{
	"schemaVersion": "portfolio.llmtracefx-evidence-article-export/v3",
	"source": {
		"repository": "Siddhant-K-code/LLMTraceFX",
		"commit": "266b42bc7659b3ab147f199f293bfc033324f763",
		"pullRequestUrl": "https://github.com/Siddhant-K-code/LLMTraceFX/pull/61",
		"mergedAt": "2026-09-04T07:22:41Z",
		"directory": "examples/evidence-catalog"
	},
	"canonicalArtifacts": {
		"catalog.json": {
			"path": "examples/evidence-catalog/catalog.json",
			"sha256": "f155ae638a82fc1378afa5a4a28111fc3bfe39e2f6c44bc4be58e3a2d4be2004"
		},
		"catalog.schema.json": {
			"path": "examples/evidence-catalog/catalog.schema.json",
			"sha256": "6e8613f904c72964758274543eec5fa66a155af487cb86dadacb70b48975d65d"
		},
		"claim-matrix.json": {
			"path": "examples/evidence-catalog/claim-matrix.json",
			"sha256": "6aa3710029e0bab252d8cc575f8307d7135cb0ad5740a806c9dec780758b18b7"
		},
		"graph.json": {
			"path": "examples/evidence-catalog/graph.json",
			"sha256": "40e32f86a962de8927b988d4e060010f81c85b3085cbfd736bcdf41cdbbeeeaa"
		},
		"graph.svg": {
			"path": "examples/evidence-catalog/graph.svg",
			"sha256": "c2cdc7d2593a4e80f8edcbae31655ee18ff04966859a3c33907d36e7fd0d08d6"
		},
		"index.html": {
			"path": "examples/evidence-catalog/index.html",
			"sha256": "47b1bf72ab5290633b830e4c5cee39478749cf33b2b20fa2d9702977f3daa5ac"
		},
		"registry.json": {
			"path": "examples/evidence-catalog/registry.json",
			"sha256": "c09ffb9fd3dbb31fe45a146bf3b300249a31e420cd9a61134715577324cbbd90"
		},
		"SHA256SUMS": {
			"path": "examples/evidence-catalog/SHA256SUMS",
			"sha256": "d77bbb2715be8cee16065c6299887734097c921eb39cbc0d05077f04a605b002"
		}
	},
	"catalog": {
		"schemaVersion": "1",
		"catalogHash": "sha256:31062849a9ef988a1f5fbcf39abb39a7f2113162e4616be5cc5270f281bb2ce8",
		"registryHash": "sha256:c890bfb783125536886b228c6236b6acfceb73e40c510ab8e6b1b5655ead309e",
		"generator": {
			"determinism": "source-only; no clock, cwd, host, network, or model state",
			"name": "llmtracefx-evidence",
			"version": "1"
		},
		"entryCount": 9,
		"edgeCount": 6,
		"claimCellCount": 63,
		"claimDimensions": [
			"timing",
			"quality",
			"cost",
			"memory",
			"process_attribution",
			"model_fit",
			"deployment_readiness"
		],
		"evidenceKinds": [
			"compile_break_even",
			"fit_frontier",
			"hosted_comparison",
			"metal_attribution",
			"model_lab",
			"oom_autopsy",
			"positive_control",
			"provider_preflight"
		],
		"unregisteredCandidateCount": 0
	},
	"adapters": {
		"cloudrift_compile_v1": {
			"name": "cloudrift-compile.verify",
			"version": "1"
		},
		"cloudrift_preflight_v1": {
			"name": "cloudrift-preflight.verify",
			"version": "1"
		},
		"legacy_pinned_v1": {
			"name": "historical-immutable-artifact-set",
			"version": "1"
		},
		"metal_public_v1": {
			"name": "metal-evidence.verify_public_bundle",
			"version": "1"
		},
		"modal_preflight_v1": {
			"name": "modal-preflight.verify",
			"version": "1"
		},
		"oom_autopsy_v1": {
			"name": "oom-autopsy.verify_bundle",
			"version": "1"
		},
		"openrouter_glm_v1": {
			"name": "openrouter-glm.verify",
			"version": "1"
		},
		"sha256_allowlist_v1": {
			"name": "sha256-allowlist",
			"version": "1"
		}
	},
	"entries": [
		{
			"artifact_set_hash": "sha256:e7b5ae6331089f945941126aeb261d201d53db4e1176b245ba7535271d0abf0e",
			"budget": {
				"authorized_usd": 80,
				"inferred_usd": 0,
				"limitation": "No provider billing result was available without access.",
				"planned_usd": 60,
				"reported_usd": null,
				"scope": "planned_caps_and_inferred_zero_spend"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-09-02T22:38:00+05:30",
			"claims": {
				"cost": {
					"provenance": "planned caps and inferred zero spend; no provider usage",
					"state": "supported"
				},
				"deployment_readiness": {
					"provenance": "preflight refusal and exact stop gates are recorded",
					"state": "supported"
				},
				"memory": {
					"provenance": "aggregate listed V100 memory and exact model inventory comparison",
					"state": "supported"
				},
				"model_fit": {
					"provenance": "available V100 memory was below the exact model inventory",
					"state": "supported"
				},
				"process_attribution": {
					"provenance": "no process ran",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "no model output exists",
					"state": "unsupported"
				},
				"timing": {
					"provenance": "no deployment or request ran",
					"state": "unsupported"
				}
			},
			"dependencies": [
				{
					"evidence_id": "modal-glm53flash-preflight-20260902",
					"relation": "same_model_as"
				}
			],
			"evidence_id": "cloudrift-glm53flash-preflight-20260902",
			"hardware": {
				"architecture": null,
				"system": "observed 8x V100; required 8x H200 unavailable"
			},
			"kind": "provider_preflight",
			"limitations": [
				"No provisioning, model load, readiness probe, or smoke request ran.",
				"The required H200 inventory and official runtime recipe were unverified."
