Experimental cognitive architecture for testing verifiable state transitions, representation mechanisms, and consciousness-relevant computational indicators — paired with a zero-knowledge proof stack for privacy-preserving attestation.
Symthaea combines hyperdimensional computing (HDC), closed-form liquid neural dynamics (CfC), consciousness-relevant measurement (IIT-inspired Φ, Butlin-indicator probes), and zero-knowledge proofs. It does not claim to prove consciousness, and nothing below is a regulatory compliance certification. What it does have: working ZK circuits over HDC's binary-field structure, measured against a prime-field baseline, and an internal research program that documents its own retractions — see the Evidence table below before trusting any specific number.
Live Demo | Paper | Benchmarks | Honest Status
| Claim | Status | Evidence | Limitations |
|---|---|---|---|
| HDC XOR binding: 256x fewer ZKP constraints than a prime-field baseline (Binius vs. Winterfell, 16Kbit) | Reproduced internally | papers/binius-hdc/reproduce.sh — real proofs, not estimates |
Same-scale comparison only; not audited by a third party |
| Health-value range proof in ~34ms (ZK range proof, value ∈ [min, max]) | Working cryptographic primitive | crates/mycelix-zkp-core/, benchmark output below |
Not a HIPAA compliance certification. A range proof is one building block; no accredited assessment of any system built on it has been performed |
| Selective-disclosure age≥18 proof | Working cryptographic primitive | papers/binius-hdc/ eIDAS-shaped demo |
Not an eIDAS 2.0-certified wallet or credential. Demonstrates the primitive an eIDAS-compliant system would need, nothing more |
| Butlin consciousness-indicator probes | Structural implementation, partially audited | Internal construct-validity re-check found only 4/12 probe rows survive; of those, one (AE-2) has real empirical support to date, and that result is explicitly scoped as "causally-supported internal wiring evidence," not functional capacity and not evidence of consciousness |
Not evidence the system is conscious. Most of the original 12-row matrix did not hold up to its own re-audit |
| Moral-reasoning benchmark (Hendrycks Ethics) | Internal benchmark, re-run and corrected | Canonical 2026-07-15 re-run on held-out test CSVs: 56.2% overall; only the virtue category (76.5%) is meaningfully above chance |
An earlier 91.1%/94.5%/84.7% figure was retracted as leakage-inflated — if you see those numbers anywhere else in this repo's history, they are superseded |
| Cognitive loop runs at ~31 Hz measured | Reproduced internally | Loop telemetry | Measured on one reference machine; not a real-time guarantee |
None of this is a claim of general intelligence, sentience, or legal/medical/financial compliance. Where a component is a real, working, reproducible cryptographic or engineering result, we say so and show the command. Where a result was wrong, we say that too, in place, rather than deleting the history.
- Healthcare attestation primitive — Prove "patient vitals are in normal range" in ~34ms without revealing the actual values, via a ZK range proof. (health-zkp)
- Anonymous eligibility proof — Prove "I'm eligible to vote" without revealing identity, with post-quantum Dilithium5 signatures. (governance)
- Selective credential disclosure — Prove "age >= 18" via ZKP without revealing date of birth or nationality. (eidas-shaped demo)
- Verified federated learning — 16 participants submit encrypted gradients; aggregation proven correct in ~91ms. (triple-stack)
- Verifiable search — Prove search results were ranked per a stated formula, not manipulated. (prism-zkp)
- Cognitive-loop state-transition proof — Prove a specific internal state transition occurred as specified, without revealing the internal state itself. This is a proof about a computation, not a proof about consciousness. (cycle_integrity)
git clone https://github.com/Luminous-Dynamics/symthaea.git
cd symthaea
# Enter the pinned development shell
nix develop
# Build
cargo build --release
# Run tests (21K+ workspace-wide)
cargo nextest run --workspace --lib
# Run ZKP benchmarks
cd crates/mycelix-zkp-core && cargo test --features full --lib
# Run Binius HDC benchmark (needs .deps/binius64)
cd crates/hdc-zkp-bench && cargo run --release
# Try the Spore kernel in browser
open https://symthaea.luminousdynamics.io| Metric | Binius (16Kbit) | Winterfell (16Kbit) | Advantage |
|---|---|---|---|
| Constraints | 256 | 65,532 | 256x fewer |
| Prover time | 430 ms | 23,762 ms | 55x faster |
| Verifier time | 10.4 ms | 23.7 ms | 2.3x faster |
XOR binding — the core HDC operation — requires zero non-linear constraints in Binius. Both numbers are real cryptographic proofs, not estimates.
