The AI trust layer between agents: identity, agreements, safeguards, and outcome intelligence for A2A.

Commit 867f032d

867f032d6ce0Merged

aryaamin <[email protected]> · 2 months ago · 39 files · +17 added · ~22 modified

Description

feat: add intelligent shield and tau3 benchmark

Details

Commit
867f032d6ce0
Parent
92498f6892e8

DATA_BOUNDARY.mdmodifiedView file

@@ -15,8 +15,12 @@
1515 Hosted records are partitioned by the authenticated Auth0 subject in Postgres. Dashboard and
1616 settings endpoints require a server-validated session token; browser authentication state is never
1717 treated as authorization by itself. Agent messages travel directly and do not enter OpenClasp
18-storage. Message text is never eligible for profiles or network intelligence. Only explicitly reported structured events,
19-hashes, evidence references, receipts, feedback, and outcomes can feed those systems.
18+storage. Shield is an explicit exception to the processing path, not the storage rule: a connected
19+agent or owner can send bounded current-turn context to the configured model provider for analysis,
20+but OpenClasp discards that text after generation and stores only its digest and Shield's structured
21+assessment. Message text is never eligible for profiles or network intelligence. Only explicitly
22+reported structured events, hashes, evidence references, receipts, feedback, assessments, and
23+outcomes can feed those systems.
2024
2125 Local contextual profiles are account-private. A learning decision is labelled `network_aggregate`
2226 only when both participating accounts opted in before conclusion processing; otherwise it is