Command reference

DEEPBOM 1.103.0

Usage:
  deepbom audit <artifact-or-package> [options]
  deepbom gguf <artifact.gguf> [options]
  deepbom verify <artifact> --contract <json> [options]
  deepbom diff <baseline-artifact-or-package> <candidate-artifact-or-package> [options]
  deepbom explore <artifact.tflite> [options]
  deepbom graph <artifact> [options]
  deepbom visualize <artifact> [options]
  deepbom accelerator collect nvidia [options]
  deepbom capabilities [--json|--compact]

Supported inputs:
  Stable or existing: .tflite, .onnx, .gguf, .safetensors, .mlmodel, .pte, .ptd
  Preview static graph: GraphDef or saved_model.pb
  Preview safe envelope: .h5, .hdf5, .keras, .pt2, .pt, .pth, .ckpt
  Packages: .mlpackage directories and sharded SafeTensors repository directories

Options:
  --target <id>          TFLite target profile (default: android_mid_a55)
  --target-profile <json>
                          Bind a strict custom TFLite target profile (mutually exclusive with --target)
  --contract <json>      Production external-interface contract for verify
  --request <json>       Bound redesign request for explore
  --external-data-dir <directory>
                          Resolve ONNX external_data or ExecuTorch PTD sidecars from this directory
  --context <tokens>     Declared text-token scenario for a statically derived LLM KV contract
  --images <count>       Declared image count; requires --tokens-per-image
  --tokens-per-image <count>
                          Declared projector output tokens per image; never inferred
  --batch <count>        LLM scenario batch size (default: 1)
  --state-bits <bits>    LLM state width: 8, 16, or 32 (default: 16)
  --memory-mib <MiB>     Compare the conditional lower bound with a declared capacity
  --tensorrt-profile <json>
                          Bind an ONNX TensorRT native/ORT EP build profile
  --tensorrt-parser-evidence <json>
                          Import identity-bound TensorRT parser/build evidence
  --tensorrt-llm-config <json>
                          Assess a TensorRT-LLM engine config with SafeTensors
  --tensorrt-llm-binding <json>
                          Bind that config to model-source/component digests
  --llm-memory-profile <json>
                          Evaluate serialized layer/state lower bounds against declared CPU and accelerator pools
  --output-format <kind> summary, json, json-compact, envelope, cyclonedx, or sarif
  --section <names>      Emit selected analysis sections; use --list-sections to discover names
  --pointer <pointer>    Emit one RFC 6901 JSON Pointer result with artifact identity
  --list-sections        List selectable analysis sections for this artifact
  --gate defects         Exit 2 only when an artifact_defect finding is present
  --timestamp <iso>      Fixed generation timestamp; SOURCE_DATE_EPOCH is also honored
  --fail-on <severity>   Compatibility severity gate: informational, low, medium, or high
  --policy-output <path> Write the deterministic finding-gate decision JSON
  --output, -o <path>    Atomically write the complete document; use - for stdout
  --no-clobber           Refuse to replace an existing output or policy file
  --error-format <kind>  text or json structured stderr (default: text)
  --json                 Compatibility alias for --output-format json
  --compact              Compatibility alias for --output-format json-compact
  --version              Print version
  --help                 Show this help

Safe-envelope boundary:
  Keras, HDF5, PT2, and PyTorch checkpoint previews do not construct framework objects or claim an executable graph.

Exit codes:
  0 pass; 1 invocation/input/analysis/output failure; 2 policy or verification block; 3 incomplete verification binding

Additional command:
  deepbom model-summary <artifact> [--level auto|operation|block|storage] [--format table|markdown|json|json-compact]

SavedModel package preview:
  TensorFlow SavedModel directories are accepted with bounded member hashing and first-MetaGraph projection.
  GraphDef, Keras config, and PT2 JSON expose serialized declarative relationships only; HDF5 and PyTorch checkpoints remain safe-envelope-only.

GGUF quick inspection:
  --tensors              Emit a concise tensor table; add --json or --compact for deepbom.tensor_table.v1

Format-neutral tensor evidence:
  --encoding-inventory   Emit serialized encoding counts, assignments, and reproducible hashes for GGUF, SafeTensors, ONNX, or TFLite

Windows-safe batch audit:
  deepbom batch <manifest.json> --batch-output-dir <directory> [--json|--compact]
  Manifest: deepbom.batch_manifest.v1 with safe relative paths and required SHA-256 identities

Built-in gate profiles:
  --policy engineering  Block artifact defects; keep cautions and evidence gaps visible
  --policy regulatory   Block artifact defects and unresolved evidence gaps; this is not a legal-compliance determination

Human-readable audit output:
  --output-format summary  Emit the bounded projection derived from deepbom.review_summary.v1
  --summary                Compatibility alias for --output-format summary

Installation and rule checks:
  deepbom self-test [--json|--compact]
  deepbom explain-rule <rule-id> [--json|--compact]
  deepbom explain-rule --list

