ON-DEVICE Pipeline
Deployment artifact evidence for on-device AI
Local static audit
Find artifact defects, review cautions, and see which claims still need runtime or process evidence.
Drop a supported file into this panel, or use the controls alongside. Analysis runs on this device.
Drop a supported artifact anywhere in this panel
Your artifact stays on this device. Optional telemetry remains off unless you enable it.
Clinical engineering context
Model evidence alone does not characterize the final deployed configuration. Runtime, hardware, interface contracts, and workflow integration can affect behavior under intended conditions of use.
Model-level performance evidence does not fully characterize the final deployed configuration. Runtime, target hardware, input and output contracts, quantization, fallback behavior, resource constraints, and workflow integration can affect how an AI-enabled medical device behaves under its intended conditions of use.
DEEPBOM provides technical deployment evidence that can support traceability, verification, and change-impact assessment. It does not replace clinical validation, usability engineering, cybersecurity assessment, or medical-device risk management.
Trust boundary
Each profile separates its deterministic regression baseline from runtime and task claims that the artifact cannot establish.
Choose a local artifact or package to inspect.
A light local read starts only after an artifact is selected.
Artifact bytes are read fully only after Run Static Audit is clicked.
The same evidence spine is used across formats; absent execution evidence is never inferred from a tensor container.
Current artifact coverageThe current row is the selected artifact's emitted coverage. The matrix below is a capability ceiling, not a claim that every artifact closes every rule.
Bind an exact runtime binary and build manifest without treating it as serialized graph evidence.
Select an artifact to resolve format-specific coverage.
Runtime execution is not observed by a static audit.
Select an artifact to resolve the next required evidence.
Release readiness is not assessed without deployment requirements and representative task-output evidence.
The exact byte-identified subject of this audit. Static derivations and target assumptions remain separate from this record.
This record establishes artifact identity and available static evidence. It does not by itself establish runtime assignment, task performance, clinical validity, or release readiness.
Replace artifact or packageSwitch between the main summary, delegate coverage, quantization contract, roofline pressure, and stage structure.
Review target-independent artifact identity first, then the separately labeled selected-target engineering dashboard.
Select a completed module result, open the Engineering Report, or export reusable engineering evidence.
Choose a local artifact or package; deployment context follows its format.
Resolve file or package
Compute serialized evidence
Apply deployment context
Bind runtime evidence
Evaluate requirements
Run controls live here. Results and downloads appear in the report pane below after execution.
The controlled WASM module is loaded only for authorized accounts. Browser delivery controls access, not cryptographic source secrecy.
Runs input/output drift probes, backend availability checks, and bounded output comparisons locally in the browser.
Run Static Audit, select registered hardware, then queue a shape/op/dtype/quantization-equivalent reconstruction. The original artifact and weight payload are not transferred.
Runs locally against the deployed TFLite runtime path and reports finite-difference deployment-function sensitivity evidence.
The login-free Engineering Report preserves the complete synthesized evidence ledger and action queue in a watermarked copy whose visible body is bound by SHA-256. Raw reusable derivatives remain separately controlled.
Every document binds the analyzed artifact SHA-256; target-dependent requirements bind the active analyzed target.
Local deploy-artifact weight and topology descriptors. Composite values are experimental and are not training-loss, accuracy, latency, robustness, or release-readiness evidence.
Local input perturbation sensitivity. Artifact bytes and generated tensors stay in browser memory.
Local backend availability and output-drift checks.
agent.py on a real device to register it.
Runs local finite-difference probes on the deployed model function and reports research-stage sensitivity proxies, not operational-stability evidence.
Calculate exact interface endpoint counts, reference-output drift, and repeated-run variance from a local capture. No capture values are uploaded.
Waiting for an audited artifact
Run a static audit before importing a capture.
Run a static audit to generate the report preview.
Run a static audit first. Package profiles exclude model bytes and raw tensor values. Hashes and signatures detect later changes but cannot prevent copying or editing.
When local history is enabled, compact artifact-audit snapshots are stored in this browser (identity, counts, contracts, quantization, delegation, and hotspots — never artifact bytes). Re-open, re-run, or compare two artifact versions by SHA-256.
Artifact observations remain distinct from declarations, inferred lineage, conditional runtime consequences, and unassessed fields.
| Parameter | Field | Expected | Declared |
|---|---|---|---|
| Not assessed | |||
Available without sign-in: exact file identity, format and graph totals, plus the complete external I/O contract ledger.
The ZIP adds a member-digest manifest and a detached ES256 signature made with a LOCAL_BROWSER_KEY. It proves package integrity, not DEEPBOM authorship. Official release evidence uses an OFFICIAL_RELEASE_KEY resolved through the fixed signing-key registry, including scope, validity, rotation, and revocation checks.
CycloneDX record, Artifact Evidence IR, interface ledger, formulation, runtime, provenance, member digest, and detached ES256 signature.
Format-applicable artifact evidence and explicitly bounded deployment signals.
Load a supported artifact or package to resolve its format-specific evidence depth.
Component signals are reported separately. No context-free 0–100 quality or readiness score is shown.
Observed or derived conditions that deserve attention; no remediation is implied.
Verification, contract binding, or changes grouped by decision axis; modeled impact is shown only where available.
Flame-style summaries for compute concentration, static bottlenecks, predicted delegate segments, quantization risk, and target fit.
Static resource-partition scenarios with imported OS affinity evidence kept separate.
| AI cores | Housekeeping | Allocation | INT8 / FP32 ceiling | Roofline floor | Steady estimate | Cold estimate | 1-core ratio |
|---|
Measures local browser runtime latency with warmup separated from steady-state runs.
| Backend | Compile | First (cold) | p50 | p90 | p95 | p99 | Mean | Trimmed p50 | Jitter | Trend | Status |
|---|
Artifact findings ranked by priority; each card separates evidence, possible effects, interpretation limits, and recommended action.
Run the static audit to generate findings.
Connected operators and tensor contracts with deployment evidence overlays.
Serialized storage is not relabeled as trainable parameters. Display order is deterministic, not runtime order.
| # | Op | Out | Kernel | MACs | Modeled steady µs | Bound | XNNPACK | Quant | Align | Pack | Fusion | Low-norm filters |
|---|
| T# | Name | Shape | DType | Role | Prod | Fan | Consumers | Quant |
|---|
Static placement evidence is available; no observed runtime assignment is bound.
No external node/edge evidence overlay imported.
| # | Op / path | Kernel / provider evidence | Tile / occupancy | Static assignment | Interface payload | Preparation / packing | Observed runtime | Evidence |
|---|
| Op | Count |
|---|
| # | Op | Output | MACs | Bytes |
|---|
| # | Op | Intensity | Logical traffic | Row working set | Modeled posture | Action |
|---|