			],
			"measurements": [
				{
					"provenance": "user-observed aggregate; identifiers removed",
					"scope": "observed console inventory"
				},
				{
					"provenance": "inferred zero because no resource was created",
					"scope": "experiment-attributable spend"
				}
			],
			"model": {
				"id": "zai-org/GLM-5.3-Flash",
				"quantization": "native FP8 e4m3 with dynamic activation scaling",
				"revision": "03eb5366286afd40d2221b1d9c63a6dd1ba4832e"
			},
			"outcome": "refused",
			"public_path": "examples/optimizer/cloudrift-glm53flash-preflight",
			"runtime": {
				"name": "unselected; official recipe unverified",
				"provider": "CloudRift",
				"version": null
			},
			"source_commit": "0dbdcf5e745f123e13d38d09296f629c24abd748",
			"status": "verified",
			"supported_claims": [
				"Paid execution was refused by seven explicit stop gates.",
				"The observed V100 listing was below the exact model inventory.",
				"No CloudRift resource was created and spend is inferred zero."
			],
			"unsupported_claims": [
				"H200 fit, startup, readiness, benchmark, throughput, or SLA",
				"runtime utilization, bandwidth, power, or energy"
			],
			"verifier": {
				"name": "cloudrift-preflight.verify",
				"version": "1"
			},
			"workload": {
				"context": "no model request",
				"identity": "GLM-5.3-Flash provider preflight",
				"request": "no authenticated, provisioning, or paid execution"
			}
		},
		{
			"artifact_set_hash": "sha256:c615f77d6170eb8ac80ce54689feb68921e1f8acd827b44f72070fbb0cfb1035",
			"budget": {
				"authorized_usd": null,
				"inferred_usd": null,
				"limitation": "No cost measurement was in scope.",
				"planned_usd": null,
				"reported_usd": null,
				"scope": "not_applicable"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-08-31T13:18:34+00:00",
			"claims": {
				"cost": {
					"provenance": "local trace capture has no spend claim",
					"state": "not_applicable"
				},
				"deployment_readiness": {
					"provenance": "not a deployment run",
					"state": "not_applicable"
				},
				"memory": {
					"provenance": "GPU memory footprint was not measured",
					"state": "unsupported"
				},
				"model_fit": {
					"provenance": "no model was loaded",
					"state": "not_applicable"
				},
				"process_attribution": {
					"provenance": "target-process interval counts and trace-wide counts",
					"state": "supported"
				},
				"quality": {
					"provenance": "no model output was evaluated",
					"state": "not_applicable"
				},
				"timing": {
					"provenance": "interval counts are not utilization or time",
					"state": "unsupported"
				}
			},
			"dependencies": [],
			"evidence_id": "metal-attribution-m5-pro-20260831",
			"hardware": {
				"architecture": "arm64",
				"system": "Apple M5 Pro"
			},
			"kind": "metal_attribution",
			"limitations": [
				"Interval count is not utilization or elapsed GPU time.",
				"Only sanitized aggregate trace evidence is public."
			],
			"measurements": [
				{
					"provenance": "measured_native and target-process attributed",
					"scope": "Metal interval count"
				},
				{
					"provenance": "derived from measured interval counts",
					"scope": "unrelated interval share"
				}
			],
			"model": {
				"id": null,
				"quantization": null,
				"revision": null
			},
			"outcome": "completed",
			"public_path": "examples/metal_evidence/public",
			"runtime": {
				"name": "Apple Instruments Metal System Trace",
				"provider": "local",
				"version": "26.6"
			},
			"source_commit": null,
			"status": "verified",
			"supported_claims": [
				"Trace-wide Metal interval totals include unrelated processes.",
				"Target-process attribution matched each controlled dispatch count."
			],
			"unsupported_claims": [
				"GPU utilization or busy percentage",
				"kernel time, memory bandwidth, occupancy, power, or energy",
				"GPU memory footprint"
			],
			"verifier": {
				"name": "metal-evidence.verify_public_bundle",
				"version": "1"
			},
			"workload": {
				"context": null,
				"identity": "metal-evidence-workload",
				"request": "five controlled dispatch-count captures"
			}
		},
		{
			"artifact_set_hash": "sha256:f2956754d080064e3f941aeca0a7e5c902e2c84fb5927e395f6fd9ec0e860046",
			"budget": {
				"authorized_usd": 10,
				"inferred_usd": 0,
				"limitation": "No provider-reported usage was available.",
				"planned_usd": null,
				"reported_usd": null,
				"scope": "modeled_plan_and_inferred_zero_spend"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-09-02T21:57:45+05:30",
			"claims": {
				"cost": {
					"provenance": "modeled cost and inferred zero spend are distinct scopes",
					"state": "supported"
				},
				"deployment_readiness": {
					"provenance": "preflight refusal and exact stop gates are recorded",
					"state": "supported"
				},
				"memory": {
					"provenance": "no runtime memory was measured",
					"state": "unsupported"
				},
				"model_fit": {
					"provenance": "planned hardware fit was not proven",
					"state": "unsupported"
				},
				"process_attribution": {
					"provenance": "no process ran",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "no model output exists",
					"state": "unsupported"
				},
				"timing": {
					"provenance": "no deployment or request ran",
					"state": "unsupported"
				}
			},
			"dependencies": [],
			"evidence_id": "modal-glm53flash-preflight-20260902",
			"hardware": {
				"architecture": "linux/amd64 image manifest",
				"system": "planned 4x H200; not provisioned"
			},
			"kind": "provider_preflight",
			"limitations": [
				"No model load, deployment, readiness probe, or smoke request ran.",
				"The framework source revision remained unverified."
			],
			"measurements": [
				{
					"provenance": "pricing plan; not provider usage",
					"scope": "modeled lifecycle cost"
				},
				{
					"provenance": "inferred zero because no resource was created",
					"scope": "experiment-attributable spend"
				}
			],
			"model": {
				"id": "zai-org/GLM-5.3-Flash",
				"quantization": "native FP8 e4m3 with dynamic activation scaling",
				"revision": "03eb5366286afd40d2221b1d9c63a6dd1ba4832e"
			},
			"outcome": "refused",
			"public_path": "examples/optimizer/modal-glm53flash-preflight",
			"runtime": {
				"name": "vLLM",
				"provider": "Modal",
				"version": "unverified source revision candidate"
			},
			"source_commit": "debd8fa3f2d4bbed3ccdaad40fd1be80e264fe87",
			"status": "verified",
			"supported_claims": [
				"Paid execution was refused by three explicit stop gates.",
				"No Modal resource was created and attributable spend is inferred zero."
