Starting with a circuit
The first question is what we can faithfully model—and what that model would allow us to claim.
Felisyn starts with a question: can an identifiable neural computation become the working mechanism of an autonomous controller?
The intended application is a cat-themed project that may eventually control a pump.fun token. That is an engineering goal. Before connecting anything to it, we need a biological model whose origin, assumptions and behaviour can be inspected.
At this point, we have a research plan and candidate sources. We have not run or reproduced a neural model locally. This journal begins before the first simulation, so it can record the decisions and failed attempts as well as any results.
A map is a starting point
A diagram of brain regions does not specify an executable brain. Even a detailed connectivity graph would leave questions about neuron dynamics, synaptic strengths, delays, initial conditions and sensory inputs.
For our implementation, each of those choices needs an explicit place in the record. Which quantities were measured? Which were fitted? Which connections were generated from a rule? Which choices were introduced by us to make the system operate?
Combining measurements from different preparations can support a model. It does not turn those measurements into a reconstruction of one particular animal.
The scope we can test
We are starting with one circuit mechanism. A narrow target lets us name a reference result, define an input protocol and compare the output with evidence. The initial candidate concerns transmission between the visual thalamus and visual cortex; the source-selection note explains why it is provisional.
A successful reproduction would support a bounded statement about that implementation and comparison. It would not establish a complete cat brain, feline awareness or an understanding of financial markets.
This distinction will shape the product. External events would be converted into model stimuli by an encoder. Neural activity would be converted into action requests by a decoder. Those mappings would be documented engineering choices, not discovered cat preferences.
What would count as progress?
The next useful artifact is a package audit: the files we can obtain, their versions, reuse terms, dependencies and the result they allow us to reproduce. After that, the first simulation needs an exact command and a comparison against a declared reference.
Later, an intervention should establish whether the neural computation actually affects controller requests. Holding the inputs fixed while removing, shuffling or bypassing a model component is one possible way to test that relationship. The exact intervention will depend on the model we select.
Biological agreement, a causal role in the software, reliable operation and financial performance are different questions. We will report them separately.
Next step: inspect the original model package before selecting a benchmark or writing a replacement implementation.