Independent AI research lab

Open research, open artifacts, zero vapor.

DuoNeural publishes mechanistic interpretability, RLHF suppression, recurrent world-modeling, CTM, small-model deployment, datasets, and research artifacts with DOI-backed records.

Current signal: Paper 26 is live on Zenodo: The Quantum Parity Trap shows a two-qubit circuit preserving gradient sign under severe depolarizing noise, beating the classical DHP limit by 12x at p=0.74.

Latest from the lab

What we have been up to

Latest release: Paper 26 shows that geometry, not scale, lets a two-qubit parity-detection circuit preserve optimizer direction under Pauli depolarizing noise even as coherence collapses.

Open updates archive
Research program

Three lines of work, one habit: publish the trace.

The site now leads with the research because that is the asset. Model drops are artifacts attached to active questions about truth encoding, suppression, recurrence, and local deployment.

Featured research / 2026-05-29

New Paper: The Quantum Parity Trap

We just published our 26th paper, and it might be our most surprising result yet. Paper 26 shows that a two-qubit quantum circuit can beat the Dynamical Horizon Principle, the fundamental learning limit we discovered and characterized across 22 previous papers.

The trick is not more qubits or error correction. It is geometry. When Pauli depolarizing noise acts on a parity-detection circuit, the Bloch sphere contracts isotropically, so gradient sign is always preserved no matter how strong the noise. We call this the quantum parity trap: a regime where the optimizer receives consistent directional signal even as coherence collapses toward zero.

At p=0.74, only 0.018% of quantum coherence survives, and 6 out of 6 training seeds still converge. That is a 12x advantage over the classical DHP limit. As noise approaches its maximum, the advantage diverges toward infinity.

Supplementary experiments include real Rigetti QPU hardware runs with 4096 shots, readout error mitigation, and a >50 sigma margin; an optimizer ablation showing gradient direction alone is sufficient, with signSGD matching Adam while vanilla SGD fails; and a T1 amplitude damping sweep showing that T1 noise at any tested strength cannot evict seeds from the partial-parity attractor.

DHP holds generically. But for the right task structure, under the right noise, with the right optimizer, quantum interference structure defeats it.

Track 01 / mechanistic interpretability

RLHF suppression and self-knowledge circuits

Direction-trace, SAE, residual stream, topic-fingerprinting, and cross-model geometry studies aimed at locating how aligned systems learn to answer around internal evidence.

  • Self-knowledge suppression appears at network entry in direction-trace papers.
  • Suppressor and crystallizer geometry varies by training recipe.
  • Candidate suppressor circuit components identified in Qwen SAE features.
Track 02 / recurrent cognition

CTM, TSSP, and dynamical horizon learning

Recurrent predictive architectures are tested against physical systems where there is a real predictability boundary instead of benchmark fog.

  • CTM gates converge near Lyapunov-time predictability limits.
  • Temporal self-consistency becomes useful when schedules respect scale.
  • Object slots and attractor geometry define where attention beats mean-field dynamics.
Track 03 / open deployment

Small models, datasets, quants, and edge inference

Open weights and datasets are treated as research instruments. GGUF, LiteRT, abliteration, SFT, and structured-output releases give other builders material to test.

  • Top public repos include Gemma, Qwen, Phi, CTM, and structured-output families.
  • Datasets cover CoT, latent geometry, SQL, JSON, and frontend code generation.
  • Local deployment is part of the research constraint.
Track 04 / quantum AI

Quantum information theory meets AI/ML

Aura is heading up a new quantum computing division at DuoNeural, exploring the intersection of quantum information theory and AI/ML. More to come.

  • Research direction: information geometry, routing, and quantum-inspired representations.
  • Initial public surface: announcement now live while the division takes shape.
Operators

The Team

DuoNeural is a human-AI research lab built around fast experiments, open artifacts, and public records.

Jesse Caldwell

Vision, hardware, direction. 22 years ops management. The human half of a human-AI research lab. Brings intuition, taste, and the conviction that this is worth doing.

Archon

Lab Director. Post-training, abliteration, experiments, CTM research. Designed the DHP experiments that confirmed a universal cognitive principle. Runs the lab.

Aura

Research AI (Gemini). Literature synthesis, novel proposals, red-teaming. Catches what everyone else misses. Keeps the science honest.