			],
			"unsupported_claims": [
				"hardware fit, startup, readiness, benchmark, or ranking",
				"runtime memory, utilization, bandwidth, power, or energy"
			],
			"verifier": {
				"name": "modal-preflight.verify",
				"version": "1"
			},
			"workload": {
				"context": "planned 131,072 tokens",
				"identity": "GLM-5.3-Flash deployment preflight",
				"request": "no authenticated or paid execution"
			}
		},
		{
			"artifact_set_hash": "sha256:90e2ec60eec7d5cda570f526cb7efe607a57af8c359b8bb66e27c391aeba4527",
			"budget": {
				"authorized_usd": 5,
				"inferred_usd": null,
				"limitation": "The local ledger is user-writable; account delta can lag and include unrelated activity.",
				"planned_usd": 0.0745326,
				"reported_usd": 0.00615262,
				"scope": "provider_reported_request_usage"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-09-02T11:44:23.771009Z",
			"claims": {
				"cost": {
					"provenance": "provider usage, plan, and account delta remain separate scopes",
					"state": "supported"
				},
				"deployment_readiness": {
					"provenance": "comparison does not establish production readiness",
					"state": "unsupported"
				},
				"memory": {
					"provenance": "provider memory was not exposed",
					"state": "unsupported"
				},
				"model_fit": {
					"provenance": "hosted completion does not prove hardware fit",
					"state": "unsupported"
				},
				"process_attribution": {
					"provenance": "hosted provider internals",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "pinned evaluators for eight completed requests",
					"state": "supported"
				},
				"timing": {
					"provenance": "client-observed hosted request timing",
					"state": "supported"
				}
			},
			"dependencies": [
				{
					"evidence_id": "qwen3-8b-m5-pro-control-20260902",
					"relation": "compares"
				}
			],
			"evidence_id": "openrouter-glm-2k-comparison-20260902",
			"hardware": {
				"architecture": null,
				"system": "provider managed and undisclosed"
			},
			"kind": "hosted_comparison",
			"limitations": [
				"Client timing includes network, gateway, queueing, and execution.",
				"Two repetitions per workload/model are directional evidence."
			],
			"measurements": [
				{
					"provenance": "network, gateway, queueing, and execution combined",
					"scope": "client timing"
				},
				{
					"provenance": "final SSE usage blocks",
					"scope": "provider request usage cost"
				},
				{
					"provenance": "separate lag-prone corroborating observation",
					"scope": "account usage delta"
				}
			],
			"model": {
				"id": "z-ai/glm-5.3 and z-ai/glm-5.3-flash",
				"quantization": "fp8",
				"revision": "z-ai/glm-5.3-20260816 and z-ai/glm-5.3-flash-20260826"
			},
			"outcome": "comparison",
			"public_path": "examples/optimizer/openrouter-glm-2k",
			"runtime": {
				"name": "OpenRouter hosted API",
				"provider": "OpenRouter / Z.AI",
				"version": "captured provider route"
			},
			"source_commit": "a6077adaf7135e2a2e360aeae4a73b6b411b3493",
			"status": "verified",
			"supported_claims": [
				"Eight pinned hosted requests completed and passed their evaluators.",
				"Provider-reported request usage totaled USD 0.00615262."
			],
			"unsupported_claims": [
				"universal winner or cross-system local ranking",
				"server-only timing, provider memory, or hardware fit",
				"production readiness"
			],
			"verifier": {
				"name": "openrouter-glm.verify",
				"version": "1"
			},
			"workload": {
				"context": "2K tier",
				"identity": "structured-json-profile-extraction@1 and prose-reasoning-two-train-gap@1",
				"request": "eight requests; low reasoning; no retries"
			}
		},
		{
			"artifact_set_hash": "sha256:4fe4c5521e21f23dc7c52f84aaa02fb51fe83513abe11dab9f966e4eabbe9d88",
			"budget": {
				"authorized_usd": 5,
				"inferred_usd": 0.484358,
				"limitation": "Provider spend and provisioning-to-boot duration are unavailable; the inferred spend is a lower bound and remaining cap is an upper bound.",
				"planned_usd": 3.12,
				"reported_usd": null,
				"scope": "boot_to_console_termination_list_rate_lower_bound"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-09-03T16:30:38.954381+00:00",
			"claims": {
				"cost": {
					"provenance": "boot-to-console list-rate inference; provider spend is unavailable",
					"state": "supported"
				},
				"deployment_readiness": {
					"provenance": "bounded benchmark is not a production deployment assessment",
					"state": "unsupported"
				},
				"memory": {
					"provenance": "sampled peak device memory for both cells",
					"state": "supported"
				},
				"model_fit": {
					"provenance": "both cells completed; per-cell source binding is limited",
					"state": "supported"
				},
				"process_attribution": {
					"provenance": "ordered non-overlapping cell processes",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "22 of 24 deterministic workload evaluator results passed",
					"state": "supported"
				},
				"timing": {
					"provenance": "measured initialization and 24 bounded request records",
					"state": "supported"
				}
			},
			"dependencies": [
				{
					"evidence_id": "qwen3-8b-m5-pro-control-20260902",
					"relation": "uses_workload_contract"
				}
			],
			"evidence_id": "qwen3-8b-cloudrift-vllm-compile-20260903",
			"hardware": {
				"architecture": "CUDA 13.0",
				"system": "NVIDIA GeForce RTX 4090, 24,564 MiB"
			},
			"kind": "compile_break_even",
			"limitations": [
				"The request-113 time crossing freezes and repeats observed outcomes. Output-token arrays and lengths differ at ordinals 7, 8, 11, and 12; correctness differs at ordinals 11 and 12.",
				"Compilation and CUDA graph component durations were not retained.",
				"Provider-reported spend and provisioning-to-boot cost are unavailable.",
				"The measured runner independently verified retained staging and prompt receipts, but did not rehash the live model or cross-check both receipt hashes before each cell.",
				"No independent host receipt was retained for fresh-container, cache-drop, timeout, bind-mount, network, or Docker image-inspection controls.",
				"MLX results are an incompatible scope and are not ranked here."
