# Pilot v1 — fixed before the first scored run

This is a feasibility and software pilot, not a claim that a fly can program. The first question is whether a completely frozen, full traced-neuron connectome plus a learned output layer can generate executable solutions to held-out elementary coding tasks. A negative result constrains this implementation and budget, not every possible use of a connectome.

1. Import MaleCNS v1.0 locally. Select `status == Traced`, keep isolated nodes, keep every positive count between selected endpoints, including count-one edges and self-edges. Do not rewire, prune by strength, sample neurons, or optimize edges. Save exact integer body IDs, edge direction, raw counts and SHA-256 manifests. The published approximately 166,700-neuron count is not substituted for the cached annotation count.
2. Use a simple rate network, with one real number per neuron. For each input byte: `h_next = 0.5*h + 0.5*tanh(g*W*h + fixed_input(byte))`, with `W[target,source]` equal to the recorded count and `g = 0.9 / max_target(sum_source(W))`. This global conversion preserves all count ratios and yields a recurrent infinity-norm bound of 0.9. The corresponding leaky update is contractive with bound 0.95. This helps stability but may erase long-range memory. It is an explicit initial hypothesis, not an optimized setting.
3. Counts are not measured electrical weights. This initial unsigned rate model does not reproduce transmitter-dependent inhibition, spike timing, neuromodulation, sensory encoding or an animal's learned experience. Input byte injection is artificial, fixed and seeded. A fixed signed pooling map reduces all neurons to 1,024 output features; no prompt bypass goes directly into the readout.
4. Fit only a 1,024-to-257 linear softmax head, with train-only feature normalization. Predict bytes plus EOS, using cross entropy on solution bytes only. Teacher forcing uses reference code only during training and validation. Greedy test generation consumes its own previous bytes and cannot access reference solutions or test outputs. Retain raw generated code, never repair it after tests.
5. Train a no-recurrence control with the same input, leak, pooling and readout. It removes only the `W*h` term in a separate model, never edits the connectome files. This tests whether the recurrent graph contributes beyond the input encoder and readout. It does not establish fly-specific topology advantages; a topology control would require a separately authorized future protocol.
6. Train a two-layer, width-96 causal byte Transformer from scratch on the same 12 examples and same answer-byte exposures per update. It is a tiny architecture control, not a pretrained LLM. Models use 200 updates for this pilot, seed 7, AdamW, clipped gradients, CPU and two threads. Readout lr 0.003 and Transformer lr 0.001 are declared defaults, not tuning outcomes. Parameter counts, elapsed time and examples/exposures are reported, rather than claiming equal compute or equal capacity.
7. Select the checkpoint with lowest answer-byte validation cross entropy, including initialization and every 25 updates plus the final update. No checkpoint is selected from test scores. Validation has two elementary tasks. Test contains six different algorithm families: pair search, palindrome, prefix sum, stack parsing, stable deduplication, maximum subarray. These are original LeetCode-style formulations, not licensed LeetCode benchmark data. Test statements and expected outputs are not training inputs. Test files are visible to the experimenter, so this is not a sealed confirmatory test.
8. One greedy candidate per task, 384 new-byte limit, no tools, repairs or test feedback. Measure task pass@1 (all provided tests must pass), syntax/runtime/unsupported outcomes, latency, generated length, and descriptive Wilson 95% task intervals. The six tasks are hand selected and do not estimate performance across all LeetCode. Judge failures are unmeasured, not wrong answers or zeros.
9. Optional pretrained LLM comparison uses exactly the same problem text and function contract. A pinned model identifier, complete request/response, returned model, usage, latency, output cap and judge are retained. Its full output and the common 384-byte budget are scored separately; provider tokens are not equivalent to bytes. Unknown pretraining contamination and vastly different training budgets make this a practical reference, not a controlled architecture comparison.
10. No reinforcement learning in this pilot. A code-test reward is meaningful only with reliable execution, adequate training tasks and a model able to emit useful programs. Later policy optimization can update the readout while freezing the graph, but this has not been implemented or shown to work here.

## Success and next experiment

A passing held-out task is evidence of that program's functional correctness under the provided tests, not proof of a biological brain acquiring programming. Lower training loss only establishes learnability of the readout on these examples. A larger study should preregister a fresh benchmark, use a licensed code corpus, compare multiple seeds, vary memory/gain using validation only, add pretrained open models and a containerized full-Python judge, and report all failures. Only after basic syntax and curriculum learning succeed should the study scale to broad algorithmic reasoning or execution-reward training.
