SBOM Conformance Report

"signature": {
  "algorithm": "ES256",
  "value": "MEUCIQD…"
}

Reference: cyclonedx.org

Write down how data moves through the system: where an input comes from, what the model receives, and where its output goes. Confirm the description matches how the system is actually deployed.

Reference: www.bsi.bund.de

State what the system does with the data it receives — whether it is stored, passed to another component, or used for further training. Check this against the privacy notice you publish.

Reference: www.bsi.bund.de

Describe the input and output the system accepts and produces — formats, size limits, and anything it refuses. A person confirms this because a model card lists modalities, not the system's actual interface.

Reference: www.bsi.bund.de

State the field the system is intended for and the uses that are out of scope. This is a decision, not a fact a tool can read, and regulators ask for it directly.

Reference: www.bsi.bund.de

Give the size and shape of the training data — record counts, class balance, and any known skew. These numbers explain the model's limits, so state them even when they are unflattering.

Reference: huggingface.co

Judge whether the training data holds personal information or third-party copyrighted work, and record what you filtered out and how. Where the answer is unclear, take it to your privacy and legal contacts before release.

Reference: www.bsi.bund.de

"externalReferences": [
  { "type": "bom", "url": "https://example.com/hbom.json" }
]

Reference: cyclonedx.org

List the controls that protect the model and its data — who can reach the weights, how the training environment is isolated, how artefacts are signed. These live in your operations, not in the SBOM.

Reference: www.bsi.bund.de

Name the security standard or internal policy this model was built under, and who signed off. Point at the assessment record rather than restating it here.

Reference: www.bsi.bund.de

Give the route for reporting a vulnerability in this model and who answers it. A published model needs a contact that outlives the person who trained it.

Reference: www.bsi.bund.de

Record what you measured on the security side — adversarial or jailbreak testing, refusal rates, red-team results — with the method and the date. If nothing was measured, say so plainly.

Reference: www.bsi.bund.de

Write down how data moves through the system: where an input comes from, what the model receives, and where its output goes. Confirm the description matches how the system is actually deployed.

Reference: www.bsi.bund.de

State what the system does with the data it receives — whether it is stored, passed to another component, or used for further training. Check this against the privacy notice you publish.

Reference: www.bsi.bund.de

Describe the input and output the system accepts and produces — formats, size limits, and anything it refuses. A person confirms this because a model card lists modalities, not the system's actual interface.

Reference: www.bsi.bund.de

State the field the system is intended for and the uses that are out of scope. This is a decision, not a fact a tool can read, and regulators ask for it directly.

Reference: www.bsi.bund.de

Give the size and shape of the training data — record counts, class balance, and any known skew. These numbers explain the model's limits, so state them even when they are unflattering.

Reference: huggingface.co

Judge whether the training data holds personal information or third-party copyrighted work, and record what you filtered out and how. Where the answer is unclear, take it to your privacy and legal contacts before release.

Reference: www.bsi.bund.de

"externalReferences": [
  { "type": "bom", "url": "https://example.com/hbom.json" }
]

Reference: cyclonedx.org

List the controls that protect the model and its data — who can reach the weights, how the training environment is isolated, how artefacts are signed. These live in your operations, not in the SBOM.

Reference: www.bsi.bund.de

Give the route for reporting a vulnerability in this model and who answers it. A published model needs a contact that outlives the person who trained it.

Reference: www.bsi.bund.de

Record what you measured on the security side — adversarial or jailbreak testing, refusal rates, red-team results — with the method and the date. If nothing was measured, say so plainly.

Reference: www.bsi.bund.de

Write down how data moves through the system: where an input comes from, what the model receives, and where its output goes. Confirm the description matches how the system is actually deployed.

Reference: www.bsi.bund.de

State what the system does with the data it receives — whether it is stored, passed to another component, or used for further training. Check this against the privacy notice you publish.

Reference: www.bsi.bund.de

State the field the system is intended for and the uses that are out of scope. This is a decision, not a fact a tool can read, and regulators ask for it directly.

Reference: www.bsi.bund.de

Judge whether the training data holds personal information or third-party copyrighted work, and record what you filtered out and how. Where the answer is unclear, take it to your privacy and legal contacts before release.

Reference: www.bsi.bund.de

List the controls that protect the model and its data — who can reach the weights, how the training environment is isolated, how artefacts are signed. These live in your operations, not in the SBOM.

Reference: www.bsi.bund.de

Give the route for reporting a vulnerability in this model and who answers it. A published model needs a contact that outlives the person who trained it.

Reference: www.bsi.bund.de

Record what you measured on the security side — adversarial or jailbreak testing, refusal rates, red-team results — with the method and the date. If nothing was measured, say so plainly.

Reference: www.bsi.bund.de