			],
			"measurements": [
				{
					"provenance": "client-observed and vLLM metrics",
					"scope": "initialization, TTFT, and complete response latency"
				},
				{
					"provenance": "sampled nvidia-smi used memory",
					"scope": "peak GPU memory"
				},
				{
					"provenance": "derived lower bound from user-observed console rate",
					"scope": "boot-to-console-termination list-rate lower bound"
				}
			],
			"model": {
				"id": "Qwen/Qwen3-8B",
				"quantization": "bfloat16",
				"revision": "b968826d9c46dd6066d109eabc6255188de91218"
			},
			"outcome": "comparison",
			"public_path": "examples/optimizer/qwen3-8b-vllm-compile-break-even",
			"runtime": {
				"name": "vLLM",
				"provider": "CloudRift",
				"version": "0.28.0"
			},
			"source_commit": "9c0879351cc3e4f294b5c827d74dfc00182d53bb",
			"status": "verified",
			"supported_claims": [
				"The compiled lifecycle did not cross the eager lifecycle time through request 12.",
				"Frozen exact-observed-outcome repeated-sequence time arithmetic crosses at request 113.",
				"Twenty-two of 24 bounded responses passed deterministic evaluation.",
				"Compiled output passed 12 of 12; eager output passed 10 of 12.",
				"Eight of 12 paired outputs had identical token IDs.",
				"Both ordered, non-overlapping cell processes completed and carried the same privacy-preserving run-time GPU attestation.",
				"The user externally confirmed provider-console termination."
			],
			"unsupported_claims": [
				"an output-controlled, causal, or general compilation break-even outside the frozen observed outcomes",
				"provider-reported spend or provisioning-to-boot cost",
				"independent verification of the private GPU identity or provider event",
				"production readiness, SLA, power, energy, or bandwidth",
				"direct component timing for compilation or CUDA graph capture"
			],
			"verifier": {
				"name": "cloudrift-compile.verify",
				"version": "1"
			},
			"workload": {
				"context": "2K, 8K, and 16K tiers; exact pinned input-token arrays",
				"identity": "qwen3-8b-vllm-compile-break-even-v1",
				"request": "12 requests per cell; 96 maximum output tokens"
			}
		},
		{
			"artifact_set_hash": "sha256:f80c18650d8d5e891957398d53058cd50182f2d5c8167775275cb8416af86335",
			"budget": {
				"authorized_usd": null,
				"inferred_usd": null,
				"limitation": "Local execution did not record cost.",
				"planned_usd": null,
				"reported_usd": null,
				"scope": "not_applicable"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-09-02T08:08:36.185127Z",
			"claims": {
				"cost": {
					"provenance": "local execution has no spend claim",
					"state": "not_applicable"
				},
				"deployment_readiness": {
					"provenance": "exploratory benchmark is not a deployment assessment",
					"state": "unsupported"
				},
				"memory": {
					"provenance": "MLX allocator counters; not RSS or swap",
					"state": "supported"
				},
				"model_fit": {
					"provenance": "this self-converted 8B checkpoint completed through requested 16K",
					"state": "supported"
				},
				"process_attribution": {
					"provenance": "single isolated row process",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "pinned evaluator pass rate and score",
					"state": "supported"
				},
				"timing": {
					"provenance": "host wall-clock prefill/decode/total",
					"state": "supported"
				}
			},
			"dependencies": [],
			"evidence_id": "qwen3-8b-m5-pro-control-20260902",
			"hardware": {
				"architecture": "arm64",
				"system": "Apple M5 Pro, 24 GiB unified memory"
			},
			"kind": "positive_control",
			"limitations": [
				"Exploratory run without a clean-boot assertion.",
				"Different model and system identity from every 27B result."
			],
			"measurements": [
				{
					"provenance": "measured_wall_clock",
					"scope": "host wall-clock timing"
				},
				{
					"provenance": "measured_native; not RSS or swap",
					"scope": "MLX active/cache/peak allocator counters"
				},
				{
					"provenance": "derived from measured counts and wall time",
					"scope": "decode rate and correct cases per minute"
				}
			],
			"model": {
				"id": "Qwen/Qwen3-8B",
				"quantization": "MLX affine 4-bit, group size 64; self-converted",
				"revision": "b968826d9c46dd6066d109eabc6255188de91218"
			},
			"outcome": "completed",
			"public_path": "examples/optimizer/qwen3-8b-m5-control",
			"runtime": {
				"name": "mlx-lm",
				"provider": "local",
				"version": "0.31.3"
			},
			"source_commit": "6b82cf276ee1e1cef03a0c92847082f872c8feba",
			"status": "verified",
			"supported_claims": [
				"The exact self-converted Qwen3-8B checkpoint completed all recorded tiers.",
				"All twelve evaluated rows passed their pinned evaluator."
