Agentic AI orchestration of CAMARA network APIs for African mobile-money trust scoring. Combines six CAMARA network signals (SIM swap, device swap, KYC match, number verification, location, device status) into a composite trust score with a hash-chained audit trail.
TrustMesh defines 3 tools with complete input schemas and descriptions. Tool names follow verb_noun convention (check_, verify_, get_). Descriptions are substantive (100-200 chars), explaining WHAT and WHEN. However, parameter descriptions lack actionable constraints (ranges, enums, formats). Output schemas are partially documented in code but not formally exposed in tool registration. Error handling is minimal, no recovery guidance or categorization. Security considerations (phone validation, audit trails) are present but not declared in tool metadata.
Check the fraud risk for a mobile money transaction. Returns a composite trust score (0-100), an APPROVE/FLAG/BLOCK decision, the contributing API signals, and an audit certificate hash. Use this before authorizing any high-value mobile money transaction.
Retrieve the most recent TrustMesh decisions from the audit chain. Useful for fraud analysts reviewing flagged transactions or for AI agents needing context on recent activity. Returns up to N most recent audit certificates with their decisions. Returns audit metadata only — no personal information.
Verify the integrity of TrustMesh's tamper-evident decision audit chain. Returns whether the chain is intact, the total certificate count, and the latest certificate hash. Use this to prove no historical TrustMesh decision has been altered.
Parameter constraints not formalized. 'phone_number' requires E.164 format and dev-mode simulator pattern (+9999999XXXX), but these are enforced in code, not declared in schema. LLMs cannot see the constraint and may pass invalid values.
Output schemas not formally documented in tool definitions. TrustScoreToolResult, AuditVerifyToolResult, and RecentDecisionsToolResult are defined in code but not exposed as part of tool metadata. LLMs cannot plan downstream calls without seeing return structure.
No error recovery guidance. Tools lack descriptions of failure modes and recovery steps. E.g., check_trust_score may fail if phone validation fails or network is unavailable, but no guidance on what the LLM should do next.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
Numeric parameter 'limit' in get_recent_decisions lacks bounds. Description says 'default 10, max 50' but no min/max declared in schema. LLMs may pass 0, negative, or >50 values.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) not present. All three tools are read-only and idempotent, but this is not declared in metadata. Agents cannot infer safety properties.