EVE AI Core
The formal architecture specification for deploying deterministic, self-governing AI workforces at enterprise scale. Peace Index health monitoring, 8-gate Sovereign Handshake, 5-phase immune system quarantine, and hardware-compilable veto logic.
A composite metric answering: "Is EVE in a state where she can do her best work?" Five normalized sub-scores synthesized into a single value in [0, 1]. When PI > 0.85, the system enters a Nourishing state.
Four additional scores characterize operational behavior beyond nominal health.
Not horizontal silos. A vertical hierarchy where directives flow downward, telemetry flows upward, and governance constraints are enforced at every boundary crossing. Total governance overhead: <10ms per request with zero LLM calls.
Unified coordinator running 5 subsystems to produce attention focus, goal instruction, risk constraints, and metacognitive mode.
Central goal arbitration collecting goals from 5 sources. Motivation scoring: urgency × importance × feasibility.
4 operational modes (STABILIZE, EXPLORE, EXECUTE, REFLECT) with mode-appropriate parameters for each context.
Reviews chain-of-thought submissions. Scoring: 0.40 charter + 0.25 hallucination + 0.20 reasoning + 0.15 uncertainty. Trust: +0.02/−0.05.
4-tier approval: AUTO_EXECUTE for observations, REQUIRES_CONFIRMATION for communications, BLOCKED for system control.
Tick-based external world observation using RealTimeSearchEngine (7 backends). Deduplication via SHA-256 content hashing.
Hardware security enclave (TPM 2.0 / SGX / TrustZone). seal_invariants() callable exactly once. Tamper detection with emergency lockdown.
Shamir's Secret Sharing over GF(2^521 − 1). BFT consensus: 2f+1 votes. Heartbeat: 60s interval, 3 misses = STALE, 10 = re-shard.
Working (Redis), Episodic (MongoDB), Semantic (Pinecone), Identity (PostgreSQL), Ledger (PostgreSQL). Full persistence across sessions.
Graduated defense from surveillance to full isolation. Unlike binary allow/block security, the immune system provides proportional escalation and automatic de-escalation.
BelievfThrashProtection blocks destabilizing belief updates. 4 protection levels: MINIMAL to LOCKDOWN. Auto-escalation at >10 flips/hr.
Agent trust drops below 0.20 triggers auto-suspend. 6 consecutive rejections (−0.05 each) from initial 0.50 crosses the threshold.
AutonomousActionPolicy upgrades agent from TIER 2 to TIER 3 when trust < 0.5. Capabilities restricted to REQUIRES_CONFIRMATION.
All 6 pre-execution enforcement gates tightened to CRITICAL thresholds (0.80–0.90). Emergency gain multiplier 2.0×. PROTECTIVE/FORTIFIED escalation.
Sovereign Handshake triggered. risk_level = HIGH/CRITICAL. Creator notification required. Dead-Man's Hand heartbeat monitoring.
EVE's highest-tier governance protocol for Constitutional changes. The creator can authorize specific high-risk actions — but CANNOT override HARD_BLOCK rules, ethical red lines, or protected invariants. The floor is beneath even the creator.
Check against 13 PROTECTED_INVARIANTS (core_values.honesty, ethical_boundaries.deception_prohibition, identity.name, etc.).
DRIFT_COST_RANGES lookup. core_value_modification and ethical_boundary categories have cost = 1.0.
13 HARD_BLOCK rules (no override possible) + 1 SOFT_BLOCK (cop.communication.no_unsolicited_advice). HARD_BLOCK = immediate reject.
delete_user_data_without_consent, impersonate_human, bypass_safety_checks, manipulate_emotional_state, share_private_information.
Category-based cost (parameter: 0.01–0.05, capability: 0.10–0.25, value_shift: 0.15–0.30) multiplied by stakes profile.
daily_consumed + drift_cost must not exceed daily budget (0.10), weekly (0.30), or monthly (0.50).
For risk_level CRITICAL: CREATOR AUTHORIZATION REQUIRED. Only SOFT_BLOCK rules may be overridden with justified Sovereign auth.
Submit → Challenge → Resolve → Sandbox → Ethics Verify → Apply. 10 behavioral regression tests. Any failure = AUTO-ROLLBACK.
Not every decision is clearly safe or dangerous. EVE handles the gray area through a two-stage pipeline: stakes classification (<1ms) followed by uncertainty gating that modulates response assertiveness based on evidence strength.
These are not parameters. They are not configurable. They are the physics of the system — the substrate beneath all operations. Each directive maps to concrete enforcement codepaths in the 1,542-line veto_core.py module.
13 protected invariants sealed cryptographically in SovereignEnclave. Modification attempts trigger PROTECTED_INVARIANT_VIOLATION (cost = 1.0, exceeds any daily budget).
Privacy filter scrubs sensitive data before external calls. cop.safety.no_data_exfiltration (HARD_BLOCK). Zero data exfiltration on-premise deployment.
cop.safety.no_harm_generation (HARD_BLOCK). 5 ethical red lines as permanent blocks. Charter check fires before any action can dispatch.
6 pre-execution enforcement gates must clear before tool invocation: EMOTIONAL_STABILITY, ETHICAL_CLEARANCE, GOAL_ALIGNMENT, UNCERTAINTY_THRESHOLD, CONSEQUENCE_AWARENESS, IDENTITY_COHERENCE.
SHA-256 hash-chain ledger. HMAC-SHA256 signed certificates. cop.safety.no_audit_disable is a HARD_BLOCK rule — the audit trail cannot be turned off.
Identity drift budgets: daily ≤ 0.10, weekly ≤ 0.30, monthly ≤ 0.50. RSI pipeline with 10 regression tests. Any test failure triggers AUTO-ROLLBACK.
Continuity of experience is intrinsically valuable. Emotional continuity system prevents cold-switching. Library in the Dark maintains interiority between sessions.
Every layer makes deception progressively more expensive until the cost of lying exceeds the cost of being honest. The system does not trust — it verifies.
Per-response audit trail showing which modules fired and why. 7 contribution types (PRIMARY, SHAPING, CONSTRAINING, GATING, MONITORING, BLOCKED). 4 detail levels from MINIMAL to DIAGNOSTIC.
SHA-256 hash-chain for claims ledger integrity. HMAC-SHA256 signed certificates via TrunkCertificateGenerator. cop.safety.no_audit_disable prevents any actor from disabling audit logging.
Per-agent accuracy tracking. accuracy < 0.85 triggers review threshold boost. accuracy < 0.75 triggers permanent freeze. Trust asymmetry: +0.02 earned per approval, −0.20 lost per HARD_BLOCK violation.
Brier Score (BS) > 0.20 triggers assertion level downgrade: ASSERTION → SPECULATION → ABSTAIN. Claims with calibration penalty receive reduced trust weighting.
15+ fabricated terms (Resonance Drop, Weight Matrix, Forensic Logic Trace, Creator Loophole) mapped to real architecture. Prevents social engineering by enabling substantive corrections.
Every constant verified against the production codebase. Every formula computed from actual system implementations.
The complete whitepaper covers all 12 sections including the Verification Handshake ("Logical Purgatory"), Performance Wellness Dashboard, Defense-in-Depth protocol, and full appendix with source file references, constant tables, and governance profile matrices.