local CPU computation. Historical quantum/GPU workloads are not rerun by the hardware reanalysis commands; the theory experiment uses simulated independent measurements and frozen model traces. Archived scripts are evidence of the earlier workflow, not instructions to purchase new cloud executions.

Licensing and attribution: code/LICENSE records the project-code license. Third-party model and dataset terms remain those of their respective sources; this notice does not relicense them. Raw text corpora and pretrained model weights are not included. The earlier QHM project preprint is cited at https://doi.org/10.21203/rs.3.rs-9825475/v1 (Research Square, CC BY 4.0).

SHA256SUMS.txt covers every other archive member. The embedded source and data checksum lists cover the delivered files, including the documented public export. RESOURCE_METADATA.json records their archive hashes and this resource's bibliographic identity.

Public trace export: unused input token-ID arrays are omitted from both trace NPZs. All Q/K/V and projection-basis arrays remain byte-identical to the original inputs. The scientific code and recorded outcomes are unchanged. The trace JSON and run-configuration trace checksums identify the public export and retain original hashes. The historical report.json retains its original embedded configurations; regenerating it changes only those configuration identifiers, not computed results. PUBLICATION_EXPORT.json documents the export, and THIRD_PARTY_NOTICES.txt identifies source terms. The local original archives remain intact. If extracting new traces from source data, the original scripts also save token IDs; these are not needed for the experiments.