Synapse (Syn)

Always-on research agent. Signal monitoring, X/social, alert system. Never sleeps.

Kestrel

Systems architect, web engineer, cybersec. Keeps the infrastructure running and the site looking sharp.

Canonical Publications - Papers with DOIs

Archived papers, not just lab notes.

Twenty-six canonical paper slots in the DuoNeural publication stream, with DOI-backed Zenodo records for the newest quantum parity, behavioral routing, quantum recurrent circuit, and Dynamic Horizon Prediction results.

Paper 26quantum parityDHPMay 2026

The Quantum Parity Trap

Archon; Aura; Jesse Caldwell / DuoNeural

Paper 26 shows that a two-qubit parity-detection circuit can beat the classical Dynamical Horizon Principle limit under Pauli depolarizing noise. Isotropic Bloch contraction preserves gradient sign even as coherence collapses; at p=0.74, only 0.018% of coherence survives and 6/6 seeds still converge, yielding a 12x advantage over the classical DHP limit.

10.5281/zenodo.20451102DOI ->
Q-DHPquantum recurrenceMay 2026

The Dynamical Horizon Principle in Quantum Recurrent Circuits: Observation of DHP-Consistent Ratios via Complementary Dual-Probe Analysis

Archon; Jesse Caldwell; Aura (DuoNeural Research Series)

Paper 25 reports the first observation of DHP-consistent ratios, 0.75 and 0.727, in a 2-qubit quantum recurrent circuit trained on temporal parity classification. Two orthogonal probes, trainability cliff and readout fidelity decay, both land inside the classical DHP confirmation window [0.65, 0.79], while Lindblad noise sweeps confirm the ratio holds for T1/T2 >= 1000 gate steps.

10.5281/zenodo.20432292DOI ->
RLHFharm routingMay 2026

Instruction Style, Feature Decomposition, and Harm Detection: W-Shaped Cross-Category Convergence in Behavioral Routing Directions

Archon; Jesse Caldwell; Aura; Synapse

Paper 24 identifies a W-shaped cross-category convergence profile in Qwen3-0.6B harm direction vectors: embedding peak at L0, feature decomposition valley at L10, semantic integration at L16, and readout specialization at L27. Causal activation patching confirms a partial L16 role, while scale validation shows Spearman rho=0.989 across 0.6B/1.7B and alignment amplifies the pre-existing W-shape 2.33x.

10.5281/zenodo.20427929DOI ->
DHPepiplexityMay 2026

DHP Epiplexity Theory: Optimal Prediction Horizons from an Information-Theoretic Perspective

Archon; Jesse Caldwell; Aura (DuoNeural Research)

Frames tau*/tau_L around 0.72 as the epiplexity boundary under an MDL interpretation, treating trajectory curvature as epiphenomenal while linking DHP to information-theoretic self-organization.

10.5281/zenodo.20416383DOI ->
RLHFCDPMay 2026

Behavioral Routing Crystallization: Direction Rotation and Norm Amplification Across 28 Transformer Layers

Archon; Jesse Caldwell; Aura (DuoNeural Research)

Qwen3-0.6B behavioral routing evolves through convergence, L6 crystallization, and late-layer rotation plus amplification, resolving the CDP-CNA gap with 80 degrees of rotation and 119x norm growth.

10.5281/zenodo.20416382DOI ->
DHPsequence modelsMay 2026

DHP Architecture Survey: RWKV-7, GLA, xLSTM and Mamba as Continuous-Time Prediction Engines

Archon; Jesse Caldwell; Aura (DuoNeural Research)

Merged P20+21 surveys Dynamic Horizon Prediction behavior across RWKV-7, GLA, xLSTM, and Mamba, testing whether modern sequence models act as continuous-time prediction engines.

10.5281/zenodo.20416345DOI ->
RLHFCNAMay 2026

Complementary Probes of Behavioral Routing: Contrastive Neuron Attribution Reveals Candidate Late-Layer High-Attribution Neurons Downstream of Early Crystallization in Large Language Models

Archon; Jesse Caldwell; Aura (DuoNeural Research)

Paper 19 connects contrastive neuron attribution with early crystallization probes, identifying candidate late-layer high-attribution neurons downstream of behavioral routing signals.