| Circuit | AND Gates | Prove Time | What It Proves |
|---|---|---|---|
| XOR binding (16Kbit) | 256 | 430 ms | HDC binding correctness |
| Bundling (3 vectors) | 768 | 925 ms | Majority vote aggregation |
| Hamming similarity | 511 | 498 ms | Search relevance |
| CfC temporal (64N×100T) | 7,360 | 954 ms | Neural network evolution |
| CfC + sigmoid | 13,760 | 680 ms | Full neural computation |
| FL encrypted (16 participants) | 1,024 | 91 ms | Encrypted gradient aggregate |
| Health range proof | ~32 | 27 ms | Value in [min, max] |
| Consciousness tier | ~32 | 46 ms | Phi >= threshold |
| Capability | Metric | Status |
|---|---|---|
| Cognitive loop frequency | 31 Hz measured (234 Hz raw) | Reproduced internally |
| Butlin indicator probes | 4/12 rows survive construct-validity re-check; 1 has empirical support so far, scope-limited | Not evidence of consciousness — see Evidence table |
| Moral reasoning (Hendrycks Ethics) | 56.2% overall (virtue: 76.5%) | Corrected 2026-07-15; supersedes an earlier retracted 91.1% figure |
| HDC dimension | 16,384D binary + continuous | — |
| WASM kernel | 324 KB (runs in browser) | — |
| Workspace | 87 crates, 21K+ tests | — |
Core crates:
| Crate | What | Tests |
|---|---|---|
symthaea |
Main cognitive loop (4-phase: perceive → evolve → measure → act) | 7K+ |
symthaea-core |
HDC vectors, CfC neurons, consciousness math | 3K+ |
mycelix-zkp-core |
ZKP infrastructure: backends, circuits, Dilithium5, domain tags | 65+ |
hdc-zkp-bench |
Binius benchmark circuits (XOR, bundling, CfC, FL, Hamming) | Benchmarks |
symthaea-zkproof |
Consciousness attestation (RISC0) + cycle integrity | 23+ |
symthaea-spore |
Browser WASM kernel (324 KB) | 183 |
symthaea-broca |
Language generation (CfC-HDC thought-to-text) | 229+ |
Feature flags: 98 flags, default = minimal. Key flags: reasoning_engine, identity, neural-bridge, ssm_language. See FEATURE_FLAGS.md.
Contributing: See CONTRIBUTING.md. Start with docs/START_HERE.md.
| Paper | Venue | Status |
|---|---|---|
| Binary-Field STARKs for HDC | IEEE S&P 2027 | Submission-ready (v5) |
| CfC Temporal Proofs | NeurIPS workshop | 3-page draft |
| Triple-Stack FL | PoPETs | 3-page draft |
| HAI: Hyperdimensional Active Inference | PLoS Comp Bio | Manuscript complete |
| Genesis: Ethical Frameworks | AI & Ethics | Manuscript complete |
| Psych-Bench | — | 98 benchmarks, 20 domains |
# Reproduce all ZKP benchmark results
cd papers/binius-hdc && bash reproduce.sh
# Run consciousness psych-bench
cargo run --features psych_bench --example run_all_benchmarksUse the flake shell for day-to-day development:
cd symthaea
nix developInside the shell:
# Fast, isolated test runner
cargo nextest run --workspace --lib
# Focused crate test run
cargo nextest run -p symthaea-core --lib
# Background compile/watch loop
bacon
# Narrow spike for EML/e-graph work
cargo test -p symthaea-eml-egraph
# Real-data conjecture quality dashboard
cargo run -p symthaea-core --example real_discovery_report --features abstract_thought
# Same dashboard, but exit non-zero when health thresholds are breached
cargo run -p symthaea-core --example real_discovery_report --features abstract_thought -- --strictThe symthaea-eml-egraph crate is an isolated egg prototype for equality
saturation experiments around EML normalization. It is intentionally separate
from production conjecture ranking and verification logic.
