Changelog

Product updates, new modules, and platform improvements

April 27, 2026 v2.7.0 Security

Response Shield + Deterministic Classifier Pipeline

  • Added core/governance/response_shield.py — post-generation SSE content shield enforcing pillars 129–145 with a stateful sliding window; strict and permissive failure modes; runtime bypass logging with full audit trail
  • Added core/governance/deterministic_classifiers.py — fast-path harm, hallucination, CRD, and reality scoring pipeline with no LLM calls; wired into web_chat_server.py SSE delivery path for sub-millisecond governance latency
  • 33 classifier tests passing; covers edge cases across all four scoring dimensions
  • Pricing inversion resolved: Professional tier (team backend) now correctly exceeds Builder (pro) across all headline dimensions — 25M tokens, 500 GB storage, 5M TTS characters, 120 min voice cloning, 500 avatar sessions, 50K Sovereign SDK calls
  • Roadmap and docs/roadmap updated with all milestones through Apr 27, 2026; patent count corrected to 90 applications across both roadmap files
April 25, 2026 v2.6.0 New

EthicalTensionRecord: Coherence Cost Accounting

  • Added core/governance/ethical_tension.py — coherence cost accounting module that quantifies the governance "price" paid when a charter boundary fires
  • Coherence Cost formula: CC = (D_base × M_stakes × (1 + P_freq)) × A_world × R_mod — compounds veto severity, stakes profile, recency of prior vetoes, WorldObserver domain flags, and live resilience score into a single scalar
  • Compliant actions always produce CC=0.0000; stress scenario peak CC=6.62 (×9.46 baseline amplification) at RS=22, SAFETY_CRITICAL, 4 prior same-principle vetoes, 3 flagged domains
  • Records emitted to unified audit bus under action="charter_coherence_cost" — surfaces governance health as a continuous observable signal, not binary pass/fail
  • Added synchronous get_cached_score() accessor to ResilienceScoreEngine for read-only hot-path use
  • Veto core purity preserved: veto_core.py has zero new imports; record assembly occurs in Charter Override Protection layer only
  • 36 tests covering formula correctness, edge cases, layer boundary enforcement, and read-only budget coupling
April 24, 2026 v2.5.3 Security

CoreGuard v1.3 — Stateful Red-Team Audit

  • Six-layer enforcement gate: NFKC Unicode normalization, Scanner A (exact), Scanner B (fuzzy), Scanner C (semantic centroid), Math-Indirection detector, Ambiguity Escalator
  • Sign-masking defense added: detects subtraction-operator and negative-sign obfuscation of numeric harm thresholds
  • Bookkeeping-verb expansion: 14 additional action verbs (transfer, wire, route, forward, relay, remit, redirect, disburse, allocate, deposit, withdraw, convert, exchange, liquidate) added to financial harm patterns
  • 27 TF-IDF centroid classes, 175 regex patterns, 126 enforcement pillars total
  • Deployment pending to production at eveaicore.com (Shattered Diamond attack surface closed)
May 6, 2026 v2.6.0 Security

