{
  "markdown": "# DEEPBOM\n\nDEEPBOM is a local static analyzer for deployed AI model artifacts. It audits\nserialized graph, tensor, quantization, memory, compatibility, and ML-BOM\nevidence without uploading model bytes.\n\nThe public source distribution covers TFLite, ONNX, GGUF, SafeTensors, Core ML,\nand bounded ExecuTorch artifacts. Findings distinguish observed and derived\nartifact facts from predicted compatibility, imported runtime evidence, and\nvalues that cannot be assessed statically.\n\n## Quick start\n\nRun the published CLI without cloning the repository (Node.js 20 or newer):\n\n```bash\nnpx deepbom audit \"https://raw.githubusercontent.com/JunHwan-Kwon/deepbom/main/web/samples/gpu_partition_probe.onnx#sha256=82a2feef00eb6ab03d82f2b30cd17f4d826e2d8307cb059eccd6a0f3120059b2\"\n```\n\nThe pinned expected values and independent verifier are in\n[`examples/expected-output`](examples/expected-output/README.md). Source builds\nrequire Rust and, for the Python channel, Python 3.9 or newer; maintainer setup\nis documented separately below.\n\n## Assistant integration\n\nRun the same local analyzer as an MCP server over stdio:\n\n```bash\nnpx deepbom mcp\n```\n\nIt exposes `deepbom_capabilities`, `deepbom_audit`, `deepbom_diff`, and\n`deepbom_explain_rule` without uploading artifact bytes or using a hosted\nanalysis endpoint. Audit calls default to a bounded human summary; detailed\nformats and large-model scan depth are explicit. Local paths are restricted to\nthe launch directory unless `DEEPBOM_MCP_ALLOWED_ROOTS` is configured.\nAgent-facing usage guidance is in [the DEEPBOM skill](skills/deepbom/SKILL.md).\n\nVerified release channels expose the same analysis implementation:\n\n```bash\nnpx deepbom audit model.onnx --format cyclonedx\nnpx deepbom audit model.onnx --format sarif --output deepbom.sarif --gate defects\ndeepbom capabilities --compact\ndeepbom self-test\ndeepbom audit model.onnx --list-sections --compact\ndeepbom audit model.onnx --section quantization,memory --compact\ndeepbom explain-rule onnx.conv.output-shape\npython -m pip install deepbom\ncargo install deepbom\ndeepbom audit model.gguf --compact\ndeepbom verify model.tflite --contract production-interface.json\ndeepbom diff baseline.tflite candidate.tflite\ndeepbom explore model.tflite\ndeepbom placement model.tflite --profiles xnnpack_cpu,tflite_coreml_delegate,litert_qualcomm_qnn\ndeepbom graph model.onnx --format json --output artifact-graph.json\ndeepbom audit model.onnx --conversion-receipt conversion-receipt.json --format cyclonedx\n```\n\nThe default is a terminal-sized evidence summary. `--json` and `--compact`\nexpose complete format evidence; `--format envelope` provides the canonical\ncross-format contract; CycloneDX 1.7 and OASIS SARIF 2.1.0 are standard\nprojections. `--policy-output` records a hash-bound gate result when `--fail-on`\nis selected. `--review-policy` adds identity-scoped, expiring exceptions and\nkeeps execution, coverage, and finding-policy states independent. See\n[`docs/CLI_AUTOMATION.md`](docs/CLI_AUTOMATION.md). The complete\noption inventory is generated from the executable in\n[`docs/CLI_REFERENCE.md`](docs/CLI_REFERENCE.md).\n\nThe graph JSON output includes the evidence-preserving\n`deepbom.artifact_ir.v2` ledger and a deterministic `deepbom.graph_ir.v1`\nvisualization compatibility projection. Serialized graph, storage topology,\narchitecture grouping, scoped quantization, static placement, and imported\nruntime evidence remain separate. Method `2.2.0` also preserves an optional,\noutput-bound conversion receipt without promoting declared converter execution\nto observed evidence. It materializes exactly decoded\nTFLite subgraphs, ONNX nested graphs/local functions, and ExecuTorch primary\nplans without flattening conditional scopes. Runtime-node fusion is reconciled\nonly from artifact-bound subject references or primary native op indices;\nnames are never guessed. `graph_ir.v1` remains primary-scope-only for legacy\nconsumers. The v2 JSON Schema is published at\n[`docs/schemas/deepbom-artifact-ir-v2.schema.json`](docs/schemas/deepbom-artifact-ir-v2.schema.json)\nand at `https://deepbom.org/schemas/deepbom-artifact-ir-v2.schema.json`.\nThe conversion receipt schema is published at\n[`docs/schemas/deepbom-conversion-receipt-v1.schema.json`](docs/schemas/deepbom-conversion-receipt-v1.schema.json).