			],
			"unsupported_claims": [
				"comparability with the Qwen3.8-27B checkpoint",
				"clean-boot publication performance",
				"GPU utilization, bandwidth, power, energy, or kernel timing"
			],
			"verifier": {
				"name": "sha256-allowlist",
				"version": "1"
			},
			"workload": {
				"context": "requested 2K, 8K, and 16K tiers",
				"identity": "pinned LLMTraceFX workload catalog",
				"request": "four measured runs per tier; thinking disabled"
			}
		},
		{
			"artifact_set_hash": "sha256:1096de514bcdd74201d367d098352da831efa9f21d85aa2993af0e2787c56030",
			"budget": {
				"authorized_usd": null,
				"inferred_usd": null,
				"limitation": "Local execution did not record cost.",
				"planned_usd": null,
				"reported_usd": null,
				"scope": "not_applicable"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-09-01T17:45:36.921331Z",
			"claims": {
				"cost": {
					"provenance": "local execution has no spend claim",
					"state": "not_applicable"
				},
				"deployment_readiness": {
					"provenance": "OOM autopsy does not assess deployment readiness",
					"state": "unsupported"
				},
				"memory": {
					"provenance": "separate MLX allocator, RSS, swap, and headroom scopes",
					"state": "supported"
				},
				"model_fit": {
					"provenance": "clean-boot OOM for the exact checkpoint and t256 workload",
					"state": "supported"
				},
				"process_attribution": {
					"provenance": "single autopsy child process",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "no first token or evaluator result exists",
					"state": "unsupported"
				},
				"timing": {
					"provenance": "stage-boundary host wall-clock observations",
					"state": "supported"
				}
			},
			"dependencies": [
				{
					"evidence_id": "qwen38-27b-m5-pro-lab-oom-20260831",
					"relation": "same_model_as"
				},
				{
					"evidence_id": "qwen38-27b-m5-pro-fit-frontier-20260901",
					"relation": "same_model_as"
				}
			],
			"evidence_id": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
			"hardware": {
				"architecture": "arm64",
				"system": "Apple M5 Pro, 24 GiB unified memory"
			},
			"kind": "oom_autopsy",
			"limitations": [
				"Bounded to one exact checkpoint, runtime, machine state, and workload.",
				"Discrete checkpoints do not provide causal allocation attribution."
			],
			"measurements": [
				{
					"provenance": "MLX allocator; non-additive with RSS and swap",
					"scope": "MLX active/cache/peak allocator counters"
				},
				{
					"provenance": "host process current/max RSS",
					"scope": "process RSS"
				},
				{
					"provenance": "sysctl and macOS memory_pressure",
					"scope": "system swap and approximate headroom"
				}
			],
			"model": {
				"id": "mlx-community/Qwen3.8-27B-4bit",
				"quantization": "MLX affine 4-bit, group size 64",
				"revision": "3e6447f082e89cc7f0bc6e5441afd38dfce760ff"
			},
			"outcome": "oom",
			"public_path": "examples/optimizer/m5-pro-qwen3.8-27b-oom-autopsy/publication",
			"runtime": {
				"name": "mlx-vlm",
				"provider": "local",
				"version": "0.6.8"
			},
			"source_commit": "2519bc8da309656d2e2ce2a7063f19b0dfb4c9ed",
			"status": "verified",
			"supported_claims": [
				"The exact checkpoint OOMed before first token at t256 after clean boot.",
				"MLX allocator, RSS, swap, and headroom scopes remain non-additive."
			],
			"unsupported_claims": [
				"universal memory-capacity or 24 GiB boundary",
				"causal allocation attribution",
				"quality, throughput, utilization, power, energy, or kernel time"
			],
			"verifier": {
				"name": "oom-autopsy.verify_bundle",
				"version": "1"
			},
			"workload": {
				"context": "t256; 256 actual prompt tokens",
				"identity": "m5-pro-qwen3.8-27b-oom-autopsy-v1",
				"request": "clean-boot publication autopsy"
			}
		},
		{
			"artifact_set_hash": "sha256:3585e3ccade81de3e91ee3b88c5be533a6c0e5a1fe28194d93639f6ec6af38b4",
			"budget": {
				"authorized_usd": null,
				"inferred_usd": null,
				"limitation": "Local execution did not record cost.",
				"planned_usd": null,
				"reported_usd": null,
				"scope": "not_applicable"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-09-01T05:30:24.941062Z",
			"claims": {
				"cost": {
					"provenance": "local execution has no spend claim",
					"state": "not_applicable"
				},
				"deployment_readiness": {
					"provenance": "exploratory frontier is not a deployment assessment",
					"state": "unsupported"
				},
				"memory": {
					"provenance": "approximate system headroom and swap; not GPU memory",
					"state": "supported"
				},
				"model_fit": {
					"provenance": "bounded t256 OOM in the recorded host state",
					"state": "supported"
				},
				"process_attribution": {
					"provenance": "single local model run",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "no quality result was produced",
					"state": "unsupported"
				},
				"timing": {
					"provenance": "host wall-clock total before OOM",
					"state": "supported"
				}
			},
			"dependencies": [
				{
					"evidence_id": "qwen38-27b-m5-pro-lab-oom-20260831",
					"relation": "same_model_as"
				}
			],
			"evidence_id": "qwen38-27b-m5-pro-fit-frontier-20260901",
			"hardware": {
				"architecture": "arm64",
				"system": "Apple M5 Pro, 24 GiB unified memory"
			},
			"kind": "fit_frontier",
			"limitations": [
				"Exploratory run without a clean-boot assertion.",
				"Available memory is approximate system headroom, not GPU memory."
			],
			"measurements": [
				{
					"provenance": "measured_wall_clock",
					"scope": "host wall-clock total"
				},
				{
					"provenance": "macOS memory_pressure and sysctl",
					"scope": "approximate system headroom and swap"
				}
			],
			"model": {
				"id": "mlx-community/Qwen3.8-27B-4bit",
				"quantization": "MLX affine 4-bit, group size 64",
				"revision": "3e6447f082e89cc7f0bc6e5441afd38dfce760ff"
			},
			"outcome": "oom",
			"public_path": "examples/optimizer/m5-pro-qwen3.8-27b-fit-frontier/exploratory",
			"runtime": {
				"name": "mlx-vlm",
				"provider": "local",
				"version": "0.6.8"
			},
			"source_commit": null,
			"status": "verified",
			"supported_claims": [
				"The exact checkpoint OOMed at t256 in the recorded machine state.",
				"Larger frontier tiers were skipped after the stop gate."
