{ "receipt": "exhibit thirteen — runtime pins, cloud shelf, metered output probes", "cut_at": "2026-08-16T14:48:09Z", "roster": [ { "arm_id": "local-qwen3.8-27b", "tag": "qwen3.8:27b", "vendor": "alibaba", "family": "qwen", "transport": "local-ollama", "size_class": "local ~27b", "metered": false, "sub_cap_usd": null, "note": "" }, { "arm_id": "local-qwen3.6-27b", "tag": "qwen3.6:27b", "vendor": "alibaba", "family": "qwen", "transport": "local-ollama", "size_class": "local ~27b", "metered": false, "sub_cap_usd": null, "note": "same family AND same size class as the arm above; the class column exists so a pooled figure cannot hide that." }, { "arm_id": "local-gemma4-26b", "tag": "gemma4:26b", "vendor": "google", "family": "gemma", "transport": "local-ollama", "size_class": "local ~26b", "metered": false, "sub_cap_usd": null, "note": "the-gpu-box's permanent resident pin (expires_at 2318); never released." }, { "arm_id": "local-llama3.3-70b", "tag": "llama3.3:70b", "vendor": "meta", "family": "llama", "transport": "local-ollama", "size_class": "local ~70b", "metered": false, "sub_cap_usd": null, "note": "seated by orchestrator ruling 1; the meta family's first seat on this instrument. A different size class from the other three local arms, which is why the class column is per-arm." }, { "arm_id": "cloud-glm-5.2", "tag": "glm-5.2", "vendor": "zhipu", "family": "glm", "transport": "ollama-cloud", "size_class": "frontier cloud", "metered": false, "sub_cap_usd": null, "note": "plan-included: $0.00 per token, real subscription." }, { "arm_id": "cloud-deepseek-v4-pro", "tag": "deepseek-v4-pro:preview", "vendor": "deepseek", "family": "deepseek", "transport": "ollama-cloud", "size_class": "frontier cloud", "metered": false, "sub_cap_usd": null, "note": "the tag carries ':preview'; the plan's earlier draft wrote it without, which is exactly the drift runtime_pins exists to catch." }, { "arm_id": "cloud-kimi-k3", "tag": "kimi-k3", "vendor": "moonshot", "family": "kimi", "transport": "ollama-cloud", "size_class": "frontier cloud", "metered": true, "sub_cap_usd": 0.9, "note": "METERED at 3.00/0.30/15.00 per Mtok. An earlier draft filed it as a plan seat, which is what made the $4 ceiling look comfortable." }, { "arm_id": "openai-gpt-5.5", "tag": "gpt-5.5-2026-04-23", "vendor": "openai", "family": "gpt", "transport": "openai-api", "size_class": "frontier cloud", "metered": true, "sub_cap_usd": 2.6, "note": "METERED at 5.00/30.00 per Mtok; reasoning tokens are not suppressible and bill as output, which is the least certain term in the budget." } ], "call_arithmetic": { "arms": 8, "families": 7, "conditions": 3, "items": 24, "scored_cells": 576, "warmups_discarded": 24, "flake_probe": 72, "metered_output_probes": 2, "total_wire_calls": 674 }, "local_pins": { "host": "http://the-gpu-box[:port redacted]", "ollama_version": "0.32.13", "read_at": "2026-08-16T14:48:09Z", "kind": "TRUE PIN — read before the first scored call, not after", "num_ctx": 32768, "think": false, "temperature": 0.0, "keep_alive": "never sent (house law 3)", "arms": [ { "arm": "local-qwen3.8-27b", "tag": "qwen3.8:27b", "present": true, "digest": "22130167c4c20e20c7b71454612966ca8e8171e9b3cc8ab6ce8aa6cbfec79643", "size_bytes": 17741872154, "modified_at": "2026-08-15T19:43:41.98159097-06:00", "parameter_size": "27.3B", "quantization_level": "Q4_K_M" }, { "arm": "local-qwen3.6-27b", "tag": "qwen3.6:27b", "present": true, "digest": "a50eda8ed977ab48a12431878896b27ffd5cef552c17af3317d9623b939a7f1e", "size_bytes": 17420432739, "modified_at": "2026-08-12T07:47:35.094639204-06:00", "parameter_size": "27.8B", "quantization_level": "Q4_K_M" }, { "arm": "local-gemma4-26b", "tag": "gemma4:26b", "present": true, "digest": "5571076f3d70050487b26b341705799e0ab29b808164f90d20d4cf84f699d251", "size_bytes": 17987581215, "modified_at": "2026-07-24T09:26:06.836187098-06:00", "parameter_size": "25.8B", "quantization_level": "Q4_K_M" }, { "arm": "local-llama3.3-70b", "tag": "llama3.3:70b", "present": true, "digest": "a6eb4748fd2990ad2952b2335a95a7f952d1a06119a0aa6a2df6cd052a93a3fa", "size_bytes": 42520413916, "modified_at": "2026-08-15T15:20:09.027102751-06:00", "parameter_size": "70.6B", "quantization_level": "Q4_K_M" } ] }, "cloud_shelf": { "read_at": "2026-08-16T14:48:10Z", "ollama_cloud_tags_visible": 19, "arms": [ { "arm": "cloud-glm-5.2", "tag": "glm-5.2", "present": true, "shelf": "ollama.com" }, { "arm": "cloud-deepseek-v4-pro", "tag": "deepseek-v4-pro:preview", "present": true, "shelf": "ollama.com" }, { "arm": "cloud-kimi-k3", "tag": "kimi-k3", "present": true, "shelf": "ollama.com" }, { "arm": "openai-gpt-5.5", "tag": "gpt-5.5-2026-04-23", "present": true, "shelf": "api.openai.com", "http_status": 200, "model_echo": "gpt-5.5-2026-04-23" } ] }, "slice_receipt": { "built_at": "2026-08-16T14:29:47Z", "reference_tokenizer": { "host": "http://localhost[:port redacted]", "model": "gemma4:12b", "digest": "4eb23ef187e2c5462566d6a1d3bbbc2f1346d0b4327cbb66d58fffbcc9b2b05c", "ollama_version": "0.31.1", "num_ctx": 32768, "num_predict": 1, "think": false, "template_overhead_tokens": 13, "note": "prompt_eval_count includes the chat template. Both blocks are measured the same way, so it is a constant on both sides of the comparison; it is published here so a reader can subtract it." }, "c_map": { "path": "corpus/blocks/C-MAP.txt", "rule": "apex llms.txt + one blank line + research llms.txt; no added prose", "parts": [ { "piece": "apex/llms.txt", "bytes": 2072, "tokens_measured_alone": 584 }, { "piece": "research/llms.txt", "bytes": 10007, "tokens_measured_alone": 2639 } ], "bytes": 12080, "tokens": 3211, "sha256": "c74cc74b81e7f75b63438e7ca7397d0a72da7c1dac5a52d1aef376e9e0306fd5", "note": "the sum of the parts measured alone is not the whole: the chat template is counted once for the concatenation and once per part." }, "c_html": { "path": "corpus/blocks/C-HTML.txt", "extractor_module_sha256": "a4aee74b2a6ba70f5a1b496339f99232003a97fe588c01bc5b38620a99ca1711", "extractor_rules": [ "1. ``