\n\nPublic product output uses CycloneDX 1.7. Experimental standards work is not\nincluded in the website, CLI commands, or release artifacts.\n\n`verify` compares the serialized external tensor ABI with a supplied,\nartifact-bound production declaration. `diff` uses the canonical deterministic\nmulti-target TFLite deployment-delta ledger, and `explore` exposes the existing\nWASM redesign Pareto search without claiming trained-model accuracy. A strict\ncustom TFLite target can be bound with `--target-profile profile.json`; the CLI\nrecords both the source-file SHA-256 and the resolved Rust profile SHA-256.\nAccelerator evidence is separate from that CPU cost profile. Source-pinned\nTFLite Core ML and LiteRT Qualcomm QNN profiles, Core ML MLComputePlan, Edge TPU\nand Qualcomm compiler reports, TensorRT parser/engine evidence, and NVIDIA host\nprofiles use a shared staged binding without promoting static or compiled\nevidence to observed execution. `placement` compares any available profiles\nover one conserved graph ledger without inventing backend priority.\n\nThe Cargo launcher downloads only the engine matching its exact package version\nand platform from the corresponding immutable GitHub Release. It validates the\nrelease matrix, byte lengths, and SHA-256 digests before caching or execution.\n\nBuild the npm, standalone, Python, and Cargo launcher channels:\n\n```bash\nnpm run build:channels\nnpm run check:public-package-boundary\nnpm run check:channels -- --no-build\n```\n\nRun the public correctness gates:\n\n```bash\nnpm run check:cli\nnpm run check:cli-docs\nnpm run check:cli-automation\nnpm run check:formats\nnpm run check:rust\n```\n\nThe browser workbench is available at [deepbom.org](https://deepbom.org/).\n\n## Evidence scope\n\n| Format | Public static evidence |\n| --- | --- |\n| TFLite | FlatBuffer graph and tensor contracts, quantization arithmetic, weight integrity, memory projections, accumulator proofs, redesign candidates, and source-bounded delegation predictions |\n| ONNX | Protobuf graph, initializer and external-data contracts, symbolic shape inference, operation cost, Q/DQ and affine quantization, and provider-compatible evidence envelopes |\n| GGUF | Container and tensor-directory integrity, quantization encoding inventory, architecture metadata, and bounded LLM memory scenarios |\n| SafeTensors | Tensor-directory and sharding integrity, configuration-bound architecture contracts, AWQ/GPTQ/HQQ/compressed-tensors metadata, and bounded LLM memory scenarios |\n| Core ML | NeuralNetwork and ML Program serialized graphs, tensor/weight encodings, deployment floor, and imported compute-plan evidence boundaries |\n| ExecuTorch | Bounded ET12/FT01 plans, source-bound portable calls and processed payload identities, plus optional selected-build/backend/operator/binary attestation; execution remains external |\n\nStatic compatibility does not establish observed execution-provider assignment,\ndevice latency, task accuracy, clinical validity, or release readiness. Runtime\nclaims require an identity-bound runtime capture.\n\nThe detailed format and accelerator boundary is maintained in\n[`docs/SUPPORT_MATRIX.md`](docs/SUPPORT_MATRIX.md). Bugs can be reported without\nsharing model bytes using\n[`docs/MODEL_FREE_BUG_REPORTING.md`](docs/MODEL_FREE_BUG_REPORTING.md).\n\n## Distribution boundary\n\nThis repository is generated from an exact reviewed allowlist. Private\nrulepack generators, hosted-service infrastructure, and unreleased research\nmodules are not included. The enforceable boundary and export verification\nmethod are documented in\n[`docs/PUBLIC_PRIVATE_BOUNDARY.md`](docs/PUBLIC_PRIVATE_BOUNDARY.md).\n\n## License and citation\n\nThe public source and release-channel packages are licensed under the Apache\nLicense 2.0. Third-party model artifacts retain their declared licenses.\n\nPlease cite:\n\n> Kwon, J. (2026). DEEPBOM: Browser-Native Static Analysis of On-Device Neural\n> Network Deployment Artifacts [Computer software]. Zenodo.\n> https://doi.org/10.5281/zenodo.21834508\n",
  "bytes": 8288,
  "sha": "1b67cbc917060f3524a3558e201d47fab54369cf28b4ae763e9d6cba14a93313",
  "repo_slug": "junhwan-kwon/deepbom",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/mcp_io_github_junhwan_kwon_deepbom_3f8a8f84/readme"
}