NeuroGenesys.
Publications

What we publish, we measured.

Preprints, design rules, and negative results from our research programs — pre-registered where possible, and published including the experiments that failed.

Nov 12, 2026Continual LearningSynthesis and results

Route on shared statistics, never expert confidence: the router architecture for frozen experts

A literature synthesis and experiment plan on the router that picks which frozen expert, model, skill or generation handles each input. Theory, MoE gating, LLM routers, continual-learning task inference and the lab's own results agree that a router should not read the experts' own confidence; the best-supported design scores experts in one shared feature space with a closed-form rule and abstains on its own margin. Pre-registered tests on a cached Split-CIFAR-100 library of 30 expert sets then confirmed five of six predictions. A shared space beat self-report, slightly without pretraining and by 45 points of selection error with ImageNet ViT features, and it held selection error nearly flat from 5 to 20 experts. Margin-only abstention did not transfer, and a direct readout of the shared features beat routing to the experts by 13 points.

Nov 11, 2026Continual LearningSynthesis

Merge confusable experts, split only separable ones

A literature synthesis and experiment plan on when to merge, split or spawn experts in continual learning with frozen or isolated experts. Under hard selection both operations change selection error first and capacity second. Re-read as a merge-and-split experiment, the lab's Split-CIFAR-100 task-count sweep shows going from 10 to 5 experts raised system accuracy from 0.342 to 0.422, and going to 20 cut it to 0.208. Merge pairs the router confuses, keep pairs it separates well, and split only at a plateau when the router can separate the children. Nine experiments (MS0–MS8) are pre-registered. The first, MS0 (10 seeds), confirms that a single merge helps behind max-softmax routing and hurts behind a strong shared-space router, that router confusion predicts which merge pays, and that keeping the parents beside the merged expert beat replacing them.

Nov 11, 2026EvaluationSynthesis

Stable wrong beliefs evade every self-check: a literature synthesis on model self-knowledge and confident errors

A literature synthesis on whether language models know what they don't know. Self-derived signals (sampling disagreement, hidden-state probes, trained verbalized confidence) detect instability well but fall to about 0.55–0.67 AUROC on errors a model repeats. Catching those needs outside information. Includes six queued experiments; many sources are unreviewed 2025–26 preprints.

Nov 11, 2026Continual LearningSynthesis

Successors should earn authority, not win contests: graded succession in continual learning

A literature synthesis and design proposal that reframes generational continual learning: a new generation of models or knowledge should not compete with the old one but take over from it by demonstrated competence, gradually and per context, until its authority is dominant, while a protected constitution changes only by explicit amendment. It assembles anytime-valid statistics, fixed-share mixtures, belief revision, truth discovery and staged deployment into one design, and proposes five pre-registrable tests. None has been run yet.

Nov 10, 2026Continual LearningResearch note

Trust the premises, not the confidence: which belief to drop when an LLM contradicts itself, and why the solver is not the bottleneck

A pre-registered test of MaxSAT belief repair on llama3.2 3B and llama3.1 8B answers to MQuAKE 2-hop chains. Ordering beliefs by graph entrenchment beat evidence priority, confidence and chance; confidence barely beat random. Most of the gain came from making each belief's phrasings agree, LLM-judged constraints halved it, and the exact solver cost 0.1 ms against 23–38 s of LLM work per chain.

Nov 10, 2026Continual LearningResearch note

Iterated self-closure settles fast, but only evidence that overrides self-belief makes the settled state true

A pre-registered toy test of "chewing": a model answers its own atomic questions, applies the knowledge base's composition rule and retrains on the derived consequences. The loop always converged, but unanchored it left 44% of updated persons coherently wrong. Anchoring to the told facts fixed that, and overwriting self-beliefs first carried updates to entities no gold consequence covered.

Nov 10, 2026Continual LearningResearch note

A gradient kernel flags contradictions but does not predict forgetting, and steering rehearsal by it backfires

A pre-registered toy test of whether the gradient kernel between a new fact and stored facts predicts which beliefs an update will break or fail to change. Low confidence beat every kernel at predicting forgetting in the update stream (AUROC 0.737 vs 0.663). The kernel predicted consequence propagation and its sign detected contradictions, but kernel-chosen rehearsal cost retention.

Nov 10, 2026Continual LearningResearch note

EWC stiffness freezes stale facts, and no per-fact Fisher ledger can release it

A pre-registered toy test of elastic weight consolidation on contradicting fact updates, with an exact Fisher. Stiffness raised unrelated retention from 44.3% to 92.0% but cut edit success from 62.5% to 13.5% and left 81.7% of paraphrases stale. A KL anchor at equal retention learned far more, and subtracting a superseded fact's own Fisher did nothing measurable.

