Skip to content

Finite-sample decision risk in unseeded quantum compiler benchmarking — Version 4

Latest

Choose a tag to compare

@Mormolykos Mormolykos released this 19 Sep 19:03
· 1 commit to master since this release

Version 4 is the audited release of this research object.

Scientific endpoint: 12/26 eligible circuits have a decision-risk interval excluding zero; 7/26 at risk ≥ 5%; 4/26 at risk ≥ 10%.

Identifiers

Source SHA e663912278baeda6d24787b37ad8a04243a4280b
paper.pdf SHA-256 19e37c1954368928475b4ef51d5f8ca8a28d41b896932de6df8e4edf0fe9c7a5
qvalidation-e663912278ba.zip SHA-256 07126ddda949ffb8a4d44b5763df5e22ff8488b20aac958bad088653b4e19941
Raw pre-registered tree 3114a270ccd78b0daaa2a1bc451dd15155d55c13

The reproducibility ZIP contains the exact 1,265-file frozen Git tree at that SHA. Every member was hashed individually against its Git object before release; all 1,265 are byte-identical. It carries no build artifact, cache, virtual environment or editor material. publish/paper.pdf is deliberately absent: the freeze pins the build inputs, not the rendered artifact, because the renderer stamps a creation timestamp into every build and byte-identical PDFs are therefore not promised.

Assurance

An independent hostile audit of this lineage found no release-blocking defect. The verifier runs thirteen read-only stages and a 44-fixture mutation suite in which every corruption must be rejected by a named stage, under the invariant that stage prints in its own terminal verdict — 44 of 44 rejected, none skipped, no execution errors.

The boundaries of that assurance are documented rather than implied, and the apparatus is not claimed to be exhaustive or formally verified. In particular:

  • Stage 13 binds the published PDF's word multiset to the registered manuscript surface in both directions. It does not establish semantic word order, so a rearrangement of the same words into a different claim is not caught by it.
  • The registered surface is a declared Markdown model of the manuscript, not the rendered page. Two measured differences are recorded in visible_surface.py and in V4_ASTRA_CORRECTION_LEDGER.md.
  • A documented human reading of the built PDF is a release requirement that no check in this repository can close. The final PDF attached here received human-reader sign-off.

Prior overclaims and their corrections remain documented in the paper and in the correction ledgers rather than being removed.