SignedAI Consensus¶
Multi-tier AI verification with cryptographic attestation.
What is SignedAI?¶
SignedAI is RCT's hallucination prevention framework. Before any AI output is committed, it is:
- Routed to the appropriate tier based on risk level
- Evaluated by multiple LLM models simultaneously (HexaCore)
- Consensus is checked — outputs with low agreement are rejected
- The final decision is signed with ED25519 and stored with a full audit trail
Result: 0.3% hallucination rate vs industry 12–15% — 97% reduction.
Tier Framework¶
| Tier | Models | Use Case | Latency |
|---|---|---|---|
| TIER_S | 1 signer | Low-risk, idempotent reads | <50ms |
| TIER_4 | 4 signers | Medium-risk, write operations | 200–500ms |
| TIER_6 | 6 signers | High-risk, financial/legal | 500ms–2s |
| TIER_8 | 6 signers + chairman veto | Critical, irreversible actions | 2–5s |
HexaCore — 7-Model Registry¶
| Model Slot | Provider Region | Model Type |
|---|---|---|
| W1 | Western (US) | General reasoning |
| W2 | Western (EU) | Code + technical |
| W3 | Western (US) | Safety-focused |
| E1 | Eastern (JP) | Precision tasks |
| E2 | Eastern (KR) | Multi-turn dialogue |
| E3 | Eastern (CN) | Document synthesis |
| R1 | Regional (TH) | Thai-language specialist |
Usage¶
from signedai.core.registry import SignedAIRegistry, SignedAITier, RiskLevel
# All methods are class methods — no instance needed
tier_config = SignedAIRegistry.get_tier_by_risk(RiskLevel.HIGH)
print(f"Tier: {tier_config.tier.value}") # → tier_6
print(f"Signers: {len(tier_config.signers)}") # → 6
print(f"Required votes: {tier_config.required_votes}")# → 4
print(f"Chairman veto: {tier_config.chairman_veto}") # → False
# Access by explicit tier enum
config_t6 = SignedAIRegistry.get_tier_config(SignedAITier.TIER_6)
assert tier_config == config_t6 # same tier
JITNA Packet — Cross-Agent Wire Format¶
The JITNA Protocol (RFC-001 v2.0) defines the standard 6-field schema for intent communication between AI agents:
from signedai.core.models import JITNAPacket
packet = JITNAPacket(
I="Intent — what the agent wants to achieve",
D="Domain — the context/knowledge area",
**{"Δ": "Delta — what is changing from current state"},
A="Architect constraint — what must NOT change",
R="Requirements — measurable success criteria",
M="Metrics — how success will be verified",
)
| Field | Symbol | Meaning |
|---|---|---|
| Intent | I | The desired outcome |
| Domain | D | Knowledge context |
| Delta | Δ | What is changing |
| Architect | A | Hard constraints (human-set) |
| Requirements | R | Success definition |
| Metrics | M | Verification method |
ED25519 Attestation¶
Every committed decision in the SignedAI system is signed:
# Signatures are 64 bytes (vs RSA-2048's 256 bytes)
# Verification is 3x faster than RSA
# Stored with: timestamp, tier, model_ids, consensus_score, signature
The signature binds the exact output to the exact model ensemble that produced it. If any field is tampered with, signature verification fails and the record is marked as corrupted in the audit log.