AdaptiveMind · Studio · Singh
AdaptiveMind/Research
Papers·Technical notes
MMXXVI

Research.

Papers and technical notes from AdaptiveMind. Focus: deterministic dynamics, topology, and auditable computation.

The studio publishes when there is a result that is specific, falsifiable, and repeatable on a second machine — never as a marketing artefact. Two threads run in parallel: a foundational paper on compositional attractor networks (Semantic Gravity, R/01) and a four-column reference that bridges the engineering vocabulary with the older ontologies it inherits from (Rosetta Stone, R/02). Active work extends the same machinery into a decision router for production incidents.

Every paper ships with a methodology section, a Kill-List of alternative hypotheses we tested and ruled out, a Limitations section that names the regimes where the claim does not hold, and a stable identifier for citation. The standards below are the contract.

[ Index ]

Each page includes a clean narrative version of the work, primary results with numbers, scope and limitations, and a diagram you can cite.

R/01 · 2026-03-22

AttractorsTopologyDeterministic dynamicsRoutingZenodo DOI

Semantic Gravity: Topology-Governed Attractor Dynamics

A paper on compositional attractor networks where a directed transition graph predicts basin escape, separating feasibility (topology) from cost (energy). Tests reach 99.75% escape correspondence (399 of 400 cases) on a held-out 8-atom, 60-pattern reference network at D = 10,000, with noise robustness up to σ = 0.20 and an LOO-CV R² of 0.676 (n = 262). The Zenodo DOI is 10.5281/zenodo.19097695.

Read the paper page →

R/02 · 2026-05-02

Compositional HopfieldSāṅkhyaLegitimation Code TheoryReference

Rosetta Stone: Four Readings of One Dynamical Object

A side-by-side reference table that aligns six concepts across four ontologies: the Bhāgavad Gītā and classical Sāṅkhya, modern dynamical-systems mathematics, Maton's Legitimation Code Theory (semantic gravity / semantic density / semantic waves), and the engineering primitives that ship in the io-gita reasoning engine. Lets a reader move between vocabularies without losing the underlying object.

Read the paper page →

[ Standards ] · The contract every paper here keeps

What does it mean to publish responsibly from a small studio?

A short answer: every numeric claim states its sample size; the data used to tune the model and the data used to test it come from different sources; the algorithm runs end-to-end on the second machine within an hour; and the alternative hypotheses we ruled out are written down in the paper, not a footnote. These four rules are the difference between a result and a slogan.

  1. S/01 · Open by default

    Every paper ships with the methodology written in plain prose, the equations rendered without paywall, and a stable identifier (DOI or ORCID-attached preprint) so the work can be cited and re-found.

  2. S/02 · Honest scope

    Every numeric claim states sample size, the data source used to tune vs the data source used to test, and a Limitations section that calls out where the claim does not hold. Synthetic results are labelled SYNTHETIC; in-distribution results are labelled accordingly.

  3. S/03 · Reproducible by hand

    Every result on this page is paired with a code repository (Python, with a documented entry-point script and a small reference fixture) so a second engineer can replay the run on their machine in under an hour.

  4. S/04 · Falsifiable claims

    Each paper carries a Kill-List — the alternative hypotheses we tested and ruled out. The list is part of the paper, not a footnote, because a result is only as strong as the hypothesis it survives.

[ Active ] · What is in motion right now

What is the next paper going to be about?

Two threads are running. The first is a bridge-synthesis layer that connects the SG-Engine to live incident traces from production systems, so a planner can refuse a multi-hop chain when the topology says no path exists. The second is a scaling study that tests whether the escape-prediction accuracy of the attractor stack holds as dimension and pattern count grow.

  1. A/01 · In progress

    Bridge synthesis on warehouse decision data

    Extending the io-gita SG-Engine into a decision router for production incidents. Every state transition is mapped into a directed graph at α-density 36.6%; multi-hop plans are routed through reachable basins; paths that don't exist are refused before they are followed. Test bed: incident-replay traces from WRIE and URIP. Next: validate-before-apply on a real warehouse decision dataset.

  2. A/02 · In progress

    Compositional capacity at scale

    Testing whether the Hopfield-based attractor stack holds its escape-prediction accuracy as dimension grows from D = 10,000 to D = 50,000 and as pattern count crosses one thousand. The aim is a published curve of accuracy vs storage load, with the regime where graph-prediction degrades isolated and named.