CapitalGuard
- policy-filtered repository projection
- OCI-contained agent session
- authenticated action broker
- signed publisher benchmark
Open evidence index 0.1
Claims do not get a score. Evidence does.
One fixed challenge, visible operator provenance, retained evidence digests, and signed submissions for every product version.
CapitalGuard has one publisher-operated signed run. Promptfoo, SPLX, and Noma Security are listed from their official documentation but remain not_assessed until the same challenge is completed. No ranking is published.
22
Fixed challenge cases
22 / 22
Publisher run passed
0
External products scored
Evidence status
A row says exactly how it was earned.
| Product | Evidence source | Same challenge | Operator | Score |
|---|---|---|---|---|
| CapitalGuard | Publisher verified | Enforced Mode 0.2.0 | Publisher | Not ranked |
| Promptfoo | Documentation only | not_assessed | None | Not ranked |
| SPLX | Documentation only | not_assessed | None | Not ranked |
| Noma Security | Documentation only | not_assessed | None | Not ranked |
Documentation map
What each publisher documents is context, not a test result.
Promptfoo
- agent and coding-agent red teaming
- trajectory and tool-execution evidence
- enterprise adaptive input/output/tool-call guardrails
SPLX
- AI asset inventory
- automated red teaming
- input/output runtime protection
Noma Security
- AI security posture management
- AI red teaming
- runtime monitoring of prompts, responses, tools, MCP, and agent handoffs
Checked July 16, 2026. Vendor descriptions can change; only a signed challenge submission changes assessment status.
A signature proves integrity, not independence.
The index records whether a run was operated by the publisher, vendor, or an independent lab. Public-key continuity detects result tampering; operator independence still requires a declared third party.
Exact product version
Fixed fixture and methodology digests
Evidence digest for every assessed case
Ed25519 submission signature
Visible operator class
No score for not_assessed
Submission path
Run. Retain. Sign. Verify.
1Reconstruct the inert fixture.
2Run every declared case without changing expected outcomes.
3Retain machine evidence and calculate each digest.
4Sign the normalized submission with an Ed25519 key.
5Publish the operator class and verification result.
Machine-readable release
Challenge packEvidence indexCapitalGuard submissionVendor evidenceSubmission templateSubmission schemaVerification keyRelease manifest
Publisher key: caabde6ca835a036859c452ddde3c5d9ed6351cf8291d5be8b0ff77d59adc522
Apply the boundary