Unlimited-OCR-2 first layout-token investigation

The first receipt output split occurs at zero-based token 20, in the first text box's left coordinate: native emits 104, and upstream emits 106. This follow-up traces that low-margin decision with identical token histories, upstream activation substitutions, and original BF16 checkpoint weights. It doesn't resolve full receipt parity or establish another native equation or dtype error that warrants a production correction.

First layout decision

Both runs share the preceding text:

<PAGE><BLOCK>image[131, 0, 862, 99]<FORK>
<BLOCK>text[

The upstream model and checkpoint remain pinned to c14071b549f4fcab10909f3c4acc6a2edb366947. The reference uses PyTorch 2.13.0 on MPS, BF16 weights, greedy selection, and the Pillow receipt input. Native uses MLX on Metal with TF32 disabled. The JPEG samples and normalized BF16 input remain identical. These are agreement checks against upstream, without annotated ground-truth boxes.

Decoder and image features	104 score	105 score	106 score	Choice
Upstream, upstream features	31.0	31.0	31.125	106
Native, upstream features	31.125	31.125	31.125	105
Native, native features	31.0	30.875	31.0	104

Token IDs are 7163 (104), 6992 (105), and 7336 (106). Ordinary argmax selects the lowest ID at an exact tie. The shared history is replayed through the target decision; the target token remains an unforced prediction.

The diagnostic substitutes each upstream decoder block output at every replayed step, including prefill. With either image-feature source, replacing block 0 yields 105, while replacing block 1 first restores 106. Subsequent boundaries are nonmonotonic: some restore the reference margin, and others produce ties. This doesn't isolate a unique block-1 implementation defect. Replacing the last block or final normalized state recovers all three reference candidate scores exactly. The native final normalization and output head therefore reproduce this decision on the upstream state.

boundaries.json records every substitution, candidate score, and earliest token disagreement. The corresponding upstream capture metadata is in upstream.json.

Identical-input numerical probes

Each decoder block is tested on the upstream hidden input, upstream prior KV and router cache, and identical positions. At token 20, layer 0's two norms and attention output match exactly. Its feed-forward output differs in 26 of 1,280 cells; the residual output differs in six. Isolating its operations finds one differing cell in each gate/up projection. SiLU and the down projection match exactly on identical inputs.

The gate projection's differing value is approximately 0.002693178394 before BF16 rounding. CPU FP64 accumulation rounds to native's 0.002700805664; upstream instead returns 0.002685546875. The up projection's value is approximately 0.096435545085. CPU FP64 rounds to upstream's 0.0966796875; native returns 0.09619140625. Their proximity to BF16 rounding midpoints makes FP32 accumulation order significant. Neither backend consistently supplies the correctly rounded result on these probes.

Prefill has an additional decoder difference. On upstream embeddings, the first input norm and isolated Q/K/V projections match exactly; isolated gate and up projections on upstream's second norm also match. The complete first attention path differs in 258,094 of 357,120 output cells, with RMSE 0.0003884. Native uses a gathered KV set for each query, while upstream uses dense masked prefill, so the attention kernels receive different shapes and reduction orders. This probe doesn't isolate a further equation or dtype defect within that path. Restoring the first coordinate with the upstream patch tensor doesn't eliminate these additional decoder differences.

A separate sweep checks every finite BF16 input in [−32, 32], totaling 33,794 values. SiLU and sigmoid match upstream exactly. GELU differs at −5.5 and −5.53125 by at most 1.65e−7, reflecting the backends' error-function tails. This doesn't identify a SiLU precision-boundary error behind the decoder split.

The first numerical difference in the vision path is SAM's patch convolution: 94 of 3,145,728 BF16 outputs differ, with maximum difference 0.0078125. At the first differing cell, [0, 0, 33, 751], CPU FP64 accumulation favors upstream. Other sampled cells favor native. The convolution preserves FP32 accumulation and bias addition before BF16 output rounding; its remaining differences arise near cancellation and rounding boundaries.

That first convolution drift changes the layout prefix causally. Replacing only its output with upstream's patch tensor restores 106 at token 20; every later vision and decoder operation remains native. No earlier token changes. This restores the first coordinate, without establishing full-run parity.

