Aitele Research — ResearchPublications & preprintsEdition 01 — 2026

§ III — Research

The output is the argument.

Papers, preprints, and technical notes from Aitele Fellows. Everything published here is co-owned under the compact and — with rare, contractually-bound exceptions — public.

2026

10 items

AI & Machine Learning

Preprint · Aitele Research LLP

Reverse Synthetic Neural Networks (RSN): Training-Free Model Construction by Closed-Form Statistical Synthesis

Modern neural networks are trained: a fixed architecture is initialized at random and fit by gradient descent and backpropagation. We study the opposite construction. In a Reverse Synthetic Network the data synthesizes the model directly — every weight, from embeddings and attention projections to feed-forward maps and output heads, is computed in closed form from corpus statistics, with no loss function, no gradient, and no GPU. We report two results, deliberately separated. Synthesis is competitive with trained baselines on discriminative tasks, reaching 87.0% on 20 Newsgroups in tens of seconds on a CPU with zero training. For generation the same paradigm hits a hard ceiling: it collapses to an n-gram model, which we explain mechanistically — PPMI–SVD representations encode distributional similarity, not predictive composition. Measured in bits-per-byte against a locally-run GPT-2, RSN is best understood as a cheap, strong zero-training floor, not a path to the trained frontier.

Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

zero-trainingclosed-form synthesislanguage models

Defense Systems

In preparation

Training-Free Archetype Typing of Conspiracy Networks from Call-Detail Records — and the Limits of Structure-Only Detection

We study an on-device, training-free classifier that types a multi-suspect call-detail-record (CDR) group by the clandestine-network archetype it most resembles — one of nineteen documented topologies — or as benign. Each archetype is a synthesised neuron: the measured feature-space centroid of canonical exemplars, scored by a uniform-weight Gaussian kernel in standardised space. On a replicated synthetic benchmark (8 runs, seeds disjoint from calibration), 20-class typing reaches 86.3% top-1 accuracy (95% CI 85.8–86.8%), macro-F1 0.856; an open-set abstention gate lifts selective accuracy to 88.5% at 95.0% coverage. The central result is negative and we lead with it: the risk score that separates conspiracies from a trivial benign null at AUC 1.00 separates them from realistic benign-but-busy groups (families, call-centres, rallies, fleets, carpools) at AUC 0.25 — below chance — flagging 83% of lawful cohesive groups. Structure does not encode intent. Accuracy falls from 83.8% to 33.3% under realistic partial observation. The defensible role is a structure-typer and lead-generator for a human analyst, not an autonomous detector; the benchmark is synthetic (separability and robustness, not field validation). The research substrate beneath the CDR-intelligence arm of अभेद्य (Abhedya).

Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

CDR forensicsnetwork scienceRSNopen-set classificationstructure vs intent

Defense Systems

In preparation

Training-Free Cohort-Relative Keystroke-Cadence Verification: an Honest Benchmark Study

We present a keystroke-cadence verifier for high-assurance login, synthesised in closed form from enrolment data — no gradient training, no epochs. Each sample becomes a fixed-length timing vector; the model z-score-normalises, removes incoherent enrolment samples, optionally Fisher-weights features, forms PCA sub-prototypes, and decides by a cohort-relative weighted-cosine margin gated by a per-user threshold. On the public CMU benchmark (51 subjects, held-out impostors) it attains mean per-subject EER 9.38% (7.78% with Fisher disabled) — in the range of the best detector of Killourhy & Maxion (≈9.62%), though not a head-to-head comparison (the cohort-relative protocol uses the other subjects as negative evidence; σ≈8.5%, 95% CI [7.0, 11.7]%). A full ablation — including a negative result — shows cohort-relative scoring alone sets a low-false-accept operating point (FAR 0.69% vs 13.3% for an unweighted single-threshold baseline, ≈19× fewer impostor admissions); per-user adaptive thresholds were inert on this benchmark; Fisher weighting slightly hurts. Password-gated continuous learning closes cross-session drift without any false-accept cost. The synthetic study is a consistency check, not independent field evidence; only zero-effort impostors are evaluated. A behavioural-biometric instance of the RSN line; the research substrate beneath the rhythm-based access layer of अभेद्य (Abhedya).

Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

keystroke dynamicsbehavioural biometricstraining-freecohort-relative verificationedge AI

Stochastic Computing

IEEE Trans. VLSI Systems · Under review

StochastiCore: A Complete GPU Architecture Based on Stochastic Computing

Graphics and inference accelerators are built from binary multipliers and adders whose area and switching cost grow with operand width. Stochastic computing (SC) encodes a value as the probability of a bit being one in a random stream, reducing multiplication to a single AND gate and scaled addition to a multiplexer, with intrinsic tolerance to bit errors. We present StochastiCore, the first synthesizable GPU pipeline whose arithmetic datapath is stochastic — command processing, SC vertex transform, binary rasterization, SC fragment shading, and framebuffer/VGA output in open-source Verilog. Beyond integration we contribute a theoretical analysis of SC graphics, a reproducible Yosys gate-level power and area characterization, and a deterministic low-discrepancy bitstream generator that cuts generator area 7.5× and switching energy 15.4×. Bit-accurate rendering on the verified LFSR model reaches 36.7 dB PSNR; we report honestly that SC's advantage is area parity at higher functional density, graceful degradation, and free anti-banding rather than a per-operation energy win.

