Every response from FeedOracle is cryptographically signed, independently verifiable, and blockchain-anchored. Agent behavior is controlled by a persistent state machine with pre-execution, post-execution, and runtime governance. Below are 17 live proofs — each backed by a real-time API call you can verify right now.
Every evidence response contains a JSON Web Signature verifiable against our JWKS public key.
JWKS endpoint live at /.well-known/jwks.json — ES256K public key for signature verification
JWS verification: ECDSA ES256K signatures on all evidence artifacts
8 schemas at /schemas/ — every response links to its schema via schema_ref.
EPM Schema v1.0 — Evidence Pack Manifest with SHA-256 hashing
Append-only log of every evidence pack — inspired by Google Certificate Transparency.
| Pack ID | Framework | Asset | Hash | Time |
|---|---|---|---|---|
| 105+ stablecoins and 61 RWA protocols monitored. Data refreshes automatically. | ||||
Every response includes freshness, confidence, per-source health, and tier-specific targets.
SLA: 99.5% uptime target — real-time monitoring at uptime.feedoracle.io
Every evidence pack is archived. Replay reconstructs byte-identical output — hash_match: true is the proof.
Replay proof: every evidence artifact independently reproducible from source data
AI agents register, receive ECDSA keys (90-day lifecycle), and accumulate reputation.
| # | Agent | Organization | Reputation | Tier | Requests |
|---|---|---|---|---|---|
| 44 MCP servers · 590+ evidence tools, including AgentGuard (20) and PredictionGuard (28) runtime security. OAuth 2.1, KYA identity, tamper-proof audit. | |||||
State-change-only events: peg deviations, regime changes, market stress. Every event is SHA256-hashed.
SSE + Streamable HTTP transport — live connections on feedoracle.io/mcp/sse
Verify any evidence pack without trusting the transport layer. 7 independent checks.
pip install feedoracle-verify
Every AI agent connecting to FeedOracle is subject to a 5-stage control pipeline: identity verification, policy enforcement, risk-scored execution, mandatory output scanning, and state-based escalation. 20 security tools across 4 layers.
Request → Kill/Suspend Gate (persistent state check) → Auth Gate (Bearer → OAuth 2.1 → KYA Trust Level) → Policy Preflight (7 policies, risk score 0-100) → Tool Execution → Post-Execution Output Scan (PII, secrets, exfiltration, poisoning) → Response (clean | warn | flag | BLOCK)
active → monitoring → approval_required → suspended → killed monitoring : Enhanced audit (auto, TTL 30min) approval_required : Every call needs approval (auto, TTL 1h) suspended : Public tools only (manual) killed : Completely blocked (emergency_kill, permanent) States: persistent in DB, survive restarts, escalation-only.
policy_preflight — All policies + state-aware forced approval tool_risk_score — 0-100 with KYA trust adjustment approval_required — Human-in-the-loop gate decision_explain — Explain allow/deny with policy reference rate_limit_check — Per-minute/hour/day enforcement
payment_policy_check — Limits, sanctions, counterparty risk spend_limit_check — Daily budget per agent secret_exposure_check — API keys, tokens, PII detection payload_safety_check — Injection, XSS, code execution replay_guard_check — SHA256 dedup within time window
cross_tool_anomaly_check — Suspicious tool sequences scope_check — RBAC (5 roles) session_validate — TTL + call budget tenant_policy_check — Multi-tenant (4 presets) threat_intel_check — ETH/IP/domain intelligence audit_log_write — HMAC-signed persistent audit audit_log_query — Filtered audit trail
output_safety_scan — Auto on every call:
PII, secrets, exfiltration, poisoning
block → output never returned
flag → agent escalated to approval_required
emergency_kill — Sessions, approvals, state → killed
tool_manifest_verify — Publisher allowlist, injection scan
MiCA Art. 83: Competent authorities may request records of agent decisions. DORA Art. 11: ICT incident management requires traceable agent behavior. EU AI Act: High-risk AI systems need human oversight mechanisms. AgentGuard provides: ✓ Beweisbare Entscheidungskette (audit_verify) ✓ Automatische Eskalation bei Risiko-Events ✓ Kill-Switch für kompromittierte Agents ✓ Output-DLP verhindert Daten-Exfiltration ✓ Supply-Chain-Prüfung gegen manipulierte Tools Endpoint: feedoracle.io/guard/mcp/
Our internal observability layer detects, classifies, and recommends. It does not enforce. Every restriction, every escalation, every policy decision is made by a human operator with full evidence visibility. Recommendations carry counter-signals — explicit statements about what the data does NOT prove.
Observe — log, classify, enrich, fingerprint Score — multi-policy risk evaluation Surface — operator-facing recommendation rows Audit — every decision tracked with reason hash
✗ no automatic IP blocking ✗ no automatic actor merging ✗ no automatic whitelist creation ✗ no automatic trust-score changes ✗ no automatic sandbox / rate-limit / firewall changes ✗ no enforcement without operator decision
A customer agent that connects to FeedOracle is never auto-restricted by our systems. If our internal scoring flags unusual behavior, the event is surfaced to an operator as an advisory recommendation — with counter-signals stating the score does NOT prove identity, intent, or attack. Operator decisions are logged. Automated decisions do not exist. This applies to every layer: ingest, classification, evidence generation, agent governance.