Skip to content

Original R reference and current Python replay

The reference bundle preserves the complete authenticated DISQOVER 0.9.14 numerical certificate produced on the original source commit 56997bbcfb5fcf7088035a16080b6a9416acc6e5. Its 124 members include raw R vectors, executed scripts, shared inputs, manifests, dependency and runtime identities, command outcomes, and Python expectations explicitly distinguished from R outputs. The bundle is 1,536,418 bytes and its certificate SHA-256 is 881c1f65b62c890fbb583c499493c21318f21857960998ca662fbb1cf2b94ae3.

The current replay evaluates the current Python implementation against those original R vectors. It first checks the pinned certificate, exact bundle inventory, source lock and actual original execution outcomes. It requires matching shared setup, R case expressions, case labels, modes and prospective tolerances. Decimal and hexadecimal literals compare by their exact binary64 values; strings and integer suffixes remain distinct. Every source-domain parameter and ring geometry remains matched.

Results and scope

  • 66 current cases considered: 48 evaluated, 48 passed, zero numerical failures, and 18 refused because the original bundle contains no R result for them.
  • All 25 original core cases match, covering dispersal kernels, deterministic and supplied-draw REVEALS, analytical LOVE, optimized candidate replay, and objective intermediates. All 23 distinct source-domain/corrected-forward cases match. Source refusals remain distinct from corrected matrix behavior.
  • Maximum absolute numerical difference: 3.197442310920451e-14. Each case retains its original tolerance and any current exact-coordinate assertions. NaN and infinity positions cannot become zero or disappear.
  • The current-source replay passed 48/48 cases with the same maximum difference and hashes all 85 current inputs, including the fixed-site and joint-draw modules. Nine previously recorded code inputs changed and one new native-parity module entered the source set; the numerical case outcomes remained identical. The deterministic REVEALS cases C0015–C0017, supplied-draw REVEALS C0018, and analytical LOVE C0023 all pass with their declared labels, modes, and tolerances.
  • Four focused regression tests passed; strict mypy and Ruff check/format passed for both producer and tests.

The 18 absent results belong to three newer prospective optimized-LOVE fixtures, each with six intermediate, axis, wrapper and nonfinite/refusal checks. Their absence is specific to this historical certificate. Separate current native R and source-closed Linux R runs executed all 43 core cases, including those 18, and all 23 source-domain cases with zero failures. Neither run changes the original certificate's 25-case denominator. Candidate replay establishes objective evaluation and selected-candidate parity, without an optimizer trajectory or global-convergence claim.

No fresh R execution occurred during this replay. The retained Linux R runtime and library remain local, while the Docker daemon was unavailable at inspection. This result is a current Python comparison to historical authenticated R outputs, not a current source-commit release certificate, independent vegetation validation, or complete publication replay. The original certificate continues to identify its original source commit and runtime.

Reproduction

From the repository root, choose an absent output below artifacts/:

PYTHONDONTWRITEBYTECODE=1 OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 NUMEXPR_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 \
  .venv/bin/python -m bijux_pollenomics_dev.lra_validation.reference_replay \
  --repository . \
  --bundle docs/report/method-certification/original-r-reference/reference-bundle \
  --output artifacts/jobs/retained-r-reference/reproduction

The producer validates the original bundle before computation, records its current dependency hashes and installed numerical versions, and rechecks the hashes afterward. Failed numerical comparisons remain in the denominator. A changed source expression or altered reference certificate refuses reuse.

The bundled original DISQOVER source archive and its license remain governed by the LRA NOTICE. Preserving this bounded certificate transfers numerical evidence out of disposable run directories; it does not copy the installed R runtime or library into the repository.