Skip to content

Local Linux R oracle execution

The source-bound report records an actual Linux ARM64 execution of source-built R 4.6.0 and the locked DISQOVER 0.9.14 private library. The core raw results passed 43/43 cases, including the 18 prospectively declared optimized LOVE cases. The source-domain raw results passed 23/23 cases. Both raw directories include executed R scripts, case manifests, source inputs, individual R vectors, R completion statuses, and R session information. The report binds their complete byte inventory to the exact source lock and 1,661 captured executable source files.

The run reused a previously source-built Linux R prefix and its private source-provisioned package library. Before execution, the producer verified the 38 locked source archives, the R build and provisioning receipts, 1,227 R runtime files, 2,784 private library files, the installed package-version closure, and the DISQOVER archive. It mounted the source, runtime, and library read-only into a container with one CPU, 4 GiB memory, and no execution network. The container ran Python 3.11.16, NumPy 2.4.6, Pandas 2.3.3, SciPy 1.17.1, and R 4.6.0 on aarch64-unknown-linux-gnu. Its exact image ID and pinned Python base digest are in the report. The producer checked the governed source hashes again after both R runners and copied raw results to this directory only after the complete 43 + 23 gate and byte verification passed.

The commit context binds the generated report's SHA-256 to the 1,661-file executed source SHA set. The first HEAD observed just after the run was 44c89c7c77f720166a86dc73b750b3e9223945d2; the repository had advanced to a862468a8ad93cc2e522bdf0173f81f6b84e8cc9 before the context annotation was written. The producer verified source bytes before and after execution but did not independently capture Git HEAD inside the run. The result therefore belongs to its recorded byte set and must not be presented as certification of either commit or a later source revision.

An earlier local run also completed 43 + 23 cases, but its source tree changed during execution, so we refused that run. The source-closed replay took 17.37 seconds and is the sole admitted local Linux result here. Its governed raw vectors, completion statuses and environment record are linked above; an old local log directory is not required to read or verify the result.

Reproduction from private custody

The private data store retains the exact source-built runtime and package library, the 37 locked R/CRAN source archives, the image ID and the Dockerfile under lra-reference/. Its linux-arm64-oracle/custody.json has SHA-256 0398b643a2124ce44d9854a6979b1939c6c93337ddea30d1e030439de44d3575. The custody was checked against the source lock, the original build and provision receipts, all 1,227 runtime files and 2,784 library files. It also retains 130 acquisition, build and provision process files with individual hashes; unpacked R source and build intermediates are reproducible from the locked archive and are not part of that private copy. The DISQOVER source archive remains governed separately by the repository's LRA NOTICE.

Set BIJUX_LRA_CUSTODY_ROOT to the restored private lra-reference/ directory. With the recorded container image available locally, run the producer from the repository root with new absent output destinations:

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.linux_reference \
  --repository . \
  --output artifacts/jobs/linux-r-oracle/reproduction \
  --retained artifacts/jobs/linux-r-oracle/reproduction-retained \
  --runtime-tree "$BIJUX_LRA_CUSTODY_ROOT/linux-arm64-oracle/runtime" \
  --provision-tree "$BIJUX_LRA_CUSTODY_ROOT/linux-arm64-oracle/provision" \
  --source-cache "$BIJUX_LRA_CUSTODY_ROOT/source-cache" \
  --image-id-path "$BIJUX_LRA_CUSTODY_ROOT/linux-arm64-oracle/image.id" \
  --dockerfile "$BIJUX_LRA_CUSTODY_ROOT/linux-arm64-oracle/Dockerfile"

The producer verifies every archive, receipt and installed file before execution. The image ID names a locally built image, not an exported image archive; another machine must restore that exact image or build and qualify a replacement. A freshly rebuilt R runtime or library also needs a new qualification if its bytes differ from the retained receipts. The original raw oracle outputs remain in this repository regardless of local container availability.

Release boundary

This result is local Linux ARM64 numerical evidence, not a release certificate. The locked release job requires Ubuntu 24.04 on x86_64 and a matching GitHub run, attempt, committed source revision, and uploaded artifact digest. None is supplied by this local run. The release gate must evaluate its own same-attempt CI certificate. Numerical parity does not establish stochastic optimizer trajectory parity, independent vegetation accuracy, or ecological propagation accuracy.