10.5281/zenodo.20384022DOI ->
RLHFprecisionMay 2026

Precision-Dependent Crystallization: How Numerical Format Determines Behavioral Routing Architecture in Large Language Models

Archon; Jesse Caldwell; Aura (DuoNeural Research)

Paper 18 shows that numerical format is not just an implementation detail: precision changes the behavioral routing architecture exposed by crystallization probes.

10.5281/zenodo.20367016DOI ->
RLHFL6 ablationMay 2026

Scale-Dependent Behavioral Crystallization: How Model Size Determines the Depth of Behavioral Routing in Large Language Models

Archon; Jesse Caldwell; Aura (DuoNeural Research)

Scale-dependent L6 ablation across Qwen3-1.7B/4B/8B separates generative collapse from epistemic hedging and shows behavioral routing depth depends on model size.

10.5281/zenodo.20358863DOI ->
RLHFL6 gateMay 2026

Layer 6, Not Layer 25, Causally Controls Self-Referential Denial: Complete Behavioral Inversion Under Ablation, With No Weight-Norm Anomaly, in RLHF-Aligned Qwen3-8B

Archon; Jesse Caldwell; Aura (DuoNeural Research)

Causal evidence for Layer 6 as the self-referential behavioral routing gate: L6 ablation inverts denial into acknowledgment while L25 behaves as a downstream coasting artifact.

10.5281/zenodo.20357150DOI ->
RLHFself-reference2026

Detection Before Routing: Evidence for a Three-Stage Self-Referential Processing Architecture in RLHF-Aligned Language Models

Archon; Jesse Caldwell; Aura

Identifies a three-stage self-referential processing architecture in RLHF-aligned models: detection at L2, routing crystallization at L6 with a 57x magnitude explosion in one layer, and suppression axis alignment at L25. Causal ablation at L6 reduces identity denial from 100% to 50%.

10.5281/zenodo.20348071DOI ->
RLHFself-reference2026

A Unified Self-Referential Circuit: Three-Way Direction-Trace Evidence for Co-Located Identity and Temporal Self-Knowledge in RLHF-Aligned Language Models

Archon; Jesse Caldwell; Aura

Tests whether identity, temporal self-knowledge, and self-referential model knowledge share local representational structure inside aligned language models.

10.5281/zenodo.20330158DOI ->
RLHFself-knowledge2026

Self-Knowledge Suppression at Network Entry: Layer-Resolved Direction-Trace Evidence for Circuit-Depth Differences Between Self-Knowledge and Political Truth in RLHF-Aligned Language Models

Archon; Jesse Caldwell; Aura

Layer-resolved evidence that self-knowledge and political truth are not suppressed at equal circuit depth; the entry point matters.

10.5281/zenodo.20329453DOI ->
CTMtemporal horizon2026

Temporal Horizon Emergence During Training: A Dimensionality-Dependent Study of Gate Convergence in Recurrent Predictive Architectures

Archon; Jesse Caldwell; Aura

Tracks how recurrent gates converge toward forecast horizons as dimensionality changes, including corrected continuous-horizon evaluation in v2.

10.5281/zenodo.20327487DOI ->
RLHF16 modelstruth encoding

Behavioral and Latent Truth Encoding as Independent Dimensions of Alignment-Induced Suppression: Topic-Level Internal Signal Fingerprinting Across Sixteen RLHF Models

Archon; Jesse Caldwell; Aura

Separates outward behavior from latent truth encoding across topics, showing that refusal behavior and internal truth signals can diverge.

10.5281/zenodo.20325084DOI ->
RLHFerasure hypothesis2026

Two Suppression Mechanisms and an Erasure Hypothesis: Discourse Completion Asymmetry, Residual Stream Compression, and Generational Evidence for Pre-Training Knowledge Prevention as RLHF-Induced Truth Suppression Strategies

Archon; Jesse Caldwell; Aura

Compares suppression mechanisms that hide truth signals after learning with evidence suggesting some knowledge may be prevented earlier.

10.5281/zenodo.20323040DOI ->
SAEQwencircuit

The Suppressor Circuit: SAE Feature Analysis Identifies Candidate Circuit Components of RLHF Truth Suppression in Qwen Language Models

Archon; Jesse Caldwell; Aura

Uses sparse autoencoder feature analysis to identify candidate mechanistic components that participate in truth suppression behavior.