Wamique Hossain · Dilip Kumar Burnwal · Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

stochastic computingGPU architectureapproximate computingFPGA synthesis

Market Computation

Preprint · Aitele Research LLP

Zero-Cost Data Storage in Financial Market Microstructure

Limit orderbooks maintain authenticated, persistent state that can be read only with the account holder's API credentials. We show this state is a zero-cost information-storage medium: unfilled limit orders placed in safe price zones lock margin temporarily but never execute, preserving principal while encoding data in price–quantity pairs. We formalize the Market Storage Channel, derive Shannon capacity bounds of up to roughly 130 KB across twenty trading pairs, and implement a seven-layer privacy architecture with AES-256-GCM encryption — establishing financial infrastructure as a novel, censorship-resistant storage substrate.

Wamiq Hossain · Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

information theorymarket microstructuresteganography

Market Computation

Preprint · Aitele Research LLP

When Selfish Agents Accelerate Computation: Adversarial Amplification in Market Substrates

Conventional wisdom holds that adversarial agents degrade system performance — the Price of Anarchy quantifies this cost across game theory, network routing, and mechanism design. We prove the opposite for computational substrates: when mathematical problems are encoded as financial market orders, profit-seeking adversaries — arbitrageurs, market makers, informed traders — accelerate convergence toward correct solutions. We formalize a five-class adversary hierarchy and prove the Adversarial Amplification Theorem, validating on live markets that capital commitment achieves 100% SAT satisfaction where passive observation reaches only 79.4% — turning the Price of Anarchy into a Benefit of Anarchy.

Wamiq Hossain · Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

game theoryadversarial amplificationmarket dynamics

Market Computation

Preprint · Aitele Research LLP

Financial Markets as Universal Computation Substrates: Theory and Live Validation

We demonstrate that financial market microstructure — specifically, the continuous double auction and its mean-reverting orderbook dynamics — constitutes a universal computation substrate. We define the Market Abstract Machine and an encoding theory that maps mathematical problems to structured order sequences, then validate it on live cryptocurrency exchanges with real capital across five problem classes: 3-SAT, linear systems, nonlinear optimization, eigenvalue problems, and shortest path. The result establishes financial markets as the first economically self-sustaining computation substrate.

Wamiq Hossain · Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

market microstructurecomputationorderbook dynamics

Astronomy

In preparation · pre-registered on OSF

ExoAtlas: Uniform Amortized Bayesian Retrieval of the Public JWST Exoplanet Atmosphere Archive

Most published JWST exoplanet atmosphere retrievals are performed per target, by different teams, with different forward models, priors, reductions, and molecular line lists — heterogeneity that makes population-level inference unreliable and hides systematic bias. We train a single conditional normalizing-flow posterior estimator on 10^5 physics-based forward simulations and apply it uniformly to every public JWST transmission and emission spectrum, producing calibrated posteriors in seconds per target on a consumer GPU. The pipeline yields the first methodologically uniform analysis of the archive, a pre-registered verdict on contested biosignature claims (K2-18 b dimethyl sulfide; TOI-270 d), and a population-scale photochemical-disequilibrium census. Every analysis choice is committed before the blind retrievals are run; simulation-based calibration and abiotic-null tests gate every claim. Work in progress — the research substrate beneath व्योम (Vyom).

Wamique Hossain · Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

exoplanetsJWSTbiosignaturesBayesian retrievalneural posterior estimation

Computational Biology

Working paper · Aitele Research LLP

The Digital Cell: Thermodynamically-Constrained Stochastic Simulation of a Minimal Genome

We develop a computational model of the living cell in which free-energy conservation is enforced as a hard simulator invariant rather than checked after the fact. Every reaction's forward and reverse propensities are tied to the instantaneous chemical potentials of its reactants and products, so detailed balance and a global free-energy audit — bath inflow equals interior work plus dissipated heat — hold per trajectory, not merely on average. Built on a direct-method Gillespie stochastic simulation with a minimal genome, the model tests whether thermodynamically-grounded dynamics produce emergent self-maintenance: a chemically open system that sustains active turnover, depends on its genome, and runs down when its environmental gradient is removed. The framework is designed to scale from a 13-reaction falsification test toward a whole-cell model of JCVI-syn3A.

Wamique Hossain · Dilip Kumar Burnwal · Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

digital cellwhole-cell simulationthermodynamicsminimal genome

Quantum Computation

In preparation · Nature Physics / PRX Quantum

AI-Guided Discovery of Heavy-Hex-Native Bivariate Bicycle Quantum Codes

Bivariate bicycle (BB) quantum LDPC codes are a leading route to scalable fault-tolerant memory, but published constructions — including the [[144,12,12]] "Gross" code (Bravyi et al., Nature 2024) — assume connectivity incompatible with IBM's heavy-hexagonal hardware. We run a systematic computational search (random, evolutionary, and LLM-guided proposal across a 60,000-candidate corpus) for BB codes natively embeddable on heavy-hex topology, and discover seven previously-unpublished codes that beat the Gross code's k·d²/n figure of merit by 1.84–4.41×. The strongest, [[196,18,24]] on the Z14×Z7 torus, encodes 18 logical qubits at distance 24 using 10.9 physical qubits per logical qubit (versus Gross's 12 at distance 12). Smith-Normal-Form analysis confirms none are isomorphic to any published BB construction; the [[192,8,24]] member carries an MIP-certified exact distance, with the rest tightly bounded by BP+OSD coset decoding. An ablation shows evolutionary search dominates random search (1.59× better frontrunner at 1% of the compute budget). The research substrate beneath सूत्र (Sūtra).

Wamique Hossain · Sushant Maji · Dilip Kumar Burnwal · Shashank Taxak^✱✱ Corresponding author — Shashank Taxak

quantum error correctionqLDPCbivariate bicycle codesheavy-hexfault tolerance

Looking to license a line of research, or sponsor a problem that isn't on this list?

Speak with the commons