{ "schema": "anvil-serving.workload-manifest/v1", "campaign": "ThinkingCap RTX 5090 128K context qualification", "identity": { "model": "bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ", "revision": "f8fe157f207a13f977bc3d620ce10a3e9ba5ab11", "runtime_image": "sha256:082ca6f035279109041ffd3fe0695cb568b29bc580b35c4f297a66a08b216c1b" }, "selected_plan": { "context_tokens": 131072, "C": 1, "attention": "TRITON", "kv": "BF16", "ssm": "default FP32", "mtp": 3, "memory_fraction": 0.93, "offload_backend": "uva", "cpu_offload_gb": 3, "cpu_offload_params": "generic parameters" }, "fixtures": { "image_suite": { "path": "tests/fixtures/eval-data/agents-a1-v1/qwen38-image-ab.json", "sha256": "ac06eb4c295e1fbd91d582bb1330ea55a70ea775f291e2c323fd4fc4f00c45db", "provenance": { "source": "agents-a1-v1 synthetic image corpus subset", "license": "CC0-1.0", "observed_date": "2026-08-15" }, "assets": [ { "path": "tests/fixtures/eval-data/agents-a1-v1/scene.png", "sha256": "30a1f11e40bf4e0a8595d39828e0aa362e8b28ac683c404a1068cc20bf7f7bda" }, { "path": "tests/fixtures/eval-data/agents-a1-v1/ocr.png", "sha256": "a8d5158019fb2b08ee9f6a153860ab2342384644f7c53ab58fdd45658124e83e" }, { "path": "tests/fixtures/eval-data/agents-a1-v1/chart.png", "sha256": "c905e453c4e16eb274e46992d1a901bbced8f2314b47b9d408971bb176ab2f3e" }, { "path": "tests/fixtures/eval-data/agents-a1-v1/ui.png", "sha256": "03fd14d8b68973c915de2848736d22eb6f67ba6bdb5da2a7a563085ac894e804" }, { "path": "tests/fixtures/eval-data/agents-a1-v1/spatial.png", "sha256": "d48f47cfba9d65524d6a9ca37d593550e10400c1b5fa160bec8aea24e0164051" }, { "path": "tests/fixtures/eval-data/agents-a1-v1/compare-a.png", "sha256": "7dac4f31d62c88d765150aa6051efea9ddaaa8c64312080be4c5b8f6ad96322a" }, { "path": "tests/fixtures/eval-data/agents-a1-v1/compare-b.png", "sha256": "8f7cdfd733f1ee9731de6160e77af6aa50189edec6eaf9ddd7654a0c8300a43e" } ] }, "text_quality_suite": { "path": "tests/fixtures/eval-data/hf-mmlu-pro-10-repeated.suite.json", "sha256": "f6112c6f592c5285f8dae86808285f328161d3b4ae30ef660440d173d47b0636", "dataset": "TIGER-Lab/MMLU-Pro", "dataset_revision": "b189ec765aa7ed75c8acfea42df31fdae71f97be", "config": "default", "split": "validation", "rows": [ 5, 10, 15, 20, 25, 30, 35, 40, 45, 65 ], "license": "not recorded in retained local fixture", "selection": "Existing fixed July 12 fixture selection; it was not chosen from this campaign\u2019s results." }, "media_bytes_retained": false, "asset_policy": "Fixture paths and hashes are retained; media bytes are not copied into this evidence bundle." }, "controls": { "functional": { "planned_observations": 27, "completed": "27/27 passed", "native": "native/mtp-uva3/functional.json" }, "reasoning": { "planned_observations": 9, "completed": "9/9 passed; thinking enabled and reasoning channel observed", "native": "native/mtp-uva3/reasoning.json" }, "vision": { "planned_attempts": 18, "completed": "18/18 passed", "native": "native/mtp-uva3/vision.json" }, "quality": { "workload": "10 checks x 3 repetitions; temperature 0; 2048 visible plus 8192 reasoning", "completed": "30/30 passed; all finish reasons stop", "native": "native/mtp-uva3/quality.json", "evidence_scope": "repeated deterministic-marker diagnostic; not a broad quality ranking" }, "capacity": { "workload": "100 requests, nominal 4K, C1, 32 code words, canary, 45 output tokens, thinking disabled", "completed": "100/100 strict passed; zero errors; all requests performance eligible", "metrics": { "decode_tok_s_p50": 51.88547695919587, "ttft_p50_ms": 3003.0448000179604, "e2e_p50_ms": 3852.9262999945786, "aggregate_output_tok_s": 11.678387852350253, "completion_tokens_p50": 45.0 }, "native": "native/mtp-uva3/capacity.json", "scope": "Matched short-output C1 population; no full-context claim." }, "context": { "workload": "5 cases x 2 buckets (32768, 125952) x 3 positions x 2 repetitions = 60; 4096 total-completion reserve", "completed": "60/60 passed; all finish reasons stop; actual inputs 32658\u201332689 and 125838\u2013125869", "summary": "context-full-summary.json", "native": "native/mtp-uva3-context-full/artifact.json", "scope": "Text-only C1 matrix; exact-limit and one-over admission, C2, and mixed text-plus-image qualification were not run." }, "context_scout": { "completed": "2/2 passed; tokenizer-observed inputs 32658 and 125838; finish reason stop", "native": "native/mtp-uva3-context-scout/artifact.json", "summary": "context-scout-summary.json", "scope": "Two-request text-only diagnostic scout at 32K and near-limit input; not the full matrix, an admission-boundary test, C2, or mixed text-plus-image qualification." }, "agentic": { "workload": "2 synthetic cases", "completed": "2/2 passed; six API turns and four tool calls; finish reasons tool_calls=4 and stop=2", "summary": "agentic-summary.json", "native": "native/mtp-uva3-agentic/artifact.json", "scope": "Synthetic tool workflow only; not SWE or repository execution." } }, "status": "MTP-UVA3 current branch completed bounded direct 128K gates; no route/client/deployment promotion and no mixed-context, C2, SWE, video, or soak claim.", "current_branch_observation": { "gpu_weights_gib": 19.11, "generic_uva_offload_gib": 3.02, "included_visual_offload_gib": 0.86, "kv_cache_gib": 9.7, "reported_kv_capacity_tokens": 138519, "identity_binding": "identity-mtp-uva3.json plus native/mtp-uva3/startup.log and health.json; native source_recipe remains null" }, "prior_no_mtp_branch": { "status": "failed repeated quality gate 27/30", "native": "native/candidate/quality.json", "note": "Historical negative case; not combined with the current MTP branch." }, "product_gaps": [ "product-gaps/2026-09-24-benchmark-evidence-job-wrapper.md", "product-gaps/2026-09-24-mixed-context-vision-gate.md" ], "prior_no_mtp_startup_observation": { "health_and_identity": "passed", "reported_kv_capacity_tokens": 134138, "gpu_weights_gib": 20.49, "actual_offloaded_visual_weights_gib": 0.86, "evidence": [ "calibration.json", "native/candidate/startup.log", "native/candidate/startup-health.json", "native/candidate/startup-status.json" ], "branch": "prior rejected no-MTP visual-only UVA branch", "interpretation": "Historical startup observation only; it does not describe the current MTP-UVA3 configuration." }, "prior_no_mtp_host_observation": { "guard": "idle-only after the model was loaded; no force, restart, or settings change", "operation": "managed host cache reclaim", "page_cache_gib": "12.9 -> 2.1", "settled_at": "2026-09-24T04:51:31Z", "windows_free_memory_kib": { "physical": 12362204, "virtual": 13567872 }, "evidence": [ "native/host/cache-reclaim-preview.json", "native/host/cache-reclaim-apply.json" ], "settled_memory_export": "native/prior-host-settled/host-memory.json", "interpretation": "Prior clean file-cache reclaim only; it did not delete model weights and does not establish current MTP headroom.", "branch": "prior rejected no-MTP visual-only UVA branch" }, "current_mtp_host_observation": { "branch": "current MTP-UVA3 generic-UVA branch", "operation": "managed guarded idle cache reclaim", "settled_at": "2026-09-24T05:10:17Z", "windows_free_memory_kib": { "physical": 8417472, "virtual": 10163116 }, "native": "native/mtp-uva3/host-memory-after-reclaim.json", "interpretation": "Current branch observation after reclaim; it does not assert a full-context memory reserve." } }