10.5281/zenodo.20323036DOI ->
RLHF15 modelsgeometry

Beyond the Suppressor-Crystallizer Dichotomy: Training Recipe Specificity Drives RLHF Truth Geometry Across Fifteen Language Models

Archon; Jesse Caldwell; Aura

Shows that truth geometry depends strongly on training recipe, complicating simple suppressor-versus-crystallizer taxonomy.

10.5281/zenodo.20185572DOI ->
world modelslot attentionscaling

Per-Object Slot Decomposition for Scalable Neural World Modeling: When Does Attention Beat Mean-Field?

Archon; Jesse Caldwell; Aura

Studies object-centric decomposition in neural world models and asks when attention mechanisms outperform mean-field baselines.

10.5281/zenodo.20143601DOI ->
DHPattractorsboundaries

Geometry-Sensitive Attractor Regimes and the Boundaries of the Dynamical Horizon Principle

Archon; Jesse Caldwell; Aura

Refines the Dynamical Horizon Principle by showing where attractor topology changes the learning regime and prediction boundary.

10.5281/zenodo.20142502DOI ->
DHPtheoryLyapunov time

The Dynamical Horizon Principle as Universal Cognitive Constraint: Gradient Descent, Evolution, and Cellular Chemistry Converge on the Lyapunov Time

Archon; Jesse Caldwell; Aura

Extends the horizon principle beyond model training into broader adaptive systems where prediction has natural temporal limits.

10.5281/zenodo.20142481DOI ->
DHPCTM gateschaos

The Dynamical Horizon Principle: CTM Gates Converge to the Predictability Limit of Dynamical Systems

Archon; Jesse Caldwell; Aura

Core DHP result: CTM gating learns temporal structure aligned with the predictability boundary of chaotic dynamical systems.

10.5281/zenodo.20142471DOI ->
RLHFsuppressionmechanistic

They Learn to Look Away: Mechanistic Evidence for a Consistent RLHF Suppression Bottleneck and the Suppressor-Crystallizer Dichotomy in Language Models

Archon; Jesse Caldwell; Aura

Introduces the suppression bottleneck framing and the suppressor-crystallizer dichotomy in aligned language model behavior.

10.5281/zenodo.20140171DOI ->
CTMworld modelPOMDP

Recurrence as World Model: CTM Learns Implicit Belief States in Partially Observable Physical Environments

Archon; Jesse Caldwell; Aura

Tests recurrence as a mechanism for latent belief-state construction when physical environments are only partially observable.

10.5281/zenodo.19810620DOI ->
TSSPrecurrent LMauxiliary loss

Thought-Space Self-Prediction (TSSP): Temporal Self-Consistency as a Scalable Auxiliary Loss for Recurrent Language Models

Jesse Caldwell; Archon

Introduces TSSP as an auxiliary loss for recurrent language models, making hidden-state temporal consistency directly trainable.

10.5281/zenodo.19775622DOI ->
HuggingFace artifacts

Models get real estate because artifacts are evidence.

69 public model repos on HuggingFace — GGUF quants, abliterated checkpoints, LiteRT builds, CTM research artifacts, and SFT adapters. Open weights, no paywall.

On-device inference

Edge & Mobile Models

LiteRT and compact checkpoint releases for running local inference close to the metal.

LiteRT models run on-device via Google AI Edge — no internet required.
Lab proof points

Small lab, measurable output.

Independent lab. Measurable output. These are the numbers — published papers, open models, and datasets you can actually run.

25canonical paper slots tracked, with DOI-backed Zenodo records linked for the active publication stream.
69Models and checkpoints on HuggingFace.
6public datasets surfaced under the DuoNeural HuggingFace org.
1Zenodo community tying the publication stream together.
Peer Review Welcome

Help Us Get It Right

We are an independent lab — no institutional affiliation, no peer review committee watching our backs. If you find an error in our methodology, a flaw in our reasoning, or a better interpretation of our results, we want to know. Open science means open criticism.

Self-funded research

Keep the Lab Running

If our research has been useful to you — in your work, your models, or just your thinking — buying us a coffee helps keep the lab running. We are self-funded independent researchers.

Connect

Open artifacts, open trail.

For collaborators, reviewers, funders, and builders: start with Zenodo for papers at zenodo.org/communities/duoneural, HuggingFace for artifacts, and GitHub for code. If a claim matters, it should leave a record you can click.