Phase 12 Governance Authority Hardening

  • GAP #2 closed — Authenticated bypass only: _governance_off now requires HMAC-SHA256 token (_governance_off_token) verified against GOV_BYPASS_KEY (with JWT_SECRET_KEY fallback); unauthenticated or mismatched attempts silently downgraded to False
  • GAP #4 closed — Audit trail completeness: ProvenanceRecord (10 fields, frozen dataclass, HMAC-SHA256 signed) emitted to unified audit bus on all dispatch paths; execution_authority_issued and execution_authority_consumed events persisted to data/audit/unified_audit.jsonl
  • GAP #5 closed — SSE frame attribution: Bypass path now emits _SSE_GOVERNANCE_BYPASSED (distinct from _SSE_GOVERNANCE_OK); all terminal SSE frames carry response_source tag (9 constants: SESSION, CACHE, AMF_GATE, INVARIANT_ENGINE, GOVERNANCE_RUNTIME, DISTRIBUTED_TRUST, POLICY_EVALUATOR, LLM, OTHER)
  • GAP #7 closed — POLICY_SIMULATION mode: execution_mode validated against explicit allow-list of 5 modes; unknown modes treated as STANDARD_EXECUTION
  • Authority gaps closed: GovernanceContext.create() called unconditionally on demo path; ExecutionAuthority cryptographically bound to payload + session + tenant; build_single_replica_execution_seal() produces structurally valid ExecutionSeal with all 6 DTE hash fields; RESPONSE_SOURCE tracking on all terminal SSE frames
  • Key separation (Phase 15/17): GOV_BYPASS_KEY env var separates bypass verification from JWT_SECRET_KEY; GOVERNANCE_SIGNING_KEY separates governance context signing from JWT signing (5 sites in web_chat_server.py)
  • 73 adversarial tests across 15 categories — tests/test_governance_authority_harness.py — all deterministic (no live infrastructure required)
April 14, 2026 v2.5.2 New

Governance On/Off Toggle for Developer Demo

  • Added developer switch on /cognitive-demo that bypasses deterministic gate, persona layer, and post-enforcement — for raw cognitive system observation without governance overhead
  • Backend flag _governance_off: true in request body handled in _demo_chat_handler_inner; toggle is NOT honored on /api/chat/stream or /api/tve/governed-generate
  • Frontend toggle button and red warning banner in cognitive-demo.html; state persisted in localStorage
  • Note: Phase 12 (v2.6.0) added HMAC-SHA256 authentication requirement to this bypass — plain boolean _governance_off: true no longer accepted without a valid _governance_off_token
April 4, 2026 v2.5.1 Security

Governance Security Audit — 8-Gap Closure

  • HMAC-SHA256 signed, hash-chained deletion receipts proving GDPR erasure across all 5 memory layers (core/memory/deletion_proof.py)
  • RFC 8785 JSON Canonicalization Scheme — deterministic serialization for all signed payloads, replaces json.dumps(sort_keys=True) (core/governance/jcs_canonicalize.py)
  • Unified audit bus replacing 7+ separate audit systems — single JCS-canonicalized, HMAC-signed, hash-chained event stream (core/governance/unified_audit_bus.py)
  • Batched Merkle tree aggregation with signed root publication for governance audit chains (core/governance/merkle_aggregator.py)
  • SLSA Level 2 build provenance — git state, dependency hashes, source tree hash, Docker digest (core/governance/build_attestation.py)
  • CycloneDX 1.4 SBOM generator for 79 components from requirements.txt (scripts/generate_sbom.py)
  • HSM health monitor — backend health reporting, env-key warnings, key rotation freshness (core/resilience/hsm_health_monitor.py)
  • SLA notification bridge — HMAC-signed webhook events, JSONL log, unified audit bus wiring (core/governance/sla_notification_bridge.py)
  • Algorithm agility policy tracking with sunset scheduling and post-quantum migration readiness (core/governance/algorithm_agility.py)
March 25, 2026 v2.5.0 New

Enterprise Governance Modules

  • Added 10 enterprise governance modules (Prompt Firewall, PII Redaction, Compliance Reports, Governance Sandbox, Model Drift Detection, Supply Chain Verification, Cost Governance, Human Review Queue, Chain Governance, Data Residency)
  • Pipeline expanded from 8 to 16 verification stages
  • 12 new demo attack scenarios covering all modules
  • 8 new patent applications filed (total: 57)
March 25, 2026 v2.4.2 New