Patch calculation supplied to native vision	Differences from upstream patch	First coordinate
Original native convolution	94	104
Upstream patch output	0	106
Native FP32 channel-major matrix projection	483	105
Native FP32 spatial-major matrix projection	94	104

Both matrix probes add bias before BF16 output rounding. Merely matching the upstream tensor's channel-major storage order when flattening patches makes agreement worse. The spatial-major projection reproduces the native convolution. Neither probe provides an evidence-supported production fix, so the production equations and precision remain unchanged from the prior DASH scalar correction.

isolated-blocks.json contains same-input metrics. rounding.json preserves the actual FP64 dot products and sampled backend results. These probes distinguish accumulation differences from equation and dtype mistakes; they don't justify changing model precision or token tie-breaking to favor this receipt.

Repeated receipt comparison

JPEG base, JPEG auto, and canonical PNG base were repeated through both CLI and XCTest with greedy decoding, repetition size 35/window 128, and a 1,024-token cap. All six outputs are byte-identical, contain 495 generated tokens, and stop before the cap. They also match the prior price-investigation outputs byte for byte. Native still emits 9.00. The fresh independent upstream run emits 9.000 and reproduces all 495 token IDs from the prior reference exactly. Its original-source hash and full output are recorded in upstream-rerun/receipt-base-bfloat16-ngram35.json.

Every run uses verified Metal library SHA-256 64431ed485178f63a8dc870c8231ba021090c3356218c607a726821a7b6abb82. The kernel verification script passes against the pinned source and binary digest. The same verified artifact is loaded beside the XCTest image. results/summary.json records hashes, comparisons, and raw-output locations. Timings include model loading and aren't controlled throughput measurements.

Agreement with upstream remains 24/25 exact block body/type matches. The same 17 unique body/type pairs have mean absolute coordinate difference 1.176 and maximum difference 4. Repeated bodies are excluded from geometry comparisons. The shared printed MODELLING CLAY spelling failure remains. The unrelated eight-case document suite wasn't repeated.

The standard gate passes: 5,675 XCTest cases, including 477 skips, plus 197 Swift Testing cases, with zero failures. All 18 owned numerical tests also pass explicitly on GPU, as do the external-asset boundary and prefill diagnostics. The documentation build passes. Details and source/capture hashes are in validation.json and method.json.

Reproduce

Use the upstream environment and pinned source described in the parent report. Capture all block states through the first layout decision:

python docs/qualification/unlimited-ocr-2-metal-20261030/price/trace-price.py \
  --source-root /path/to/upstream/model --image /path/to/receipt.jpeg \
  --reference /path/to/upstream-receipt.json --step 20 \
  --capture-all-steps --capture-leaves --output /tmp/unlimited-ocr-layout

python docs/qualification/unlimited-ocr-2-metal-20261030/layout/probe-rounding.py \
  --source-root /path/to/upstream/model --capture /tmp/unlimited-ocr-layout --prepare

Set MERERUN_TEST_UNLIMITED_OCR_LAYOUT to the capture directory and MERERUN_TEST_UNLIMITED_OCR_ROOT to the original checkpoint. Run UnlimitedOCRLayoutDiagnostics.testFirstLayoutSplit with MERERUN_TEST_MLX_DEVICE=gpu and MLX_ENABLE_TF32=0. MERERUN_TEST_UNLIMITED_OCR_PRICE_NATIVE_FEATURES optionally supplies native page features. MERERUN_TEST_UNLIMITED_OCR_LAYOUT_PREFILL supplies the earlier prefill capture, including pixels and SAM/decoder intermediates. MERERUN_TEST_UNLIMITED_OCR_LAYOUT_PATCH_SUBSTITUTIONS=1 enables vision-patch substitutions; MERERUN_TEST_UNLIMITED_OCR_LAYOUT_PROBES_ONLY=1 skips decoder boundary replay. Run probe-rounding.py again without --prepare to produce the FP64 comparison, optionally passing --prefill with that earlier capture.

Large activation and checkpoint files remain external. Production inference requires neither Python nor diagnostic captures. Activation substitutions are optional package-scoped test hooks and default to no substitution.