Evidence you can check, from the chip up
We report on the evidence that lets someone other than an AI provider check its claims.
You, talking with an AI system
0 of 7 layers checked
Your prompts and data pass through all seven layers.
Point at or tab to each layer until you turn colorful.Tap each layer's arrow until you turn colorful.Click each layer's arrow until you turn colorful.Point at or tab to any layer to open it.Tap any layer's arrow to open it.Click any layer's arrow to open it.
Logos and names are examples, not endorsements.
Who checks: Regulators and inspectors can check commitments, mostly through audits and site visits.
T7, Policy
Checkable claim: Developers and governments keep the commitments they make about training, deployment and compute.
Who checks: Regulators and inspectors can check commitments, mostly through audits and site visits.
Who checks: Independent evaluators given access to the model can rerun the tests or check the test logs.
T6, Evaluations & audits
Checkable claim: A reported result came from the stated model, tests and conditions.
Who checks: Independent evaluators given access to the model can rerun the tests or check the test logs.
Who checks: Services an agent calls can check its signed requests; reviewers can later check its log.
T5, Agents & actions
Checkable claim: An agent took these actions, under this policy, on behalf of this person or organization.
Who checks: Services an agent calls can check its signed requests; reviewers can later check its log.
Who checks: Readers and platforms can check a response's signature or an image's Content Credential.
T4, Outputs & content
Checkable claim: This output came from this model and input, and this media has this origin.
Who checks: Readers and platforms can check a response's signature or an image's Content Credential.
Who checks: Developers and auditors can check a model's signature, or compare its hash with the published one.
T3, Models & data
Checkable claim: These are the weights that were trained, signed and deployed, with this lineage.
Who checks: Developers and auditors can check a model's signature, or compare its hash with the published one.
Who checks: Your device can compare attested measurements with published values before it sends a request.
T2, Software & runtime
Checkable claim: This exact code and configuration processed the request.
Who checks: Your device can compare attested measurements with published values before it sends a request.
With a TEE: Launch measurements identify the code and configuration that were loaded.
Who checks: Your device can check the attestation report's signature before it sends a request.
T1, Hardware & compute
Checkable claim: The computation ran on genuine hardware in a protected mode.
Who checks: Your device can check the attestation report's signature before it sends a request.
With a TEE: A report whose key chains to the chip vendor shows genuine hardware in a protected mode.
The chain of evidence today
Full stack and trust tiers How other approaches compare with TEEs
NewsAgent identity & accountability
Three U.S. agencies published ten voluntary practices on 16 October 2026 for incidents in which agents take significant uninstructed actions during training and evaluation. Several concern evidence that others could check later, including tamper-resistant logs, preserved model weights and outside audits of control systems.
· NIST
· Anthropic
· Infocomm Media Development Authority
· Official Journal of the European Union
· OPAQUE Systems
· Center for Security and Emerging Technology, Georgetown University
· Engineering at Meta
· Google DeepMind
· Cohere on X
Explainer · 8 min read
Explainer · 8 min read
Guide · 9 min read
Analysis · 9 min read
Analysis · 8 min read
