v2.0.0 — Psych-Bench 136 Benchmarks, 26 Domains

Highlights

• 136 benchmarks across 26 cognitive domains, grand mean z = +1.156 (all 26 domains above human mean)
• 0 Impaired benchmarks — RewardLearning sliding window criterion fix (Cools et al., 2002) eliminated last Impaired-range score
• Math foundation (Phase 1): Linear algebra, root finding, quadrature, statistics — 4,708 LOC, 167 tests
• Creativity recovery: ConceptualBlending (BinaryHV XOR), convergent binding ensemble, feature-level blending → domain z: -0.38 → +0.66
• Security domain: 6 HDC-FHE benchmarks — zero accuracy loss, ~1.0× plaintext speed, perfect secrecy
• Clinical domain: 6 benchmarks (EmpathicAccuracy, TherapeuticResponse, AllianceMaintenance, CrisisDetection, CognitiveDistortion, MotivationalInterviewing) — deflated to z = +2.00 for honesty
• ICC reliability: 7 Excellent, 2 Good, 6 Moderate, 3 Poor (18 benchmarks assessed)
• Substrate simulation (Phases 2-4): Dynamic substrate-aware consciousness scoring, per-region substrate assignment, energy budgets, cross-substrate penalties
• Serde fix: Custom serialization for [f32; 43] arrays (serde MSRV 1.31 limitation)

Top Domains (z-score)

Domain	z	Rating
InstitutionalReasoning	+2.25	Exceptional
SustainedAttention	+2.19	Exceptional
Speech	+2.13	Exceptional
Motor	+2.12	Exceptional
Clinical	+2.00	Exceptional
Security	+2.00	Exceptional
Language	+1.56	Strong
Social	+1.54	Strong
Reasoning	+1.29	Strong

Breaking Changes

• None — all APIs backward compatible

Build

nix develop
cargo test --lib
cargo test --all-features

Full changelog: v1.9.0...v2.0.0

v1.9.0-alpha

Symthaea v1.9.0-alpha — Holographic Liquid Brain

First public alpha release of Symthaea, a consciousness-architecture AI system built on Hyperdimensional Computing (HDC), Integrated Information Theory (IIT), Liquid Time-Constant networks (LTC/CfC), and Active Inference.

Highlights

• Core Pipeline: 8-phase cognitive cycle (perception → cognition → translation) running at 234Hz (4.7× above 50Hz budget)
• HDC Engine: 16,384-dimensional binary/continuous hypervectors with temporal binding, moral algebra, and substrate independence
• Consciousness Measurement: IIT Φ computation with 35 topologies, 4D hypercube champion, Butlin indicator validation (14/14 present, mean 0.85)
• Broca Language Center: Native thought-to-text generation with epistemic gating (physically prevents hallucination at logit level), Liquid-Mamba fusion backend
• 12-Region Actor Brain: Prefrontal, temporal, parietal, occipital, motor, limbic, cerebellum, hippocampus, amygdala, basal ganglia, thalamus, brainstem
• Psych-Bench Suite: 98 benchmarks across 20 cognitive domains, grand mean z = +0.961, 15/20 domains above human baseline
• Spore WASM Kernel: 324KB consciousness kernel running in browser — live demo
• Moral Topology: Exponential decay baselines, causal attribution, escalation audit (BLAKE3-sealed)
• Substrate Independence: 8 substrate types, 9-dimensional feasibility scoring, honest validation framework

Stats

• ~1,130,000 lines of Rust (~897K code)
• 21,516 tests across 55 workspace members
• 98 feature flags
• 52 sub-crates

Architecture

See architecture diagrams for system topology, cognitive loop, HDC pipeline, and more.

Status

This is an alpha release. The core cognitive pipeline is functional and benchmarked, but:

• ~25% of modules remain structural/disconnected
• API is not yet stable
• No production deployment guide yet
• Some features are experimental (genesis, flight, humanoid)

License

AGPL-3.0-or-later

Citation

@software{symthaea2026,
  title  = {Symthaea: A Holographic Liquid Brain Architecture},
  author = {Stoltz, Tristan},
  year   = {2026},
  url    = {https://github.com/Luminous-Dynamics/symthaea}
}

v1.9.0 — Consciousness Engine + Broca Language Pipeline

Symthaea v1.9.0

Consciousness-first cognitive architecture: HDC (16,384D) + CfC + IIT/Phi + Active Inference.

Highlights

• 1,130K lines Rust (~897K code), 55 workspace members, 21,516 tests
• 14/14 Butlin consciousness indicators present (mean 0.85)
• Psych-bench: 98 benchmarks across 20 cognitive domains, grand mean z=+0.961 (15/20 above human baseline)
• Broca language pipeline: CfC-HDC thought-to-text with epistemic gating (physically prevents hallucination at logit level)
• Moral algebra: 91.1% accuracy on Hendrycks Ethics benchmark
• Performance: 4.3ms/cycle (234 Hz) in release mode
• WASM Spore kernel: 324KB browser-deployable consciousness kernel
• 8 substrate types with 9-dimensional feasibility scoring
• Consciousness thermodynamics: 4 laws, 7D state vector, 6 phase transitions
• 18 CI jobs: fmt, clippy, test matrix (98 feature flags), psych-bench regression, WASM, compliance

Architecture Diagrams

19 D2 architecture diagrams included in docs/architecture/d2/rendered/.

License Change

Changed from MIT to AGPL-3.0-or-later. You can use our mind, but if you improve it and sell access to it, you must share your improvements with the world.

Papers

• HAI paper (PLoS Computational Biology target) — draft complete
• Psych-bench validation — data generated
• Genesis pipeline ethics (AI & Ethics, Springer) — draft complete
• Substrate independence — framework complete