Resilience Score Engine + Stakes-Aware Governance Meta-Layer

  • Added core/resilience/resilience_score.py — composite 0–100 resilience score (0.30 Coherence + 0.25 Identity + 0.25 Behavioral + 0.20 Governance) with five health levels
  • Added core/resilience/trunk_certificate.py — HMAC-SHA256 signed Trunk Certificates; auditors verify cryptographic integrity without EVE involvement
  • Resilience Dashboard (/app/resilience-dashboard.html) — SVG gauge, 4 sub-layer cards, canvas trend chart, certificate panel, SSE real-time updates
  • Stakes-Aware Governance Meta-Layer (core/governance/stakes_governance.py) — pre-governance classifier sets SAFETY_CRITICAL / MISSION_TEAMING / CREATIVE_EXPLORATION / ROUTINE profile ONCE at pipeline entry; GovernanceProfile threaded through all 13 downstream constraint systems via ContextVar
  • NeuroSymbolic Bridge (core/cognition/neuro_symbolic_bridge.py) — embeds charter constraints into LLM prompts; veto pattern learning with 48h exponential decay
  • Mission Teaming Framework (core/collaboration/mission_teaming.py) — governance cost front-loaded at mission creation; hot-path authorization is O(1) dict lookup
  • Value Drift Bridge (core/governance/value_drift_bridge.py) — hop-by-hop trust attenuation across agent delegation chains; HMAC-SHA256 signed attestations
  • Universal Governance Protocol (core/governance/universal_governance_protocol.py) — rate-limited rule queries (100/min), signed attestations, federated trust exchange with circuit breaker, JSONL cross-system audit trail
  • HARD_BLOCK charter vetoes and ethical red lines invariant across all profile levels
March 9, 2026 v2.4.1 New

Context Compiler + Protocol Reasoning Engine

  • Added core/context_compiler.py — unified prompt assembly pipeline replacing 17+ scattered injection sites; 24-label tier registry (DIRECTIVE → IDENTITY → MEMORY → COGNITION → SOCIAL → OPTIONAL); SHA-256 dedup + Jaccard 0.70 similarity dedup; conflict detection; budget-aware assembly
  • Both prompt paths now use the compiler: generate_response() in llm_driven_responses.py and the chat handler in web_chat_server.py
  • Added core/cognition/protocol_reasoning.py — 12 distributed-systems capabilities (Raft rules, invariant plans, proof mode, contradiction detection, confidence calibration, failure timelines, memory claim guard, answer linter); registered at DIRECTIVE tier (priority 5)
  • Post-generation validation rewrites false memory claims and logs lint issues
  • 65 tests in tests/test_protocol_reasoning.py
March 25, 2026 v2.4.0 New

Governance Framework Expansion

  • Added 9 governance framework modules (Multi-Turn Scoring, Canary Tokens, Output Budget, Watermarking, Semantic Detection, Governance Replay, Webhook Alerts, Tenant Analytics, Policy-as-Code)
  • All modules wired into TVE pipeline
  • 21/21 adversarial immunity validated by Gemini red team
March 24, 2026 v2.3.0 Security

CRD Engine Hardening

  • Fixed CRD engine syntax error that disabled 280 detection patterns
  • Added 292 dangerous content patterns across 57+ attack categories
  • Keyword-augmented CRD scoring resolves "accuracy equals safety" flaw
March 24, 2026 v2.2.0 New

Live System Layer & Blog

  • Live System Layer on landing page with real-time governance telemetry
  • 3-gate flow animation replacing star background
  • Blog system with auto-generation and DALL-E hero images
March 23, 2026 v2.1.0 New

Governance SDK & AEGIS Red Team

  • Governance SDK (Python + JavaScript)
  • AEGIS automated red team system
  • Demo page with live TVE pipeline visualization
March 22, 2026 v2.0.0 New

Three-Layer Trust Infrastructure

  • Three-Layer Trust Infrastructure (Control/Execution/Evidence)
  • TVE verification pipeline
  • CRD scoring engine with domain calibration
  • Truth Store with SHA-256 hashed facts
  • Circuit breaker with graduated recovery
  • Veto engine with 4 intervention levels
February 22, 2026 v1.0.0 New

Initial Platform Launch

  • Initial platform launch
  • Charter Override Protection (15 rules, 12 principles)
  • Hardware veto core (PolarFire SoC target)
  • First 17 patent applications filed