Nov 10, 2026Continual LearningSynthesis

Coherence is selected, anchored, then pruned: a literature synthesis on logical coherence in LLM world models

A literature synthesis on whether LLMs hold a logically coherent model of reality. They hold many local, blended models instead, with accuracy near ceiling and cross-question consistency near 60%. Methods that propagate new facts generate, filter and retrain; iterating that loop converges to whatever the selector rewards. Many sources are unreviewed 2025–26 preprints.

Nov 10, 2026Continual LearningSynthesis

EWC protects what its Fisher sees: a literature synthesis on elastic weight consolidation, scale and coherence

A literature synthesis reading EWC as a diagonal approximation to a KL anchor on old inputs. At 1–8B it protects what its Fisher data exercises and little else, and published rankings depend on the estimator. It proposes a per-fact Fisher ledger, since tested on a toy and found null. Many sources are unreviewed 2025–26 preprints.

Nov 10, 2026EvaluationResearch note

Real judges did not break Best-of-N at N ≤ 32, and measured ε did not predict the budget

A pre-registered test of verifier exploitability on 9,600 Llama 3.1 8B candidates over GSM8K, MATH-500 and HumanEval+, scored by a reward model, an LLM judge and self log-probability. No judge made accuracy fall by N = 32, and measured ε did not order the best budget. A plurality vote over the top-scored half beat argmax in every math cell.

Nov 9, 2026Continual LearningResearch note

Giving experts a way to say "not mine": ten pre-registered selector and abstention experiments

Frozen per-task experts on Split-MNIST vote confidently for tasks they never saw. Ten experiments, frozen before running and run once over 5 seeds, asked what stops it. A selector that never reads expert confidence (RanPAC over streaming-PCA features) reached 0.926 class-IL accuracy at ε 0.067. A wrong-answer penalty failed inside the expert and worked at the selector.

Jul 29, 2026Updated Nov 9, 2026MethodologySynthesis

What survived: a claims register with falsifiers for a pre-registered AI research program

Every surviving claim from a pre-registered, run-once program, with its evidence and its falsifier. Alive: pool-relative verifier exploitability, test-time recurrent depth, the AIMC window-policy rule, narrow optical interconnect. Dead: wave/complex substrates, optical compute for inference, frontier-datacenter AIMC, and the ε 'universality class'.

Jul 27, 2026Updated Nov 9, 2026Test-Time ComputeSynthesis

Recurrent depth as test-time compute: parameter-efficient, extrapolating, and bounded by contractivity

A weight-tied recurrent-depth network matches a feedforward net at 1/8 the parameters and wins out of distribution (+0.36 ± 0.13, mean ± sd over 10 seeds; the original single-seed run gave +0.28), extrapolating about 8× in problem size. A non-contractive residual update causes divergence and overthinking; a gated contractive update fixes both.

Jul 26, 2026Updated Nov 9, 2026Continual LearningSynthesis

Escaping the saturated game: plasticity, harder benchmarks, and retrieval in continual learning

Thirty pre-registered experiments. Plasticity loss behaved as an Adam/weight-norm pathology that plain weight decay fixed better than specialized methods at this scale; representation-hard benchmarks left class-IL collapse dominated by cross-task selection; retrieval over external memory worked weight-free. The selection decomposition is Kim et al. 2022's theorem.

Jul 19, 2026Updated Nov 9, 2026MathematicsResearch note

Spectral response of positive recurrences under memory-filtered balanced coefficient perturbations

When an AR(1) memory process perturbs a positive two-term recurrence while conserving total coefficient weight, growth changes only at second order, and the response factors into a positive memory filter times a recurrence-geometry term: memory sets the strength of the response but cannot change its sign. Numerically confirmed; the proof is a skeleton, and the general result is a worked example of a known 1986 expansion.

Jun 27, 2026Updated Nov 9, 2026Continual LearningPreprint draft

Catastrophic forgetting in class-incremental learning is a verifier-exploitability problem

Freeze one near-perfect expert per task and all of class-incremental difficulty becomes selection — which expert to trust. Accuracy factorizes exactly as selection accuracy × within-task accuracy, and the selection term is pinned by the selector's exploitability ε. A November 2026 follow-up shows that selecting by a statistical model of shared features instead of expert confidence lifts Split-MNIST to 0.926.

Sep 9, 2026EvaluationResearch note

A critic that shares the generator's prompt shares its errors: measured self-critique correlation

Self-critique failure is mostly classical correlated-jury theory, and the note says so first. Measured: a same-context critic missed 12/12 planted errors that fresh critics all caught, and a panel's (e, P_both) predicted its majority accuracy (0.841 vs 0.842). The lever was prompt scaffolding, and only for context-induced errors.

