DEEPBOM Artifact Evidence

Inspect AI model artifacts

Inspect attached AI model deployment files without executing model code. DEEPBOM analyzes supported TFLite, ONNX, GGUF, SafeTensors, Core ML, and ExecuTorch files in the browser and reports SHA-256, serialized structure, tensor encodings, quantization evidence, and static findings.

What you can ask

Inspect this model’s structure and quantization evidence. Which layers should I examine next?
Example path → ·

MCP connector

Server URL: https://deepbom.org/mcp
Login: None
Protocol: MCP 2025-11-25
Instructions: Use when a user attaches a supported serialized AI deployment artifact and asks for static inspection, a cross-format model summary, or a model-structure visualization. The attachment is analyzed in the ChatGPT browser sandbox; the DEEPBOM server receives only a bounded result, not model bytes. Do not present a result until returns it. Report model_summary rows as serialized operations or storage, not framework trainability or runtime order. The widget renders deterministic Model IR views locally and can export SVG, PNG, and a Word-ready bundle; do not reconstruct a graph from the bounded conversation summary or invent details absent from it. Keep artifact_defect, caution, and evidence_gap distinct. Static evidence does not establish measured latency, accuracy, clinical validity, regulatory compliance, standard conformance, or actual runtime placement. For follow-up questions missing from that summary, use with the same authorized file and previous SHA-256. Query operators, findings, improvement investigations, serialized activation fusion, or static backend profiles. Use exact subject_ref values and next_offset; wait for before answering. Source-pinned static candidates can be assessed even when actual runtime placement is unobserved. Do not label simple adjacency as backend fusion. Use to choose among broader CLI capabilities, visualizations and website interactions. Answer from existing evidence or request a supported query first; use the returned deepbom.org workflow link for richer controls and local argument templates for unsupported browser operations. The website requires selecting the model again and comparing SHA-256. Do not claim that a guide executes analysis. Use the local DEEPBOM MCP for confidential files or artifacts too large for ChatGPT.

Tools

Reads data7

Skills

1 in the ChatGPT listing

What you can ask

Inspect this model’s structure and quantization evidence. Which layers should I examine next?

Example path → ·

MCP connector

Server URL

https://deepbom.org/mcp

None

Protocol

MCP 2025-11-25

Instructions

Use when a user attaches a supported serialized AI deployment artifact and asks for static inspection, a cross-format model summary, or a model-structure visualization. The attachment is analyzed in the ChatGPT browser sandbox; the DEEPBOM server receives only a bounded result, not model bytes. Do not present a result until returns it. Report model_summary rows as serialized operations or storage, not framework trainability or runtime order. The widget renders deterministic Model IR views locally and can export SVG, PNG, and a Word-ready bundle; do not reconstruct a graph from the bounded conversation summary or invent details absent from it. Keep artifact_defect, caution, and evidence_gap distinct. Static evidence does not establish measured latency, accuracy, clinical validity, regulatory compliance, standard conformance, or actual runtime placement. For follow-up questions missing from that summary, use with the same authorized file and previous SHA-256. Query operators, findings, improvement investigations, serialized activation fusion, or static backend profiles. Use exact subject_ref values and next_offset; wait for before answering. Source-pinned static candidates can be assessed even when actual runtime placement is unobserved. Do not label simple adjacency as backend fusion. Use to choose among broader CLI capabilities, visualizations and website interactions. Answer from existing evidence or request a supported query first; use the returned deepbom.org workflow link for richer controls and local argument templates for unsupported browser operations. The website requires selecting the model again and comparing SHA-256. Do not claim that a guide executes analysis. Use the local DEEPBOM MCP for confidential files or artifacts too large for ChatGPT.