Unlimited-OCR-2 native JPEG comparison

The later execution follow-up supersedes this stage's CLI price and geometry results. It found that those CLI runs used a Metal binary inconsistent with its copied provenance stamp. Exact JPEG pixel/tensor parity remains valid; the earlier 9.000 CLI result isn't a verified-kernel parity result.

Unlimited-OCR-2 now decodes JPEG inputs with a private, statically linked libjpeg-turbo 3.2.0 decoder. The real receipt's decoded RGB bytes and resized, normalized BF16 input tensor match the pinned upstream Pillow capture exactly. The public Metal CLI produces identical output from JPEG and the canonical Pillow-decoded PNG, including the item's 9.000 price.

Receipt results

The checkpoint remains pinned to c14071b549f4fcab10909f3c4acc6a2edb366947. Runs use the original BF16 weights on an M5 Max with 128 GiB, temperature zero, TF32 disabled, a 1,024-token cap, and no-repeat size 35/window 128.

Public CLI input	Seconds	Block body/type matches with upstream	Maximum coordinate delta
Original JPEG, base	35.4	25/25	2
Original JPEG, auto	38.6	25/25	2
Canonical PNG, base	33.7	25/25	2

All three stdout files are byte-identical and match the earlier canonical PNG CLI output. Every request emits one page and stops before its token cap, with no stderr diagnostic. Raw streams, commands, hashes, and measurements are in cli/summary.json. Timings include startup and model loading; concurrent repository checks make them unsuitable as throughput benchmarks.

The body/type comparison counts a multiset of parsed blocks. Geometry uses only unique exact body/type matches: 18 pairs, six exact boxes, mean absolute coordinate difference 0.333, maximum two, on the model's 0–1,000 scale. Repeated bodies such as RM, 0.00, and 9.00 are excluded from geometry matching. These are agreement measurements against upstream output, without annotated ground-truth boxes.

The saved upstream receipt used repetition size zero. Its token trace shows that size 35/window 128 never bans a selected token. Both implementations still spell the printed MODELLING as MODELING; matching upstream doesn't establish correct transcription.

Pixel and codec checks

Receipt preprocessing difference from Pillow	Previous ImageIO path	Native JPEG path
Maximum RGB byte difference	41	0
Mean RGB byte difference	0.6918	0
Maximum normalized BF16 tensor difference	0.28125	0
Mean normalized BF16 tensor difference	0.0043135	0

See pixels.json and pixel-parity.log.txt. The decoded RGB SHA-256 and source hashes are in method.json. The identical-pixel probe checks the original receipt JPEG against the existing upstream tensor capture, including native letterboxing and resampling.

The decoder uses full-resolution 8-bit samples, integer slow DCT, and fancy chroma upsampling. RGB and grayscale produce opaque RGBA; CMYK conversion matches Pillow's inverted JPEG sample convention. ICC transforms and EXIF rotation aren't applied, matching ordinary Pillow loading without additional transforms. Unsupported precision and truncated inputs return decode errors. Other formats retain the platform decoder.

Five Swift tests pass. Seven independent Pillow 13.0.0/libjpeg-turbo 3.2.0 goldens cover RGB 4:4:4, 4:2:2, and 4:2:0, progressive 4:2:0, grayscale, CMYK, and RGB-coded JPEGs with odd image dimensions. Additional checks exercise malformed/truncated input, concurrent decodes, signature-based selection, and PNG fallback. The fixture generator and original expected bytes live in Tests/MereRunCoreTests/Fixtures/PillowJPEG/.

bash Tests/NativeJPEG/check.sh also passes 6,549 decodes under AddressSanitizer and UBSan, including every truncated prefix of those fixtures. The C error context owns allocations across longjmp; no jump crosses Swift frames. This is a bounded fixture sweep, not an exhaustive JPEG decoder audit.

XCTest execution difference

The full-checkpoint XCTest harness still emits 9.00. Its JPEG and canonical PNG outputs are byte-identical to each other, but differ from CLI output. Those paired runs share one generator instance, use the verified source files, and pass their page/output checks. Raw records are in harness-pair/; harness-comparison.json retains both paths.

The new native-JPEG harness output has 24/25 exact block body/type matches, mean coordinate difference 0.441, and maximum six among 17 unique pairs. The earlier ImageIO harness mean was 0.75, with maximum six. The matched sets are different from the CLI's set because the price text differs.

Repeating the original receipt base/auto anchor tests still fails the two MODELLING CLAY assertions shared with upstream. Those failures remain in receipt-anchor-tests.log.txt and results/. The paired test checks equal inputs/output and doesn't replace the original spelling acceptance criterion.

The execution follow-up resolves the CLI/XCTest difference. Earlier hash checks ran after the build restored test resources, so they missed the library replaced during direct XCTest execution. The binaries loaded by the two paths differed. With verified kernels, both paths agree and emit 9.00. Full upstream price and BF16 numerical parity remain unproven.

Source, ownership, and build impact

Sources/CMereRunJPEG/upstream.json pins upstream libjpeg-turbo commit c85e6b905bf237038faa936dab160ebfc5da0344, the archive digest, and hashes of 97 unmodified upstream source/header files. SwiftPM compiles 54 selected decoder translation units plus the C wrapper. Symbols have a private prefix; the decoder links without a system/Homebrew JPEG dylib. Python and CMake aren't build or runtime requirements.

MediaIO owns the new CMereRunJPEG dependency and exposes the opt-in decodeWithPillowJPEG facade. Unlimited-OCR-2 calls that facade. No new public SwiftPM product is added. Ten existing targets gain this private target in their transitive closure; the before/after inventory is in method.json. Package-policy checks pass for Darwin and both Linux manifest variants. Third-party licenses and the IJG acknowledgment accompany the source and are listed in THIRD_PARTY_NOTICES.md.

Three decoder-only clean builds take 0.726–0.922 seconds; recompiling the wrapper and relinking takes 0.149–0.185 seconds. The standalone Debug decoder probe is 523,000 bytes, compared with a 17,144-byte empty executable. The full Debug CLI is 243,005,200 bytes. These measurements retain dependency/compiler artifacts and overlap other qualification work. They don't measure a complete Swift clean build or the historical CLI size delta. See build-measurements.json and measure-build.py.

Reproduce

The required ./scripts/check.sh gate and corepack pnpm docs:build both pass. Codec tests, the strict receipt pixel/tensor probe, the paired-input harness check, and the public CLI comparison also pass. The XCTest section retains the original spelling-anchor failures. validation.json records the completed checks and log digests.

From the repository root, run:

swift test --filter PillowJPEGTests
bash Tests/NativeJPEG/check.sh
python3 docs/qualification/unlimited-ocr-2-metal-20261030/jpeg/run-receipts.py \
  --binary .build/debug/mere.run \
  --inputs .build/qualification/unlimited-ocr-2/inputs
./scripts/check.sh
corepack pnpm docs:build

The CLI probe requires the installed original checkpoint and the frozen receipt inputs described in the initial qualification. To regenerate the small codec goldens, use Pillow 13.0.0 built with libjpeg-turbo 3.2.0 and run Tests/MereRunCoreTests/Fixtures/PillowJPEG/generate.py. The optional upstream capture remains external because it includes large model tensors.