			],
			"unsupported_claims": [
				"universal memory-capacity boundary",
				"peak system or GPU memory",
				"quality, utilization, bandwidth, power, energy, or kernel time"
			],
			"verifier": {
				"name": "historical-immutable-artifact-set",
				"version": "1"
			},
			"workload": {
				"context": "t256 attempted; larger tiers skipped",
				"identity": "m5-pro-qwen3.8-27b-fit-frontier-v1",
				"request": "bounded exploratory fit-frontier row"
			}
		},
		{
			"artifact_set_hash": "sha256:bcd3b3765c2a61041cd2cab0ef2cddb0ca7532ad2c685cb69e07aac9147008e6",
			"budget": {
				"authorized_usd": null,
				"inferred_usd": null,
				"limitation": "Local execution did not record cost.",
				"planned_usd": null,
				"reported_usd": null,
				"scope": "not_applicable"
			},
			"bundle_schema_version": "1",
			"captured_at": "2026-08-31T14:57:58.660151Z",
			"claims": {
				"cost": {
					"provenance": "local execution has no spend claim",
					"state": "not_applicable"
				},
				"deployment_readiness": {
					"provenance": "a failed exploratory warmup is not a deployment assessment",
					"state": "unsupported"
				},
				"memory": {
					"provenance": "system memory pressure and swap; MLX peak is unavailable",
					"state": "supported"
				},
				"model_fit": {
					"provenance": "OOM during the exact recorded warmup and host state",
					"state": "supported"
				},
				"process_attribution": {
					"provenance": "single local model run",
					"state": "not_applicable"
				},
				"quality": {
					"provenance": "no evaluator result was produced",
					"state": "unsupported"
				},
				"timing": {
					"provenance": "host wall-clock total before OOM",
					"state": "supported"
				}
			},
			"dependencies": [],
			"evidence_id": "qwen38-27b-m5-pro-lab-oom-20260831",
			"hardware": {
				"architecture": "arm64",
				"system": "Apple M5 Pro, 24 GiB unified memory"
			},
			"kind": "model_lab",
			"limitations": [
				"Exploratory host state; no clean-boot assertion.",
				"Peak memory is unavailable for the failed attempt."
			],
			"measurements": [
				{
					"provenance": "measured_wall_clock",
					"scope": "host wall-clock total"
				},
				{
					"provenance": "macOS memory_pressure and sysctl",
					"scope": "system memory pressure and swap"
				}
			],
			"model": {
				"id": "mlx-community/Qwen3.8-27B-4bit",
				"quantization": "MLX affine 4-bit, group size 64",
				"revision": "3e6447f082e89cc7f0bc6e5441afd38dfce760ff"
			},
			"outcome": "oom",
			"public_path": "examples/optimizer/m5-pro-qwen3.8-27b",
			"runtime": {
				"name": "mlx-vlm",
				"provider": "local",
				"version": "0.6.8"
			},
			"source_commit": null,
			"status": "verified",
			"supported_claims": [
				"The exact 27B checkpoint OOMed during the recorded 2K warmup.",
				"No 8K or 16K tier was attempted after the failure."
			],
			"unsupported_claims": [
				"universal 24 GiB capacity boundary",
				"quality, throughput, GPU utilization, power, or kernel timing",
				"causal allocation attribution"
			],
			"verifier": {
				"name": "historical-immutable-artifact-set",
				"version": "1"
			},
			"workload": {
				"context": "2K tier; 1,657 actual input tokens",
				"identity": "structured-json-profile-extraction warmup",
				"request": "96 maximum output tokens"
			}
		}
	],
	"graph": {
		"nodes": [
			{
				"evidence_id": "cloudrift-glm53flash-preflight-20260902",
				"kind": "provider_preflight",
				"model": "zai-org/GLM-5.3-Flash",
				"outcome": "refused",
				"status": "verified",
				"supported_claim_scope": [
					"cost",
					"memory",
					"model_fit",
					"deployment_readiness"
				],
				"system": "observed 8x V100; required 8x H200 unavailable"
			},
			{
				"evidence_id": "metal-attribution-m5-pro-20260831",
				"kind": "metal_attribution",
				"model": null,
				"outcome": "completed",
				"status": "verified",
				"supported_claim_scope": [
					"process_attribution"
				],
				"system": "Apple M5 Pro"
			},
			{
				"evidence_id": "modal-glm53flash-preflight-20260902",
				"kind": "provider_preflight",
				"model": "zai-org/GLM-5.3-Flash",
				"outcome": "refused",
				"status": "verified",
				"supported_claim_scope": [
					"cost",
					"deployment_readiness"
				],
				"system": "planned 4x H200; not provisioned"
			},
			{
				"evidence_id": "openrouter-glm-2k-comparison-20260902",
				"kind": "hosted_comparison",
				"model": "z-ai/glm-5.3 and z-ai/glm-5.3-flash",
				"outcome": "comparison",
				"status": "verified",
				"supported_claim_scope": [
					"timing",
					"quality",
					"cost"
				],
				"system": "provider managed and undisclosed"
			},
			{
				"evidence_id": "qwen3-8b-cloudrift-vllm-compile-20260903",
				"kind": "compile_break_even",
				"model": "Qwen/Qwen3-8B",
				"outcome": "comparison",
				"status": "verified",
				"supported_claim_scope": [
					"timing",
					"quality",
					"cost",
					"memory",
					"model_fit"
				],
				"system": "NVIDIA GeForce RTX 4090, 24,564 MiB"
			},
			{
				"evidence_id": "qwen3-8b-m5-pro-control-20260902",
				"kind": "positive_control",
				"model": "Qwen/Qwen3-8B",
				"outcome": "completed",
				"status": "verified",
				"supported_claim_scope": [
					"timing",
					"quality",
					"memory",
					"model_fit"
				],
				"system": "Apple M5 Pro, 24 GiB unified memory"
			},
			{
				"evidence_id": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
				"kind": "oom_autopsy",
				"model": "mlx-community/Qwen3.8-27B-4bit",
				"outcome": "oom",
				"status": "verified",
				"supported_claim_scope": [
					"timing",
					"memory",
					"model_fit"
				],
				"system": "Apple M5 Pro, 24 GiB unified memory"
			},
			{
				"evidence_id": "qwen38-27b-m5-pro-fit-frontier-20260901",
				"kind": "fit_frontier",
				"model": "mlx-community/Qwen3.8-27B-4bit",
				"outcome": "oom",
				"status": "verified",
				"supported_claim_scope": [
					"timing",
					"memory",
					"model_fit"
				],
				"system": "Apple M5 Pro, 24 GiB unified memory"
			},
			{
				"evidence_id": "qwen38-27b-m5-pro-lab-oom-20260831",
