Skip to content
Open
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 48 additions & 0 deletions IMPLEMENTATION_PATTERNS.md
Original file line number Diff line number Diff line change
Expand Up @@ -548,6 +548,53 @@ The five core elements do not operate in isolation. Effective ATF implementation
| Observability | All Elements | Provides visibility, triggers, and audit trail |
| Incident Response | Behavior | Incidents inform demotion decisions |

### Pattern: Three-Plane Composition for Verifiable Agent Actions

High-assurance agent systems often need more than one receipt or log entry.
A complete forensic record separates three questions that are related but not
identical:

| Plane | Question answered | Typical evidence | Primary ATF links |
|-------|-------------------|------------------|--------------|
| Delegation plane | Whose authority allowed this agent to act? | Signed delegation, scoped capability, authority-chain root | Identity, Segmentation, Incident Response |
| Decision plane | What did policy allow or deny at evaluation time? | Signed policy decision receipt, policy digest, evaluated constraints | Behavior, Segmentation, audit trail |
| Execution plane | What actually went out to the downstream system? | Signed execution proof, request/response commitment, trace/span reference | Behavior, Data Governance, Incident Response |

These planes should compose by content-hash reference rather than by re-signing
or mutating another plane's artifact. Each layer keeps its own trust boundary,
while auditors can walk the references to reconstruct the action.

```mermaid
flowchart LR
A[Delegation plane\nwhose authority?] -->|authority root hash| B[Decision plane\nwhat policy allowed?]
B -->|decision receipt hash| C[Execution plane\nwhat actually happened?]
C -->|execution proof hash| D[Audit / incident review]
A -. independent verifier .-> D
B -. independent verifier .-> D
C -. independent verifier .-> D
```

Implementation guidance:

- The delegation plane should prove scoped authority and revocation status for
the actor at the time of the decision.
- The decision plane should bind the evaluated policy, input or argument hash,
and allow/deny/escalate outcome.
- The execution plane should bind the actual downstream request/response or a
privacy-preserving commitment to them.
- Cross-plane links should be hashes of canonical artifacts, not copied mutable
fields.
- Verifiers should report which plane failed: delegation invalid, policy
mismatch, execution mismatch, or broken reference. These imply different
incident workflows.

Illustrative implementations discussed by the community include APS-style
delegation chains for the delegation plane, Veritas Acta / ScopeBlind decision
receipts for the decision plane, and proxy or gateway execution proofs for the
execution plane. The pattern does not require adopting all three at once; the
hash-reference boundary lets organizations adopt one plane while preserving a
path to later composition.

### Common Integration Failures

| Failure | Symptom | Resolution |
Expand All @@ -556,6 +603,7 @@ The five core elements do not operate in isolation. Effective ATF implementation
| Behavioral monitoring siloed | Anomalies detected but not acted upon | Connect monitoring to incident response |
| Data governance separate from segmentation | Data boundaries not enforced | Unify data and resource policies |
| Observability incomplete | Gaps in audit trail | Ensure all elements log to central platform |
| Planes collapsed into one artifact | A valid signature hides missing authority or mismatched execution | Separate delegation, decision, and execution evidence; link by canonical hash |

---

Expand Down