Welcome to the Library

What this is

A working archive for VINE — a geometric AI platform built solo at NEXICOG Ltd. VINE replaces if/elif branching code with geometric settling on continuous decision surfaces: decisions become positions rather than cases, and outputs are read off the position the system settled into. The same primitive runs across ERP integration, veterinary decision support, clinical triage, a live data cruncher, a procedural world simulator, language and cognition experiments, and a transformer-internals research thread.

Two streams run in parallel here:

• Lab Portfolio — formal evidence, organised into reading-path buckets. Results, behaviour, case studies. The current shelves:
  • Orientation — what the library is and how to read it.
  • Mechanism — reference descriptions of the engine's observable behaviour, and the deterministic English-rendering layer that sits on top of it.
  • Computing Lineage — a twelve-floor survey of historical decisions that produced the current architecture of machine learning.
  • Transformer Decoupling — experiments measuring a structural signal in trained transformers that the softmax output mechanism does not read, and the failure modes that follow from its suppression.
  • Biological Architectures — nervous-system designs in biology and their analogues in the engine, from tardigrade and C. elegans up to squirrel navigation and dragonfly interception.
  • Behavioural Observations — standalone findings about trained- model behaviour, including Alice & Bob training-environment echoes and a migration experiment showing pre-trained embeddings are navigable geometric manifolds.
• Dev Log — a public lab notebook. What was built today, and why. Dated entries, informal, often scruffy.

Both streams are equal and both are public. One shows you the outcomes; the other shows you the rhythm of getting there.

How to read it

• Start with the portfolio if you want evidence that it works.
• Start with the dev log if you want to watch it being built.
• Tag chips in the corner of each entry lead to sibling work on the same theme.

A note on tone

VINE is built solo, on a budget, in a cottage. The dev log reflects that; the portfolio is a touch more measured.

The author does not come from a software-engineering background. The working register here is closer to HACCP and general health-and-safety practice: rule-based, auditable, the kind of thinking where every decision has to be traceable and every exception has to be written down in advance. The technical vocabulary in this library — basin, settling, attractor, concern-resolution — maps onto operations that could in principle be reviewed by someone who has never written a neural network, and that is the point.

The day-to-day research approach is closer to clicker training than to software engineering. Observe an animal for long enough and the rules it is already following begin to surface. The same approach applied to plants, weather, economic systems, and neural networks surfaces rules the same way. Translating what is observed into the smallest, most boring code that reproduces the behaviour is the core craft. What arrives from nature is already well-designed; the job is not to invent, but to transcribe.

Two goals drive the work:

• Auditability and reliability, deployed. Any decision the system makes should be inspectable at the moment it is made. Failures should be locatable to a specific rule or basin rather than diffused across a billion parameters. This matters because VINE is being deployed into practical domains — veterinary welfare, clinical triage, ERP economics — where the cost of an opaque wrong answer is measured in real consequences.
• Integration, not replacement. Most of humanity's investment in AI is sunk into pre-trained models. Throwing those away and starting over is expensive, wasteful, and increasingly damaging. A core aim of this work is to find methods that integrate into those existing models rather than replace them — to make the neural networks that already exist more coherent, more auditable, and more maintainable, with the long-term hope of breaking the cycle of retraining new ones and keeping the ones that already exist useful for the long run.