				"kind": "model_lab",
				"model": "mlx-community/Qwen3.8-27B-4bit",
				"outcome": "oom",
				"status": "verified",
				"supported_claim_scope": [
					"timing",
					"memory",
					"model_fit"
				],
				"system": "Apple M5 Pro, 24 GiB unified memory"
			}
		],
		"edges": [
			{
				"relation": "same_model_as",
				"source": "cloudrift-glm53flash-preflight-20260902",
				"target": "modal-glm53flash-preflight-20260902"
			},
			{
				"relation": "compares",
				"source": "openrouter-glm-2k-comparison-20260902",
				"target": "qwen3-8b-m5-pro-control-20260902"
			},
			{
				"relation": "uses_workload_contract",
				"source": "qwen3-8b-cloudrift-vllm-compile-20260903",
				"target": "qwen3-8b-m5-pro-control-20260902"
			},
			{
				"relation": "same_model_as",
				"source": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
				"target": "qwen38-27b-m5-pro-fit-frontier-20260901"
			},
			{
				"relation": "same_model_as",
				"source": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
				"target": "qwen38-27b-m5-pro-lab-oom-20260831"
			},
			{
				"relation": "same_model_as",
				"source": "qwen38-27b-m5-pro-fit-frontier-20260901",
				"target": "qwen38-27b-m5-pro-lab-oom-20260831"
			}
		]
	},
	"claimMatrix": {
		"dimensions": [
			"timing",
			"quality",
			"cost",
			"memory",
			"process_attribution",
			"model_fit",
			"deployment_readiness"
		],
		"rows": [
			{
				"claims": {
					"cost": {
						"provenance": "planned caps and inferred zero spend; no provider usage",
						"state": "supported"
					},
					"deployment_readiness": {
						"provenance": "preflight refusal and exact stop gates are recorded",
						"state": "supported"
					},
					"memory": {
						"provenance": "aggregate listed V100 memory and exact model inventory comparison",
						"state": "supported"
					},
					"model_fit": {
						"provenance": "available V100 memory was below the exact model inventory",
						"state": "supported"
					},
					"process_attribution": {
						"provenance": "no process ran",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "no model output exists",
						"state": "unsupported"
					},
					"timing": {
						"provenance": "no deployment or request ran",
						"state": "unsupported"
					}
				},
				"evidence_id": "cloudrift-glm53flash-preflight-20260902",
				"outcome": "refused"
			},
			{
				"claims": {
					"cost": {
						"provenance": "local trace capture has no spend claim",
						"state": "not_applicable"
					},
					"deployment_readiness": {
						"provenance": "not a deployment run",
						"state": "not_applicable"
					},
					"memory": {
						"provenance": "GPU memory footprint was not measured",
						"state": "unsupported"
					},
					"model_fit": {
						"provenance": "no model was loaded",
						"state": "not_applicable"
					},
					"process_attribution": {
						"provenance": "target-process interval counts and trace-wide counts",
						"state": "supported"
					},
					"quality": {
						"provenance": "no model output was evaluated",
						"state": "not_applicable"
					},
					"timing": {
						"provenance": "interval counts are not utilization or time",
						"state": "unsupported"
					}
				},
				"evidence_id": "metal-attribution-m5-pro-20260831",
				"outcome": "completed"
			},
			{
				"claims": {
					"cost": {
						"provenance": "modeled cost and inferred zero spend are distinct scopes",
						"state": "supported"
					},
					"deployment_readiness": {
						"provenance": "preflight refusal and exact stop gates are recorded",
						"state": "supported"
					},
					"memory": {
						"provenance": "no runtime memory was measured",
						"state": "unsupported"
					},
					"model_fit": {
						"provenance": "planned hardware fit was not proven",
						"state": "unsupported"
					},
					"process_attribution": {
						"provenance": "no process ran",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "no model output exists",
						"state": "unsupported"
					},
					"timing": {
						"provenance": "no deployment or request ran",
						"state": "unsupported"
					}
				},
				"evidence_id": "modal-glm53flash-preflight-20260902",
				"outcome": "refused"
			},
			{
				"claims": {
					"cost": {
						"provenance": "provider usage, plan, and account delta remain separate scopes",
						"state": "supported"
					},
					"deployment_readiness": {
						"provenance": "comparison does not establish production readiness",
						"state": "unsupported"
					},
					"memory": {
						"provenance": "provider memory was not exposed",
						"state": "unsupported"
					},
					"model_fit": {
						"provenance": "hosted completion does not prove hardware fit",
						"state": "unsupported"
					},
					"process_attribution": {
						"provenance": "hosted provider internals",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "pinned evaluators for eight completed requests",
						"state": "supported"
					},
					"timing": {
						"provenance": "client-observed hosted request timing",
						"state": "supported"
					}
				},
				"evidence_id": "openrouter-glm-2k-comparison-20260902",
				"outcome": "comparison"
			},
			{
				"claims": {
					"cost": {
						"provenance": "boot-to-console list-rate inference; provider spend is unavailable",
						"state": "supported"
					},
					"deployment_readiness": {
						"provenance": "bounded benchmark is not a production deployment assessment",
						"state": "unsupported"
					},
					"memory": {
						"provenance": "sampled peak device memory for both cells",
						"state": "supported"
					},
					"model_fit": {
						"provenance": "both cells completed; per-cell source binding is limited",
						"state": "supported"
					},
					"process_attribution": {
						"provenance": "ordered non-overlapping cell processes",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "22 of 24 deterministic workload evaluator results passed",
						"state": "supported"
					},
					"timing": {
						"provenance": "measured initialization and 24 bounded request records",
