{ "admission_policy": { "min_agreement": 0.6, "min_reading_arms": 2, "source": "ADL/O frozen operating point, as carried in probe/run_gate_evaluators.py OPERATING_POINT", "require_judge": false, "require_judge_note": "Certify scores consensus admission; live judge is enforced at backfill serve (PRODUCT_POLICY require_judge=True). Phase 4b." }, "arms": [ { "container": "python:3.11@sha256:143c7a1877acd77bd323d544cb7b39ba0a975288a7204f47586c2597130d80ad", "deps": [ "paddlepaddle-gpu==3.2.1", "paddleocr[doc-parser]==3.6.0", "huggingface_hub==1.24.0" ], "model": "PaddlePaddle/PaddleOCR-VL", "name": "paddle", "pin_note": "huggingface_hub was unpinned as run; 1.24.0 recovered from the job log of 6a66e4badb23d7a7ec1cf494", "revision": "f54aa90d389e98361cf295b7f4544bfb7452996d" }, { "container": "vllm/vllm-openai@sha256:b7a7be708c9a325107cdeddeba095e5637617716b3ab469c19d34759ec1afa39", "deps": [ "huggingface_hub==1.25.1" ], "model": "zai-org/GLM-OCR", "name": "glm_whole", "pin_note": "inference is vLLM in-container against a pinned model revision; the dependency only downloads the job script", "prompt": "Table Recognition:", "revision": "ca5d8b3e287e52589e37c28385d9655ee4372f9d", "serving": "vllm max_model_len=16384 max_tokens=8192" }, { "container": "python:3.11@sha256:143c7a1877acd77bd323d544cb7b39ba0a975288a7204f47586c2597130d80ad", "deps": [ "torch==2.6.0", "torchvision==0.21.0", "accelerate==1.6.0", "addict==2.4.0", "easydict==1.13", "einops==0.8.1", "huggingface_hub==0.34.0", "matplotlib==3.10.1", "numpy==2.2.4", "pillow==11.1.0", "requests==2.32.3", "sentencepiece==0.2.0", "tqdm==4.67.1", "transformers==4.51.3" ], "mode": "gundam base=1024 image=640 max_length=32768", "model": "baidu/Unlimited-OCR", "name": "unlimited", "pin_note": "fully pinned as run", "prompt": "document parsing.", "revision": "ee63731b6461c8afcdcc7b15352e7d2ffecc2ead" }, { "container": "vllm/vllm-openai@sha256:b7a7be708c9a325107cdeddeba095e5637617716b3ab469c19d34759ec1afa39", "deps": [ "glmocr[selfhosted]==0.1.5", "huggingface_hub==1.25.1", "pymupdf==1.28.0", "pypdfium2==5.12.1", "opencv-python-headless==5.0.0.93", "requests==2.34.2" ], "model": "zai-org/GLM-OCR", "name": "glm_pipe", "pin_note": "the ADL/M run left all six unpinned and redirected pip output inside the container, so its versions were unrecoverable. Re-run pinned (adlp-glmpipe-pinned-193133) with pip freeze printed to the job log so the pin stays recoverable. glmocr 0.1.5 is the only candidate: it is the newest release and nothing has shipped since 2026-04-08.", "revision": "ca5d8b3e287e52589e37c28385d9655ee4372f9d" } ], "denominator": "tolerant \u2014 see evidence/adl-n/decision-denominator.json (ADL/AZ #93)", "excluded_arms": [ { "endpoint": "https://api.z.ai/api/paas/v4/layout_parsing", "measured_cost": "6 cells of union keyed coverage on dev50 (0.4167 -> 0.4134); it is the only arm keying those 6.", "model": "z.ai glm-ocr", "name": "glm_api", "why": "A vendor endpoint has no container or dependency set and the served model may change without notice, so it cannot satisfy ADL/P's acceptance that every runtime identity is pinned to an immutable digest or revision." } ], "frozen_by": "Claude, under authority delegated by the owner on 2026-07-30", "frozen_on": "2026-07-30", "previously_provisional_because": "57.9% of gold cells read by no arm at the keyed level \u2014 re-measured at 0.5833 across five arms and re-characterised as row-wording divergence under a strict denominator no gate uses (ADL/AZ #93)", "rationale": "Highest measured keyed fidelity among the routes considered, and the only option yielding a confidence signal that ADL/O turns into abstention. Arm ranking on the re-golded fifty (evidence/adl-p/arm-comparison.json): paddle leads association 0.9880 and judged 0.7432; glm_whole 0.9591; unlimited 0.9223; glm_pipe 0.8806.", "route_kind": "cell_level_ensemble", "status": "FROZEN", "supersedes": "the PROVISIONAL document of the same name" }