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 Links to unpublished release documents are shown as text.

Retained preparation and storage costs

This is an object-level reconciliation for the frozen d15 main programme and its separate coalescing sentinel. No conversion, source acquisition, index build or query was run for this note. Current serving uses source commit d15b0bccdf87654e6e6bd221ce114fb4f129e42d; every preparation time below predates it. The recipe and source bytes were preserved, not regenerated with d15.

Main objects and retained originals

The CSV contains 38 distinct object paths: all 33 main serving objects (336,407,132 logical file bytes), two retained ordinary TIFF inputs not selected as main serving controls, and three distinct sentinel SKVs. It matches every main objects path, SHA256 and byte length. The CSV is an object inventory, not a sum over 588 operations or 196 cell/lane slots. Logical file bytes are not filesystem allocation and do not include packaging, source checkouts or evidence.

The four ordinary compile inputs total 67,862,601 bytes; their four SKVs total 82,213,780 bytes; retaining both totals 150,076,381 bytes. These are the same originals and SKVs across summary-on/off lanes. SKV serving is self-contained (serving_requires_original=false); original retention supports provenance and rebuilding. “Original” here means the exact normalized ordinary TIFF compile input, not every earlier download or a larger private country source. The 40-band dataset is generated, not actual real40.

The complete conversion costs match the exact-byte historical lifecycle rows. The main SKV writer library was 1f9e848402532f61eaf503e18230e0be89474741f5e196b0389d45e0ca992f3e, recorded in the original preparation manifests. This is distinct from the later R11 serving library 5b6fcdf1…60639, its build commit 507a6e9…, the later source export f54783e…, and d15. No source commit is inferred from an unlinked library hash. Any changed pixels, masks, metadata or layout require a full immutable rebuild; no incremental update saving is claimed.

The CSV also retains available upstream fixture-preparation durations: analytical generation, authorized real36 crop/encoding, and existing WorldPop ordinary encoding. They have different boundaries from downstream SKV compilation and are separate stages, not missing conversion overhead. Original lifecycle CSV blanks remain unknown at that table’s encoding grain. These one-off historical measurements do not predict OS-cold rebuild latency, network acquisition, every input, or another machine.

Controls and indexes

Direct control costs are joined by exact dataset, variant, source path and SHA256. This includes controls selected by the retained R12 evidence that were absent from some original 672 lanes. BAND256 is a tiled GTiff, not a COG. PIXEL128/256 are GDAL COG layouts; BAND128 COG carries its explicit Rasterio 1.5.1 / bundled GDAL 3.12.4 encoder provenance. Other retained fixture encoding uses GDAL 3.8.4. Their recorded elapsed_seconds is the historical rearrangement/encoding cost; d15 did not produce new competing objects. Selection is fastest observed eligible within the declared tested control class on old geometry, not a universal optimum. Native and natural exactextract selections remain separate policies.

RSI is additional to its exact boundary TIFF/COG, not a replacement for it. Existing indexes contain at most 20 bands per group. A band 23 query needs only the group starting at 20; full 36/40-band queries use both groups in one job. The CSV lists each physical RSI once and separately repeats its all-group bundle total for context: do not add that repeated total across rows. Nor should COG bytes be multiplied by index groups, transports or repetitions. Generated40 RSI controls were built with library c8a38d1e102541f6f7b9e92e4cfc2cf37d5eeb2a22e620ca10a511823598d139; other listed RSI builders use 1f9e8484…992f3e. Index preparation is additional to producing the chosen TIFF/COG, but per-group subphases are already inside their complete bundle time.

Separate sentinel layouts

The R12 group1 objects retain encoded samples and summaries while changing physical order; their exact hashes differ from main. Their historical plan pins 5b6fcdf1…60639 and the f54783e source export. The row-group40 object is an R8 historical explicit-development-library compilation (a50601aa…ade64d), not an installed d15 writer measurement. Its source-to-complete time is 7.031308 s; full recorded lifecycle including engine setup is 7.052697 s. R12 outer result wall times include independent correctness/oracle work; those envelopes are not added to the conversion clock.

Accounting boundaries and reproducibility

- Source read/decode, predictor, compression and summary milliseconds are components of complete compilation, not additional costs. Instrumented phases can overlap; no sum is used to manufacture an unexplained remainder. - source_retained_bytes is the compiler’s live source-adapter retained buffer (16MiB here), not source-file disk retention. The CSV renames it source_retained_buffer_bytes. working_bound_bytes is a bound, peak_scratch_file_bytes is a scratch-file measurement, and neither is process RSS. Do not add unrelated peaks as simultaneous memory. - Encoded payload, directory and bootstrap are components of the SKV file, not extra permanent bytes. Typed sample and mask bytes describe logical uncompressed content, not additional retained files. - Full rebuild costs are historical; modern reader/reducer/coalescing changes do not prove modern writer speed. No cloud acquisition cost, remote egress tariff, amortization threshold or OS-cold claim is inferred. - Object checks here use the already admitted source descriptors and original small receipts; no raster payload was read again. Release-package verification is separate in PRIVATE_INSTALLATION_01.md.

Evidence links (each CSV row pins its applicable receipt SHA256):

- Main freeze — SHA256 c450e98a310a38d7d4621a9656ae444419828934e78f9596789634354890e275. - Prepared source descriptors — SHA256 4eddf525ce10d591e3f2ba2f2eba0b3527b4b4cda5746b26e9cb8ba5c2c5dbb9. - Historical lifecycle table — SHA256 9814fe485a3df025c889f559dd78de10b093190848097abc43246a50fd9247d3. - R12 control map. - Object-level CSV — SHA256 1cfd2d132f93f60e82e9be2ad68fde077b72efe59bea1475aec995a8d5298e06. - prepared-final-other-r3 preparation — SHA256 c540147d452cafa03103540945ff3ecb4de8d949de48724f35b5a08c72501624. - prepared-r3-large40-group40 preparation — SHA256 407b94d81824b3e8d737135417f47f85a831b08f14ad4d752d2c36ecbe37bd5a. - prepared-large40-index-controls preparation — SHA256 ba3241ce4c45dec42cb6a702144e2247a59623b959df93c1dc947fe7116e21f1. - R12-band-group1 / analytical40-1025x1031 receipt — SHA256 613ff91215f6b27cefc0b1a5e7eefe70f22b5be4c6d8a02877dc25d4661b2642. - R12-band-group1 / real-age36 receipt — SHA256 bf98272ef03da1bf5c22dd3a03bdb25de6b3c8acc5c9d160f7130f717d3735cd. - R8-row-group40 / analytical40-1025x1031 receipt — SHA256 22b2a3ba352844b5af244131e3e21f25f2efd17a800b02666e4c540ccfe7af93.

The authoritative revision-02 lane-storage table has 196 frozen cell/lane rows. An independent cross-check matched all 196 source path/SHA/byte triples, every query-required RSI byte sum, and all 33 object identities to this inventory. Its corrected RSI accounting supersedes the preserved first derivation; no measurement changed. SHA256 ee146c84b4462bc85560fe29077564c74efe36919d2d7e628639844e6226f319. Per-object CPU/RSS and source acquisition costs are unavailable where the historical receipt exposes only whole-preparation resource envelopes; those envelopes were not allocated artificially to rows.