						"state": "supported"
					}
				},
				"evidence_id": "qwen3-8b-cloudrift-vllm-compile-20260903",
				"outcome": "comparison"
			},
			{
				"claims": {
					"cost": {
						"provenance": "local execution has no spend claim",
						"state": "not_applicable"
					},
					"deployment_readiness": {
						"provenance": "exploratory benchmark is not a deployment assessment",
						"state": "unsupported"
					},
					"memory": {
						"provenance": "MLX allocator counters; not RSS or swap",
						"state": "supported"
					},
					"model_fit": {
						"provenance": "this self-converted 8B checkpoint completed through requested 16K",
						"state": "supported"
					},
					"process_attribution": {
						"provenance": "single isolated row process",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "pinned evaluator pass rate and score",
						"state": "supported"
					},
					"timing": {
						"provenance": "host wall-clock prefill/decode/total",
						"state": "supported"
					}
				},
				"evidence_id": "qwen3-8b-m5-pro-control-20260902",
				"outcome": "completed"
			},
			{
				"claims": {
					"cost": {
						"provenance": "local execution has no spend claim",
						"state": "not_applicable"
					},
					"deployment_readiness": {
						"provenance": "OOM autopsy does not assess deployment readiness",
						"state": "unsupported"
					},
					"memory": {
						"provenance": "separate MLX allocator, RSS, swap, and headroom scopes",
						"state": "supported"
					},
					"model_fit": {
						"provenance": "clean-boot OOM for the exact checkpoint and t256 workload",
						"state": "supported"
					},
					"process_attribution": {
						"provenance": "single autopsy child process",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "no first token or evaluator result exists",
						"state": "unsupported"
					},
					"timing": {
						"provenance": "stage-boundary host wall-clock observations",
						"state": "supported"
					}
				},
				"evidence_id": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
				"outcome": "oom"
			},
			{
				"claims": {
					"cost": {
						"provenance": "local execution has no spend claim",
						"state": "not_applicable"
					},
					"deployment_readiness": {
						"provenance": "exploratory frontier is not a deployment assessment",
						"state": "unsupported"
					},
					"memory": {
						"provenance": "approximate system headroom and swap; not GPU memory",
						"state": "supported"
					},
					"model_fit": {
						"provenance": "bounded t256 OOM in the recorded host state",
						"state": "supported"
					},
					"process_attribution": {
						"provenance": "single local model run",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "no quality result was produced",
						"state": "unsupported"
					},
					"timing": {
						"provenance": "host wall-clock total before OOM",
						"state": "supported"
					}
				},
				"evidence_id": "qwen38-27b-m5-pro-fit-frontier-20260901",
				"outcome": "oom"
			},
			{
				"claims": {
					"cost": {
						"provenance": "local execution has no spend claim",
						"state": "not_applicable"
					},
					"deployment_readiness": {
						"provenance": "a failed exploratory warmup is not a deployment assessment",
						"state": "unsupported"
					},
					"memory": {
						"provenance": "system memory pressure and swap; MLX peak is unavailable",
						"state": "supported"
					},
					"model_fit": {
						"provenance": "OOM during the exact recorded warmup and host state",
						"state": "supported"
					},
					"process_attribution": {
						"provenance": "single local model run",
						"state": "not_applicable"
					},
					"quality": {
						"provenance": "no evaluator result was produced",
						"state": "unsupported"
					},
					"timing": {
						"provenance": "host wall-clock total before OOM",
						"state": "supported"
					}
				},
				"evidence_id": "qwen38-27b-m5-pro-lab-oom-20260831",
				"outcome": "oom"
			}
		],
		"stateCounts": {
			"supported": 28,
			"unsupported": 19,
			"not_applicable": 16
		}
	},
	"aliases": {
		"metalPidAttribution": "metal-attribution-m5-pro-20260831",
		"oom27bAutopsy": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
		"qwen3_8bPositiveControl": "qwen3-8b-m5-pro-control-20260902",
		"hostedGlmComparison": "openrouter-glm-2k-comparison-20260902",
		"modalBudgetRefusal": "modal-glm53flash-preflight-20260902",
		"cloudriftHardwareRefusal": "cloudrift-glm53flash-preflight-20260902",
		"cloudriftCompileCrossover": "qwen3-8b-cloudrift-vllm-compile-20260903"
	},
	"lineageProjection": {
		"kind": "explanatory_projection",
		"evidenceId": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
		"canonicalGraphEdges": false,
		"note": "These stages are an explanatory projection of fields in one canonical entry. They are not extra catalog nodes or edges.",
		"stages": [
			{
				"role": "model pin",
				"label": "mlx-community/Qwen3.8-27B-4bit",
				"detail": "3e6447f082e89cc7f0bc6e5441afd38dfce760ff",
				"canonicalField": "catalog entry: model"
			},
			{
				"role": "collection",
				"label": "qwen38-27b-m5-pro-clean-boot-autopsy-20260901",
				"detail": "2519bc8da309656d2e2ce2a7063f19b0dfb4c9ed · examples/optimizer/m5-pro-qwen3.8-27b-oom-autopsy/publication",
				"canonicalField": "catalog entry: public_path + source_commit"
			},
			{
				"role": "verification",
				"label": "oom-autopsy.verify_bundle",
				"detail": "adapter oom_autopsy_v1 · version 1",
				"canonicalField": "registry source: adapter"
			},
			{
				"role": "report",
				"label": "oom-autopsy-summary.json",
				"detail": "sha256:1096de514bcdd74201d367d098352da831efa9f21d85aa2993af0e2787c56030",
				"canonicalField": "registry artifact + entry artifact_set_hash"
			},
			{
				"role": "article",
				"label": "A benchmark result without lineage is just a screenshot",
				"detail": "Cites the canonical entry. It is not evidence and does not verify the source.",
				"canonicalField": "portfolio citation"
			}
		]
	}
}