How grading works

Every check result is deterministic: given the same directory and response headers, the checker always produces the same verdict and grade. This page is the auditable scoring reference.

3-state verdict

The verdict is shown at the top of every report. It has exactly three possible values.

Verdict What triggers it
VALID The directory exists, all keys are structurally valid (kty, crv, and a 32-byte x), and no private key material (d) is present. A custom kid label is fine; the verdict does not depend on it.
INVALID A directory was found at the well-known path, but at least one critical check failed: private d present, wrong kty or crv, x not 32 bytes, or the Ed25519 signature does not verify. A custom kid that differs from the thumbprint does not make a directory invalid.
NOT FOUND No usable directory at the well-known path: non-2xx HTTP response, response body is not JSON, the key set is empty, or the host is unreachable (including DNS failure).

Letter grade

The letter grade (A through F) is only shown when a directory exists, that is when the verdict is VALID or INVALID. There is no letter grade for NOT FOUND.

The grade is derived from a weighted score that starts at 100 and subtracts points for each failed or warned check.

Scoring weights

Checks are grouped into three tiers. The deductions differ by tier and by severity (fail vs. warn).

Tier Fail deduction Warn deduction
JWKS directory -25 per fail -6 per warn
RFC 9421 signature -30 per fail -8 per warn
Agent Card (advisory) -6 per fail -2 per warn

Letter thresholds

Letter Minimum score
A90 or above
B75 to 89
C60 to 74
D40 to 59
Fbelow 40, or capped (see below)

Grade capping rules

Certain critical failures impose a hard ceiling on the grade regardless of the raw score.

Condition Effect
Private key material present in the directory Grade is capped to F
Ed25519 signature fails to verify Grade is capped to a maximum of D
Directory not served over HTTPS Grade is capped low (treated as a critical failure)

When a cap applies, the report notes the cap reason next to the grade so the cause is always visible.

Agent Card checks are advisory

Checks in the Agent Card tier nudge the score and can raise warnings, but they never change the verdict and never trigger a grade cap. A missing or incomplete Agent Card will lower the score slightly; it will not cause a VALID directory to become INVALID.

What is checked

JWKS directory checks

  • Directory is served over HTTPS
  • Directory URL returns HTTP 200
  • Response media type. The web-bot-auth type (application/http-message-signatures-directory+json) passes. Another valid JWKS type such as application/jwk-set+json or a generic application/json is accepted but raises a warning, because web-bot-auth asks for its own media type. A non-JSON type such as text/html, or a missing media type, is a failure.
  • Response body is valid JSON
  • Key set has the expected shape (keys array)
  • Key set is non-empty
  • Key set size is within reasonable bounds
  • Caching headers are present
  • Per-key: kty=OKP
  • Per-key: crv=Ed25519
  • Per-key: x is 32 bytes (base64url-encoded)
  • Per-key: the RFC 7638 thumbprint is computed and shown. This is the value a verifier matches an incoming signature's keyid against. The JWK kid member itself is a free-form label, so a custom kid is allowed and only the signature's keyid has to be a thumbprint. This check is informational and never lowers the grade.
  • Per-key: no private key material (d not present)
  • No duplicate kid values in the set
  • Validity window (if present) is plausible

RFC 9421 signature checks

  • Signature-Agent header present
  • Signature-Input parses correctly
  • Covered components include required fields
  • Required signature parameters present
  • tag=web-bot-auth
  • created and expires within allowed clock skew
  • keyid resolves to a key in the directory
  • Ed25519 signature verifies against the resolved key

Agent Card checks (advisory)

  • client_name present
  • client_uri present
  • contacts present
  • jwks_uri present
  • purpose present
  • RFC 9309 robots.txt compliance declaration
  • Rate expectation declared