NOTE (2026-08-21, at publication): the quoted journal lines below are the run's own captured telemetry; their clock readings are re-expressed in UTC (same instants) because this site publishes UTC only. Counts, statuses, endpoints, and epochs are untouched; the lines no longer byte-match the journal they quote, and this note is the disclosure. # Contention evidence for the scored window 2026-08-21T07:22:17Z - 07:30:10Z captured_utc: 2026-08-21T07:29:56Z journal window passed as epoch seconds 1787296937 - 1787297410 (journalctl reads bare --since/--until on the host clock, so '@' is the unambiguous form) ## Every 'starting llama-server' inside the scored window Aug 21 07:22:25 the-gpu-box ollama[449651]: time=2026-08-21T07:22:25.315Z level=INFO source=llama_server.go:431 msg="starting llama-server" cmd="/usr/local/lib/ollama/llama-server --model /usr/share/ollama/.ollama/models/blobs/sha256-1278394b693672ac2799eadc9a83fd98259a6a88a40acfb1dcaa6c6fc895a606 --port 33053 --host 127.0.0.1 --no-webui --offline -c 32768 -np 1 Aug 21 07:22:34 the-gpu-box ollama[449651]: time=2026-08-21T07:22:34.411Z level=INFO source=llama_server.go:431 msg="starting llama-server" cmd="/usr/local/lib/ollama/llama-server --model /usr/share/ollama/.ollama/models/blobs/sha256-7121486771cbfe218851513210c40b35dbdee93ab1ef43fe36283c883980f0df --port 44893 --host 127.0.0.1 --no-webui --offline -c 32768 -np 1 Aug 21 07:25:04 the-gpu-box ollama[449651]: time=2026-08-21T07:25:04.563Z level=INFO source=llama_server.go:431 msg="starting llama-server" cmd="/usr/local/lib/ollama/llama-server --model /usr/share/ollama/.ollama/models/blobs/sha256-5c19f6282f4fc51cb114cb6c876d70ca2fc3b9cf0fbd0a018d9908f4fe1f63b3 --port 45657 --host 127.0.0.1 --no-webui --offline -c 32768 -np 1 ## Every non-/api/generate, non-/api/ps request inside the scored window Aug 21 07:22:34 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:22:34 UTC | 200 | 1.271279ms | [private-peer] | GET "/api/tags" Aug 21 07:22:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:22:35 UTC | 200 | 907.966µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:22:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:22:35 UTC | 200 | 1.148038ms | 127.0.0.1 | GET "/api/tags" Aug 21 07:22:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:22:35 UTC | 200 | 788.693µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:24:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:24:35 UTC | 200 | 741.824µs | [private-peer] | GET "/api/tags" Aug 21 07:24:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:24:35 UTC | 200 | 737.712µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:24:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:24:35 UTC | 200 | 678.93µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:24:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:24:35 UTC | 200 | 680.736µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:26:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:26:35 UTC | 200 | 805.523µs | [private-peer] | GET "/api/tags" Aug 21 07:26:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:26:35 UTC | 200 | 774.634µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:26:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:26:35 UTC | 200 | 776.637µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:26:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:26:35 UTC | 200 | 740.994µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:28:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:28:35 UTC | 200 | 780.506µs | [private-peer] | GET "/api/tags" Aug 21 07:28:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:28:35 UTC | 200 | 670.374µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:28:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:28:35 UTC | 200 | 646.782µs | 127.0.0.1 | GET "/api/tags" Aug 21 07:28:35 the-gpu-box ollama[449651]: [GIN] 2026/08/21 - 07:28:35 UTC | 200 | 717.146µs | 127.0.0.1 | GET "/api/tags" ## Request-path histogram inside the scored window 92 /api/generate 16 /api/tags 12 /api/ps ## Blob-digest attribution — which tag each of the 3 loads was Read from the runtime's own manifests (the 'model' layer digest per tag): gemma4:12b -> 1278394b693672ac gemma4:26b -> 7121486771cbfe21 nemotron-3.5-lightning:30b-a3b -> 5c19f6282f4fc51c qwen3.8:27b -> f5f1dd8920d417aa 07:22:25Z load of 1278394b... = gemma4:12b — THIS RUN's preflight 07:22:34Z load of 7121486771. = gemma4:26b — THIS RUN's preflight 07:25:04Z load of 5c19f6282f. = the a3b visitor — THIS RUN, mid-run, exactly where the protocol places it qwen3.8:27b was already resident before the window opened and shows NO load line at all (its per-call load_duration stayed at 0.275-0.299 s, the established no-relaunch signature). ## Accounting — every request in the window is this run's 92 /api/generate = 80 scored + 8 warm-ups + 3 preflight + 1 visitor release 12 /api/ps = 8 per-arm seating + pre-run + non-eviction + end-state + 1 operator spot-check taken mid-run 16 /api/tags = a metadata health poll on a 2-minute cadence (07:22:34/35, 07:24:35, 07:26:35, 07:28:35 x4 each). No GPU work: these read the model list and return in 0.6-1.3 ms. ZERO foreign /api/generate. ZERO foreign model loads. The window is CLEAN; contention_receipt.py's 'clean: false' flag fires on ANY load in the window and all three loads here are this run's own, itemised above.