Skip to content

AI Usage Disclosure

This document discloses the use of AI assistants in qgrav's development, as required by JOSS's 2025 submission policy (docs/research/RESEARCH_JOSS_SUBMISSION.md) and as a matter of scientific honesty. It is intended to be cited (or copied verbatim) in the methods/acknowledgements of any paper that uses qgrav.

Summary

Substantial portions of the v1.0 and v1.1 implementation, the v1.0 architectural plan, the research-prompt drafting and the supporting documentation were produced with AI assistance (Claude). Every line of physics-sensitive code, every benchmark value, every literature citation and every test was reviewed by the human author (Aditya Prakash) and verified against primary sources where possible; bugs identified by that review process were corrected in v1.0.1 (Hu 2013 short-term noise off by 10×, Ménoret 2018 short-term noise off by 1.5×) and v1.0.2 (audit of the remaining 10 reference values, finding three unit-category errors and two ambiguous entries).

Specifically

Code generation. Routine plumbing (imports, dataclasses, YAML passthroughs, parametrised tests, docstrings) was generated by AI then inspected. Non-routine physics code - the integrated-phase patch, the Bertoldi closed-form predictor, the PID servo, the correlated-vibration sampling, the calibration step - was co-designed in conversation with the AI and verified against the primary literature.

Plan generation. The development plan and documentation structure were AI-drafted under the author's direction, then reviewed and signed off by the author before being implemented.

Research-prompt drafting. The 14-topic research brief was AI-drafted, then the author ran the web research separately and recorded the results under docs/research/. Those research files are AI-assisted summaries of primary sources, with verbatim quotes; the author re-verified the most consequential numbers against the cited papers, catching one premise error (Hu 2013 is the HUST short fountain, not the Wuhan 10-m tower).

Test design. Test parametrisation, fixture design, tolerance choice and seed selection were AI-assisted; the tolerances were calibrated empirically by running the tests and adjusting until they reflected the underlying physics correctly (e.g., factor-3 envelope on Freier ASD vs factor-4.5 on the secondary benchmarks, both motivated by the V≈0.6-vs-1 inversion mismatch).

Documentation. The Markdown docs in docs/, the CHANGELOG, the README and this disclosure file were AI-drafted, then reviewed and edited by the author for accuracy and tone.

Verification posture

Aditya Prakash is the sole author of the project, takes responsibility for the code and the claims, and is the first point of contact for any reviewer or user query. The physics-review packet (docs/PHYSICS_REVIEW_PACKET.md) is the primary mechanism by which the project's core physics derivation (the integrated-phase patch and the factor-of-2 finding) will be independently validated by a domain expert external to the project; until that review is on file the FULLY_SIMULATED study-scope label should be read as "the simulation does not invoke the analytical formula", not "this has been peer-reviewed".

Concrete bugs the AI workflow introduced

Listed here so future users can decide for themselves whether the verification posture is adequate:

  • Hu 2013 short-term noise stored as 4.2e-9 m/s²/√Hz instead of 4.2e-8 (off by factor of 10). Fixed in v1.0.1.
  • Ménoret 2018 short-term noise stored as 5e-7 m/s²/√Hz; the as-operated Larzac value is 7.5e-7. Fixed in v1.0.1.
  • Kasevich & Chu 1991 stored 3e-6 (a dimensionless Δg/g at 1000 s, in sodium) as if it were an amplitude spectral density. Fixed in v1.0.2.
  • Bidel 2018 marine stored the 0.17 mGal static uncertainty as a /√Hz sensitivity; the actual sensitivity is 0.8 mGal/√Hz. Fixed in v1.0.2.
  • SG noise floor stored the 1 nGal frequency-domain detectability as an ASD; off by ~100×. Fixed in v1.0.2.
  • Hu 2013 initial assumption (Wuhan 10-m drop tower) was a different paper entirely. Corrected during Phase 16 research and execution.

The pattern (quantity-type confusions and parameter mis-identification) is consistent with AI-assisted citation work and is a stronger argument for independent physics review than for de-listing AI assistance.

License

This file is part of the qgrav distribution and is licensed under GPL-3.0-or-later (same as the rest of the project).