Saved benchmark evidence for d15b0bc. Release publication is pending. This local report includes public-safe tables and no source pixels. Links to unpublished release documents are shown as text.

Inspect and reproduce the final native benchmark

The published tables are saved observations of d15b0bc, not measurements of the newer launch candidate. The original seven-lane programme contains 588 complete operations:14 dataset/query quadrants ×2 transports ×7 lanes ×3 observations. See FROZEN_PROGRAMME.json for the exact seed10738291, geometry, control choices, numerical policy, resource limits, source hashes and runtime versions. The old eight-lane672-operation tournament is a different historical programme.

Recheck the report without raster access

python benchmarks/native-skv/verify.py
python benchmarks/native-skv/render.py --root benchmarks/native-skv \
  --output /tmp/skarve-charts-new

Rendering needs Matplotlib3.11.1 and NumPy2.5.3. Output must be new. Included PNG/SVG charts are byte-identical exports of the final chart set; regenerated image bytes can differ with renderer metadata/fonts. The numerical tables and generation script identify their inputs. All588 operation records,196 lane summaries,28 fixed-summary comparisons, source/index storage, numerical differences, summary ablations and60 sentinel observations are available under data/ and sentinel/. No timing is rerun by these commands.

PROVENANCE.json maps each public evidence file to its retained source hash. Private absolute file references became stable logical identifiers; private-reference/... is not a downloadable dataset. Timing, result, cost and loss fields remain unchanged. The raw private trace archive and source rasters are not redistributed. Tests and generated fixtures contain no production data.

Replicate using public synthetic data

Use the qualified Ubuntu24.04 toolchain with GDAL3.8.4 for ordinary fixtures, and the explicitly separate Rasterio1.5.1/GDAL3.12.4 COG encoder. Install the candidate Python package and test/benchmark dependencies first. /usr/bin/python3 needs the distribution's python3-gdal and python3-numpy; the active python needs the installed Skarve package, NumPy2.5.3, Rasterio1.5.1, exactextract0.3.0 and Shapely. Keep BLAS/GDAL worker counts at1. No cloud credentials or data download is needed.

WORK=/tmp/skarve-native-replication-new
mkdir "$WORK"
/usr/bin/python3 benchmarks/skv-v0/fixtures.py --output "$WORK/fixtures36" \
  --bands 36 --width 513 --height 519 --seed 17391
/usr/bin/python3 benchmarks/skv-v0/fixtures.py --output "$WORK/fixtures40" \
  --bands 40 --width 1025 --height 1031 --seed 17391
python benchmarks/skv-v0/modern_cog.py --fixtures "$WORK/fixtures36/fixtures.json" \
  --output "$WORK/fixtures36/modern.json" --block 128
python benchmarks/skv-v0/modern_cog.py --fixtures "$WORK/fixtures40/fixtures.json" \
  --output "$WORK/fixtures40/modern.json" --block 128
LIBRARY=$(python -c 'from raster_engine_lab import resolve_library; print(resolve_library())')
python benchmarks/skv-v0/prepare.py \
  --fixtures "$WORK/fixtures36/modern.json" "$WORK/fixtures40/modern.json" \
  --output "$WORK/prepared" --library "$LIBRARY" \
  --chunk-edge 128 --band-group 64 --predictor byte_delta_v1 \
  --payload-layout band --skip-equivalent-cog --extra-index-variants cogband128
python benchmarks/native-skv/replicate.py freeze \
  --prepared "$WORK/prepared/prepared.json" --output "$WORK/final"
# Copy the exact freeze_sha256 printed above. Do not change it after timing begins.
python benchmarks/native-skv/replicate.py run --output "$WORK/final" \
  --freeze-sha256 PRINTED_SHA256

These two analytical families produce 336 operations with the same four query quadrants, local/HTTP controls and fixed geometry. The new replication records its own actual object/runtime hashes. Source logical sample/mask hashes, interpretation and predeclared layouts must match; it refuses silent substitution. Preparation receives no polygons. --smoke --datasets analytical36 at the freeze step creates exactly two local functional operations, explicitly excluded from full benchmark claims. It is useful for testing the installed reproduction harness before a long independent run. The smoke needs only the installed Skarve runtime; full replication additionally requires the exact natural-control versions listed above. Pass --python /path/to/venv/bin/python if needed: the driver preserves that virtual environment, pins its executable, binding modules, library and dependency versions, then checks them again before execution. Activate the same environment for preparation and freeze.

The runner executes the original final native worker, numerical validators and HTTP server. Source open, identity, metadata, indexes, ranges, decoding, computation and consumed five-field output remain inside primary timing. Imports/runtime verification/engine lifecycle are reported separately. Fresh processes do not evict the OS page cache; pre-run file hashing warms it. HTTP is unthrottled loopback, not WAN/R2. Do not run simultaneous builds or other benchmarks. The fixed2GiB process address-space,1GiB job working budget,64MiB caches,180second timeout and source request/byte bounds remain unchanged. Complete-operation failures stay in the denominator; there is no retry or outcome-selected layout. complete.json reports planned, attempted and unattempted operations separately. A detected source replacement or an evidence-cap stop retains the completed prefix and counts the remaining planned operations as unavailable, never successful. Per-operation HTTP traces stay in compressed files rather than accumulating in the coordinator. External interruption preserves the attempt/worker/record prefix; it does not authorize resuming or overwriting that programme.

Real data and exact historical reproduction

Full588-operation replication additionally requires the exact real36 native crop and WorldPop source under their applicable access/data rights. No acquisition URLs or source pixels are included. Their logical hashes, native grids, query frame and physical layouts are in FROZEN_PROGRAMME.json. Supply their independently prepared manifests and explicitly include real-age36,worldpop1 in --datasets; source validation is identical. Generated40 is not a real40 result.

The old d15 binary hashes and dependency versions identify the measured implementation. A rebuilt native library or newly prepared object has a new identity even if logical values agree; new timings are replication results, never replacements for the saved report. The launch facade does not silently inherit benchmark results. New receipts label the historical source commit separately and leave the new binary source commit unspecified unless independently established by its build receipt. Historical conversion times refer to the unchanged original objects/writer; newly executed preparation writes its own timing and storage receipts.

The independent sentinel against the older b7 build preserves batch regressions in the saved public tables. It is context, not a requirement to rebuild an obsolete runtime for every public installation. No new optimization campaign is part of reproduction.

Driver-only checks (no raster operations):

python -m unittest discover -s tests -p test_native_replication.py