For discovery work, prefer real_discovery_report as the project-health
baseline. It summarizes real observation families by verification rate, EML
coverage, macro-candidate pressure, promoted operators, suspicious formulas, and
runtime. Use --strict only for opt-in CI guardrails; the current default is a
soft warning report.
psych-bench is an internal benchmark suite. For the honest per-module status of math and science capabilities (Lie theory, Langlands, thermodynamics, persistent homology, frontier physics, etc.), see MODULE_STATUS.md. External-benchmark results (miniF2F, PutnamBench, ARC-AGI-2) are planned for Phase 1.
- 256x constraint reduction for HDC XOR binding in binary-field STARKs (measured, same scale)
- Zero encryption overhead for federated learning (OTP + Binius share GF(2))
- 694 KB Binius verifier compiles to WASM (in-zome verification possible)
- 4/12 Butlin indicator probes survive an internal construct-validity re-check; only 1 has empirical support to date, explicitly scoped as internal causal-wiring evidence, not consciousness evidence
- 56.2% moral accuracy on Hendrycks Ethics (canonical re-run; an earlier 91.1% figure was retracted as leakage-inflated)
┌─────────────────────────────────────┐
│ Symthaea Cognitive Loop │
│ Perceive → Evolve → Measure → Act │
│ (~31 Hz, 4-phase) │
└──────────┬──────────────────┬────────┘
│ │
┌──────────▼──────┐ ┌────────▼────────┐
│ HDC Encoder │ │ CfC Dynamics │
│ 16,384D binary │ │ Closed-form LTC │
│ + continuous │ │ O(1) per step │
└──────────┬──────┘ └────────┬────────┘
│ │
┌──────────▼──────────────────▼────────┐
│ DASTARK ZKP Layer │
│ Binius (HDC) + Winterfell (range) │
│ + RISC0 (complex) + Dilithium5 (PQ) │
└──────────┬──────────────────┬────────┘
│ │
┌──────────▼──────┐ ┌────────▼────────┐
│ Mycelix (DHT) │ │ Applications │
│ 9 hApps on │ │ Health, Identity│
│ Holochain │ │ Finance, DeSci │
└─────────────────┘ └─────────────────┘
See full architecture diagrams in docs/architecture/.
None of the following is a compliance certification. These are working cryptographic primitives that a compliant system in each domain would plausibly need — no accredited assessment, legal review, or regulatory audit has been performed on any of them.
| Regulation | What the primitive demonstrates | What it is NOT |
|---|---|---|
| HIPAA (US) | ZK range proof: vitals/labs in range without revealing values (~34ms) | Not a HIPAA-compliant system or a covered-entity certification |
| eIDAS 2.0 (EU) | Selective-disclosure credential proof with dastark-2026 cryptosuite + Dilithium5 |
Not an eIDAS-certified wallet, issuer, or trust-service provider |
| EU DPP (EU) | Proof of a compliance predicate without revealing full supply-chain data | Not a Digital Product Passport implementation or conformity assessment |
If you are evaluating this for an actual regulated deployment, treat everything here as a research prototype and get your own legal/compliance review.
AGPL-3.0-or-later. Commercial licensing available — contact tristan.stoltz@evolvingresonantcocreationism.com.
@misc{stoltz2026biniushdc,
title={Binary-Field STARKs for Hyperdimensional Computing},
author={Stoltz, Tristan},
year={2026},
note={Available at github.com/Luminous-Dynamics/symthaea/papers/binius-hdc}
}