Verify who signed what under an explicit trust policy
Bind signatures to artifact identities, approved signers, trustworthy time evidence, and versioned roots without treating a valid signature as harmless code.
On this page
The question for today
What you’ll understand
- Separate cryptographic validity, identity verification, and authorization.
- Explain certificate lifetime and trusted signing-time evidence.
- Plan rotation and compromise response across verifier trust state.
A release has a mathematically valid signature made by an identity your deployment does not authorize. Verification has answered one question while approval still fails. A supply-chain verifier must bind the exact artifact or attestation payload to an accepted key or certified identity and then apply an explicit trust policy. The signature does not inspect model-serving code, establish benign intent, or repair a compromised builder. Its value comes from a verifiable attribution and content binding within a maintained trust system.
Verify the payload and authorize the identity
Use the signature scheme's defined verification procedure on the exact signed representation; signatures are not universally explained as decrypting a hash. For an image signature, check that the verified claims identify the intended artifact digest rather than a different release. A valid signature on unrelated bytes is irrelevant. When using a certificate-based workflow, validate the chain to approved roots and the expected identity and issuer constraints. Sigstore's documented Cosign identity-based examples specify both identity and OIDC issuer. Accepting any identity certified by a broad ecosystem is different from approving your release workflow.
Keep verification and authorization results explicit. Cryptographic success means the signature is consistent with the verified payload and key under the algorithm's assumptions. Identity validation associates that key with an asserted identity under the certificate system. Authorization decides whether that identity may make this claim about this artifact for this deployment. A stolen credential or compromised approved workflow can produce signatures that pass the first checks. Review what the identity is allowed to assert and retain source, builder, and policy context rather than converting every verified signature into a universal safe label.
Use trustworthy time evidence and maintain roots
Short-lived signing certificates create an important distinction between the verifier's current time and the time relevant to the signature. An expired certificate does not automatically prove every historical signature invalid, nor does a timestamp typed into the signed payload establish when signing occurred. A verification bundle can carry certificates, signatures, and transparency or timestamp material that the verifier checks under the applicable format and policy. The trusted time and log requirements depend on that workflow. Do not bypass chain or time checks merely because an artifact was previously accepted on another machine.
Root and key rotation must update verifier trust through an authenticated mechanism, with versions, validity, and rollback protections appropriate to the system. Sigstore's trust model describes maintaining trust material through TUF; fetching arbitrary replacement roots from the artifact's registry is not equivalent. Plan overlap and retirement deliberately so old releases are handled according to policy while compromised authorities can be restricted. A key's removal from future signing and an artifact's revocation from deployment are different actions. Record which trust version supported a decision so an incident can identify stale acceptance caches and offline verifiers.
Threshold counts need distinct approved evidence
Some organizations require several distinct approvers for a release. Count only signatures that verify, bind the intended subject, and satisfy the approval policy; repeated signatures from one identity do not create independent authorities. Threshold design must account for shared credentials and administrative control rather than assume different key names imply independence. Document exceptional recovery and trust-update procedures before an outage. The example here uses a fictional three-of-five approval rule; it is not a statement that Sigstore or every signature format implements that particular threshold rule.
The reliability lab can display the count arithmetic alone. For each separate run, Reset, set both replica count n=5 and unavailable count k, predict the remaining count and majority indicator, then inspect. With k=2, three remain and meet the instrument's majority count; with k=3, two do not. This is a deterministic availability model, not a signature verifier, threshold-cryptography implementation, or estimate of independent signer compromise. In a real approval workflow, being online is insufficient: each counted signature must also pass content, identity, time, and authorization checks.
Work it through
Filter signatures before counting approvals
A fictional release policy requires two distinct approved signer identities. Three distinct identities provide signatures over the same intended digest. All three signatures are cryptographically valid, two identities are approved, and only one of those two has acceptable time and trust evidence under this exercise’s policy.
Begin with three cryptographically valid signatures whose artifact claims match. That does not yet provide three authorized approvals.
Identity policy leaves two candidates. Applying the specified time and trust checks leaves one acceptable distinct approval.
One is less than the required two, so the release does not meet this fictional policy. Repair evidence or obtain the required valid approval through the authorized workflow; do not count the rejected signatures as availability substitutes.
There is one countable approval, so the two-identity rule is unmet. No multiplication of independent probabilities is involved in this deterministic filtering.
The living notebook
Explore the model
Interactive lab · reliability
Count remaining participants without claiming signature validity
For each view: Reset, set both n=5 and unavailable k, predict, then inspect. Compare k=2 with k=3: three versus two participants remain, crossing the instrument’s majority of three. This count model is not cryptographic verification or a universal Sigstore threshold policy.
- Available replicas
- 4 of 5
- Majority needed
- 3
- Majority available?
- Yes
A replica count and a majority threshold, without a consensus protocol. Playback highlights fixed replica states. Having a majority alone does not establish linearizability, durability or failure independence. Parameter changes pause playback and recompute the model at the current elapsed time. Reset restores both parameters and returns time to zero. Playback starts paused. Adjust either parameter, then choose Play or Step.
Look forA real approval count includes only distinct signatures satisfying all subject, identity, time, and trust requirements; online participants alone are insufficient.
Pause & reason
Check your understanding
Your turn
Try it yourself
A fictional three-distinct-approver policy receives five valid signatures: two from identity A and one each from B, C, D. A, B, C are approved, but C fails required time evidence and D is unapproved. How many approvals count?
Take a moment to make a prediction or work it out on paper. Open a hint when you need one.
Give me a hint
Deduplicate identities, then apply every policy condition.
Compare with the solution
Only A and B count, giving two approvals. The duplicate A signature adds none, C fails the required evidence rule, and D fails authorization. The required three distinct acceptable approvals are therefore absent.
Keep the idea
Take it with you
- Verify the exact signed subject and explicitly authorize the verified identity.
- Trusted time and maintained roots are part of historical verification policy.
- Rotation, compromise response, and approval counting require more than signature availability.
Read further
Original sources and deeper explanations to follow at your own pace.
Optional progress, saved in this browser when available. Revisit any lesson at any time.
Lesson not marked complete.