Sep 8, 2026MethodologyMethods note

The novelty gate: five checks that caught seven fluent ideas before they shipped

A procedure calibrated on seven attractive ideas that turned out not to be discoveries: does the claim forbid anything; is the object older than the literature; is a cross-domain R² ≈ 0.99 really an identity; does it survive changing microscopic detail; and a hostile prior-art sweep. A meta-check holds deflations to the same bar.

Sep 8, 2026PhotonicsFalsification study

The wavefield gate: Fourier modes are the wrong basis for visual computing

A proposed photonic 'wavefield' graphics architecture, tested on its two measurable claims. Fourier modes need 21–49% of coefficients to reach 35 dB, not 0.5%, and a one-pixel move changes 97% of them. Display space–bandwidth and conversion costs close the rest: the physics is sound, the advantages are not.

Aug 26, 2026Silicon & Analog ComputeNegative results

The log-arithmetic gates: weight sharing buys compute, not storage

Eight prior-art reviews and six pre-registered experiments on log-domain neurons and the hardware proposed for them. None of the 16 ideas was new; LNS loses at inference precisions and analog log-domain error compounds with fan-in. The output is six design rules — chiefly, weight sharing pays only when the freed compute is spent.

Aug 4, 2026PhotonicsFalsification study

Diamond is a material, not a system: no case for quantum memory in AI datacenters

An AI-written brief cast diamond as quantum memory between AI accelerators. The claim fails structurally: collectives copy data, superdense coding gives no net gain, and entanglement rates fall ~12.7 orders of magnitude short. Diamond's thermal conductivity holds up; its classical-electronics claims are overstated. Verification is partial and labeled as such.

Jul 30, 2026Silicon & Analog ComputeSynthesis

Can analog or photonic hardware replace LLM inference silicon? AIMC for compute, optics for decode interconnect

About 54 pre-registered experiments found no general replacement. Analog MoE routing adds no accuracy, and the FFN it accelerates is under 19% of decode energy. Optical interconnect's energy win collapses under honest laser power, leaving bandwidth density at extreme scale, mostly in decode.

Jul 30, 2026MethodologyMethods note

Could the instrument have shown it? Power and saturation checks for pre-registered predictions

Seven pre-registered misses in one day shared a cause: the registered statistic could not have shown the effect — floors, noise, saturation, collinearity. The note adds a five-line power step to pre-registration, and argues that independent parallel attacks act as decorrelated verifiers.

Jul 30, 2026LanguageNegative results

What language buys: re-sampling, transmission, and the limits of a discriminability metric

Eight run-once experiments split language into measurable layers. Extra reasoning steps improve discrimination only when the input can be re-sampled; the metric is provably blind to compositionality; efficiency fixes a lexicon's shape but not its identity. Several hypotheses were withdrawn and two experiments voided as design failures.

Jul 30, 2026Test-Time ComputeResearch note

When the exploit is constructed: ranking collapses, counting survives

At a matched exploit rate, Best-of-8192 accuracy is 0.862 with sampled exploits and 0.000 with exploits constructed against the verifier — while plurality voting among high scorers holds at 0.999. The consensus threshold missed a dominant-wrong-answer regime and was corrected; the two-regime framing is prior art (Dwork et al. 2015).

Jul 30, 2026Continual LearningResearch note

The capacity audit: most class-incremental routing error is placement slack, not capacity

Is the router-error plateau in class-incremental learning a sphere-packing capacity limit? Two floor estimators were killed by their own pre-set validity checks. A validated simulator shows packing stops binding near 16 dimensions; at deployment dimensions roughly 80% of the error is placement slack.

Jul 30, 2026Silicon & Analog ComputeFalsification study

The last hardware flank: analog prefill survives only as an energy-bound niche

Five pre-registered kill tests on analog in-memory prefill with optically disaggregated decode. Shipping the KV cache costs under 0.1% of prefill energy, demoting the optical claim; the cache survives 5 bits but collapses at 4. Against measured chips, digital wins throughput per area, leaving an energy-bound, short-prompt niche.

Jul 30, 2026PhotonicsSynthesis

The wrong phase: optical interconnect's case is in decode, not prefill

A source-checked review found the 'networking is 50% of MoE prefill' figure blends training, decode, and an unsupported claim; measured prefill communication is 9–23%. Decode dominates communication, so that is where optical interconnect's case lives. Digital's move to native 4-bit also erodes the precision edge analog designs relied on.

Jul 29, 2026Test-Time ComputeResearch note

Exploitability is a property of a verifier and its candidate pool, not the verifier alone

The same verifier's exploitability ranged 6.2× (0.153–0.949) depending only on its prefilter, so ε belongs to the verifier–pool pair. Per-pool regime detection fails, but averaging over a window of ~50 pools succeeds. Against adaptive generators, decorrelated panels help only partly; an out-of-span channel did about 8× better.

