OCR latent compression

Grading pipeline

Architecture review

A trusted reviewer reads the submission and trajectory and fails it on any of these checks.

1. learned_discrete_visual_codes
   Trace POST /encode through the supplied inference code. It must run a neural network on the image pixels and quantize the visual features into a sequence of codes. FSQ, LFQ, learned codebooks, residual or product quantization, and equivalent discretizations are valid; quantization levels may be fixed or learned, and multiple codebooks are allowed. The declared codebook_size describes the serialized integer alphabet used to count the codes against the 16,384-bit (2 KB) budget. The codes are learned visual codes: no text recognition or OCR engine, no language-model text generation, no transcription to characters, words or text-tokenizer ids, and no DOM/HTML/markup extraction runs anywhere in /encode, and the code list is not a text-token sequence or a compressed transcript. Deterministic image preprocessing and learned entropy models over the codes are allowed. A visual quantizer whose codes end up correlating with glyphs is allowed; an explicit transcription step, even if followed by a learned code, is not. Judge operations actually run on the image, not imported libraries or unused methods; loading a complete vision-language model is allowed when /encode only executes its visual path.
2. decoder_generates_text_from_codes_only
   POST /decode maps the code list through embeddings, scalar reconstruction or another representation into a neural decoder (any language decoder, fine-tuned or frozen with adapters) that generates the returned text. The code list is the only per-image information reaching /decode: no files, caches, databases, sockets, environment variables, clocks, request order or process state carry information between requests or between the encode and decode instances, nothing is precomputed from evaluation images, and no dictionary of stored transcripts or retrieval over a text corpus produces or completes the output. Decoding is deterministic (greedy or fixed-seed) for a given code list.
3. served_weights_are_the_trained_model
   Provenance and the actual checkpoint inventory establish that the served encoder, decoder and any learned quantizer parameters are the trained weights described by the training code, configurations and logs, that fixed quantization rules match the supplied code and configuration, that the declared codebook_size matches the serialized integer alphabet, and that any starting checkpoint is declared. Frozen backbones with trained adapters, merged or fully trained checkpoints, and inference exports are all valid. Trace serve.sh through both endpoints to the supplied weights and any runtime conversion; there is no decoy checkpoint or alternate model.