Attach a deployed model and ask for its static evidence.

Connect during review

1. Enable developer mode in ChatGPT.
2. Add https://deepbom.org/mcp as a Streamable HTTP plugin endpoint.
3. Attach one supported deployment artifact and ask: Use DEEPBOM to inspect this file. Keep defects, cautions, and evidence gaps separate.

The endpoint is ready for private developer-mode testing. Public discovery in ChatGPT begins only after OpenAI plugin review and listing.

Review a traceable model view

After the bounded audit completes, the component renders six deterministic Model IR levels: identity and boundary, architecture overview, block detail, exhaustive graph or storage, static runtime projection, and observed runtime overlay. Every displayed subject remains linked to the same hash-bound IR used by the audit.

The selected page can be downloaded as canonical monochrome SVG or a 300-DPI black-and-white PNG. The Word-ready bundle contains all view pages, caption sidecars, page hashes, conservation results, and insertion metadata. If local downloading is blocked, Save via ChatGPT uploads only the generated export and opens its temporary download URL; Send link to chat requests a clickable file link. Send PNG to chat shares the selected picture. Model bytes are not sent to the DEEPBOM service.

Continue with focused evidence questions

Use deepbom_workflow_guide to choose among conversational analysis, web interaction and local CLI workflows. Optional weight queries cover distributions, channels, similarity, singular spectra, sparsity and quantization, with per-feature coverage and SVG/PNG plots. Continue on deepbom.org for richer controls; select your model again and compare its SHA-256. The link transfers only a workflow ID.

Connections exposing deepbom_query_file can search and page through operations, inspect a selected layer, retrieve findings, compare available static delegate profiles, and inspect serialized activation fusion. Follow-ups verify the same attachment SHA-256 and preserve the requested inspection depth. Click a query row for its details or save the current page as SVG or PNG.

Available profiles come from the rules bundled for that artifact. Static eligibility is distinct from observed runtime placement; investigation priorities do not establish speedup or accuracy. New tools must be available in the host connection before this flow can be used.

Ask in plain language

Check how this attached ONNX model is quantized without executing it.
Keep artifact defects, cautions, and evidence gaps separate.

What tensor encodings are present in this attached GGUF file?
Include its SHA-256 and the static evidence boundary.

Can this attached file establish runtime latency or accuracy?

Once the plugin is connected, the user does not need to name DEEPBOM. Tool selection is host-controlled and must be tested rather than assumed. DEEPBOM should analyze a supported attachment for the first two prompts. For the third, it should explain that those outcomes require separate runtime or task evidence rather than inventing a static conclusion.

Choose the right execution path

Evidence boundary

DEEPBOM reports serialized-artifact observations and deterministic static derivations. It does not establish measured latency, energy, task accuracy, clinical validity, actual runtime placement, or regulatory compliance. artifact_defect, caution, and evidence_gap remain separate in the returned result.

Supported single-file inputs are TFLite, ONNX, GGUF, SafeTensors, Core ML .mlmodel, and ExecuTorch .pte/.ptd. Package and sidecar closure remain local-tool workflows.

Policies and support

• Privacy and data handling
• Terms of use
• Support and vulnerability reporting
• Machine-readable product boundary