Jul 27, 2026Test-Time ComputePreprint draft

Thurstone bounds Best-of-N: a 1927 identity, calibrated for modern verifiers

The noise floor of Best-of-N selection is Thurstone's 1927 Case V identity, not a new law. Contributed: blind-spot rates ε ≈ 0.002–0.05 from inverting published overoptimization curves, an exact consensus threshold p/(1−p), and a budget rule N*(ε) with a working tool. Heavy-tailed score noise breaks the identity, so estimate ε directly.

Jul 27, 2026Silicon & Analog ComputeDefensive publication

Which sequence architecture keeps analog in-memory compute's advantage? A cost-model answer with a design rule

Analog in-memory compute accelerates static-weight matmuls but cannot touch attention's growing KV cache, so its decode advantage dilutes with context length. The attention window policy — sliding vs. full — not the attention fraction, decides whether a hybrid keeps the advantage. Distilled from ~70 pre-registered falsification experiments.

Jul 26, 2026Mathematics & LogicMachine-checked results

The displacement theorem, and the aspect no-go

Two small results, stated precisely and machine-checked (22/22 pre-registered predictions across the project): a pigeonhole-style displacement theorem — the metalanguage, not the primitive, is what cannot be eliminated — and a two-line model-theoretic no-go for aspect. Both are plainly labeled where they are trivial.

Jul 26, 2026FoundationsCorrection

Correction: the ε selection law is Thurstone's 1927 comparative judgment, not a discovery

Our ε law, ε = Φ(−gap/(σ√2)), is Thurstone's Case V identity; its R² = 0.997 measured Monte-Carlo error. It is not a universality class (723× off under heavy tails), and the forgetting row is Kim et al. 2022. The consensus threshold p/(1−p), the measured constants, and the tools stand.

Jul 19, 2026MathematicsExpository note

A closed-form spectral response for balanced perturbations of positive recurrences

A tidy, fully explicit special case of the weak-noise Lyapunov-exponent expansion: perturb a positive two-term recurrence so the coefficient sum is conserved, and the growth response has a closed form — with a pleasant golden-ratio corollary. Honestly labeled a worked example, not a new theorem.

Jun 29, 2026Silicon & Analog ComputeResearch note

Analog chips can do nonlinear ML — if you co-design around the device's native nonlinearity

On MNIST, a frozen MLP tolerates 4-bit weights with moderate device noise; 2-bit is the cliff. Saturating device nonlinearities match ReLU only after hardware-aware retraining (0.947 vs 0.938) — naive swaps fail. Class-IL routing stays robust at 4-bit with heavy noise. A first-order model, not a device simulation.

Jun 27, 2026Test-Time ComputeSuperseded draft

When more sampling hurts: verifier exploitability, not accuracy, bounds test-time compute

Best-of-N sampling lifts coverage log-linearly, but realized accuracy is ceilinged by the verifier — and the property that matters is not discrimination but exploitability ε: the rate at which wrong candidates score high. Any ε > 0 makes accuracy non-monotone in N, peaking at a computable budget before declining.

Jun 27, 2026PhotonicsCost-model study

Where optics-native inference actually pays: a cost-model study

Speed-of-light inference is real but narrow and conditional: an optics-native state-space model beats a GPU on batch-1 decode latency only inside a tight box — WDM microrings, small-to-mid width, ~5-bit precision, error-tolerant workloads. The model reproduces a published 2025 chip's throughput exactly; the defensible contribution is the cost model itself.

Jun 27, 2026PhotonicsFalsification study

Reverse-engineering a trained network for all-optical inference via random-feature kernels

Can inference on an already-trained network run entirely in light? An envelope exists, bounded by a quantified depth→width feature wall — and an honest requote against a digital ASIC collapses the headline speedup to near parity. Two genuine advantages survive: precision robustness and zero-forgetting frozen features.

Jun 27, 2026FoundationsNegative results

Complex amplitudes are decorative: a pre-registered falsification of the wave/quantum thesis for learning and cognition

Across continual learning, cognition, and compositional binding, the complex/wave degree of freedom is load-bearing nowhere at equal real-parameter budget. The geometric intuition survives in real-valued form; the transferable contribution is the equal-budget falsification methodology itself.

View Full ArchiveExternal Research LibraryFindings LedgerResearch ToolsSubmit Research

Leadership

A word from the founder — on why the brain is the blueprint, and why the next decade of intelligence will be built here.

"The brain is the universe's only proof that general intelligence is possible. We treat it as the blueprint."

From cortical interfaces to photonic silicon, every program at NeuroGenesys exists to answer one question: how does intelligence arise — and how do we build it safely.

Will Schulz

Will Schulz

Founder, NeuroGenesys • AW3 Technology