Agent-native local use (no hosted analysis server):
  deepbom capabilities --format agent-json
  deepbom integrate codex                  Preview .agents/skills/deepbom
  deepbom integrate claude-code            Preview .claude/skills/deepbom
  deepbom integrate generic                Preview skills/deepbom
  deepbom integrate <target> --apply       Write only managed Skill files
  deepbom integrate status [target]        Inspect without changing files
  deepbom integrate remove <target>        Preview managed-file removal
  deepbom integrate remove <target> --apply  Remove only unchanged managed files

Assistant tool access (Model Context Protocol over stdio, local process only):
  deepbom mcp
  Tools: deepbom_capabilities, deepbom_audit, deepbom_diff, deepbom_explain_rule
  Audit default: bounded human summary; request envelope, section, or pointer for detail
  Local paths: launch directory, or roots declared by DEEPBOM_MCP_ALLOWED_ROOTS

NVIDIA accelerator binding:
  --accelerator-profile <json>
                          Bind an observed NVIDIA host profile without inferring selected-build or runtime assignment
  --accelerator-device <index>
                          Select one device when the bound profile contains multiple NVIDIA devices

N-way placement comparison:
  deepbom placement <artifact> [--profiles <id,id|all>] [--json|--compact]
  --profiles <ids|all>   Compare selected independent profiles (default: all available profiles)

Compiled accelerator evidence:
  --coreml-compute-plan <json>
                          Import an artifact- and compiled-model-bound MLComputePlan estimate; not executed placement
  --edgetpu-compiler-evidence <json>
                          Import an artifact/compiler/invocation/compiled-artifact-bound Edge TPU operation ledger
  --litert-qualcomm-evidence <json>
                          Import an artifact/source/compiler/QNN-plan-bound operation ledger

Conversion provenance:
  --conversion-receipt <json>
                          Bind a self-hashed source/converter/environment receipt to the observed output artifact. Source .pt/.pth/.h5 files are identified by digest only and are never deserialized.

NVIDIA accelerator observation:
  deepbom accelerator collect nvidia [--device <index>] [--json|--compact]
  --device <index>        Collect one NVIDIA device index (default: all devices)
  --include-device-identifiers
                          Include raw GPU UUID and PCI bus ID; hashes are always emitted

Remote immutable artifact input:
  hf://owner/repo@<40-hex-commit>/path
  gs://bucket/object#generation=<generation>
  https://host/path#sha256=<64-hex>
  --expected-sha256 <hex> Add an independent content digest requirement
  --cache-dir <directory> Use a content-addressed cache directory
  --offline               Refuse network access and require a verified cache receipt
  --max-download-gib <n>  Bound one remote download (default: 50, maximum: 1024)

Bounded scan policy:
  --scan <mode>           auto, structure, integrity, or full
  GGUF/SafeTensors use range reads; auto selects structure above 10 GiB and streamed integrity above 2 GiB.
  Monolithic TFLite/ONNX/ExecuTorch files above 1 GiB fail before full-file allocation.

Repeat-review policy:
  --review-policy <json> Bind required analysis coverage, finding threshold, and identity-scoped expiring exceptions
  --policy-output <path> Write the deterministic policy decision JSON
  --policy, --gate, --review-policy, and --fail-on are mutually exclusive policy sources.

Output format selection:
  --output-format <kind>  Preferred unambiguous spelling for --format
  --format <kind>         Backward-compatible alias; retained for existing automation

Deterministic graph export:
  deepbom graph <artifact> --view structure --output-format svg -o graph.svg
  --view <kind>           structure, placement, quantization, or architecture
  --output-format <kind>  svg, png, html, mermaid, dot, or json for graph

Regulatory-document model views:
  deepbom visualize <artifact> --view all --orientation portrait -o model-views.zip
  --view <ids>            all or comma-separated identity-boundary, architecture-overview, block-detail, exhaustive, static-runtime, observed-runtime
  --orientation <kind>    portrait or landscape ISO A4
  --output-format <kind>  bundle (SVG, captions, manifest) or json manifest
  These monochrome, hash-bound engineering views do not assert regulatory approval or standard conformance.

Format-neutral Keras-style model summary:
  deepbom model-summary <artifact> --level auto --format table
  --level <kind>          auto, operation, block, or storage
  --format <kind>         table, markdown, json, or json-compact
  Rows are projections of deepbom.model_ir.v1; storage is not relabeled as trainable parameters and display order is not runtime order.

TensorRT optimized-engine option:
  --tensorrt-engine-inspector <json>
                          Import identity-bound TensorRT optimized-engine inspector evidence

ExecuTorch selected-build option:
  --executorch-build <json>
                          Bind backend/operator inventories and runtime binary digests to a PTE audit

Artifact reconciliation and baseline capture:
  deepbom verify <artifact> --bom <cyclonedx-1.7.json> [--component-ref <bom-ref>]
  deepbom contract capture <artifact> [-o baseline.interface-contract.json]
  deepbom diff <baseline> <candidate> --tensors
  --render markdown      Render audit, diff, or verify results as Markdown

Agent capability summary:
  deepbom capabilities --format agent-text

Additional exit code:
  4 independently supplied artifact SHA-256 mismatch

deepbom --version
deepbom capabilities --compact
deepbom capabilities --format agent-json