# Does measured wiring help language prediction? Declared on 10 September 2026 before fitting these language controls. This is an exploratory extension of the registered WikiText-2 comparison. Its main test results and acute mechanism interventions have already been inspected; this is not a new untouched-test discovery study. The larger BabyLM queue is paused at saved checkpoints to prioritize this missing comparison. ## Question and comparison Does FLM's measured 1,024-neuron subset predict held-out language better than independently rewired graphs with the same recurrent model and matched local constraints? Separately, does the slow state improve prediction after allowing the model to retrain without it? | Condition | Graph seeds | Training seeds | Runs | |---|---|---|---:| | Measured wiring, fast and slow state | Original graph | 42, 43 | 2 existing references | | Rewired, fast and slow state | 101, 103, 107 | 42, 43 for **each** graph | 6 new | | Measured wiring, slow state disabled throughout training | Original graph | 42, 43 | 2 new | The measured references are the previously completed 6,000-update FLM runs, bound to their frozen checkpoint hashes in `reports/wikitext2/selection.json`. Their learned weights are never used to initialize controls. New models start from scratch. For each training seed, every condition starts with exactly the same named parameter tensors and sees the same sampled text. Graph buffers differ only where the protocol requires. Verify the current trainer against the reference source on a short deterministic optimizer replay before fitting. Reuse of the two references does not create additional independent runs. ## Null graphs Use `data/graphs/central-1024/graph.npz`, SHA-256 `a2f35369a5ed4981b8ef75c92df285f7df576a020582f119a1372bfbedbaebee`: 1,024 retained MaleCNS neuron identities and 76,130 directed weighted edges. This is a selected subset, not a whole fly brain. Use the signed double-edge-swap algorithm in `flm/wiring_controls.py`. In postsynaptic-row/presynaptic-column notation, propose `(i,j), (k,l) -> (i,l), (k,j)` only when sources `j,l` have the same sign. Reject duplicate edges, new self edges, unchanged proposals and invalid endpoints. Freeze every original self edge at its original weight. Stop at exactly ten accepted swaps per original edge, failing if this needs more than 100 proposals per edge. Independently initialize all three chains from the measured graph with NumPy generator seeds 101, 103 and 107; do not select graphs based on model performance. Validate exact per-node directed in/out degrees, each neuron's source sign, each target's signed incoming weight/contact multiset, node order, positions, pool membership and original self edges. Weights remain attached to their original postsynaptic slots. New contacts are artificial transferred magnitudes, not measurements of those newly connected neuron pairs. Outgoing weighted strength, distance, reciprocity, connected components and higher motifs need not be preserved. Report edge overlap, reciprocity, strong components and acceptance counts for every null. Ten swaps per edge is a finite chain budget; uniform sampling and mixing are not established. The contrast concerns higher wiring organization conditional on these preserved features. It does not test whether all anatomical information is useful versus none, because node identity, signs, degrees, input magnitudes and pooling remain anatomically derived in the controls. ## Matched learning and slow-state control Keep `docs/WIKITEXT-PROTOCOL.md` unchanged: official raw article partitions; the same training-only 4,096-ID byte-BPE tokenizer; 600 training articles; batch 16, sequence 96, first 16 positions excluded from loss; 6,000 AdamW updates; 9,216,000 presented tokens; warmup 100, learning rate 0.002 to 0.0002 by cosine decay; weight decay 0.01; gradient clipping 1; four CPU threads; seeds 42/43. Use the exact original train/validation cache hashes and fixed validation prefixes. No extra tuning, early stopping, changed tokenizer or test feedback. All models allocate 600,003 parameter entries. The `no_slow` variant sets slow state to zero from the first update and retrains all remaining effective machinery. Its 1,024 beta entries are disconnected from the objective. Keep the original pooling/readout and LayerNorm widths to avoid an additional architecture change. Normalization can make the nominal slow feature columns nonzero after centering, so those readout columns are not necessarily unused. Report this allocated-versus-effective distinction. This tests the implemented slow-state branch, not all possible capacity-matched memory alternatives. Save optimizer state and sampler/PyTorch RNG state at each validation checkpoint and resume only matching identities. Record protocol, numerical source, graph, tokenizer and cache hashes. Verify reference source compatibility, shared initial parameter hashes, sampled-token stream hashes and final exposure/RNG. ## Selection, evaluation and reporting Choose each run's checkpoint by lowest validation bits per UTF-8 byte among updates 500, 1,000, ..., 6,000, retaining the earliest on an exact tie. All eight new runs must finish and all ten selected checkpoint identities must be frozen before scoring any new control on the test split. Existing measured test scores are already public; reusing them requires exact checkpoint and data identity. Score all 60 official test articles, resetting state per article and carrying it within each article, chunk size 96. Exclude inserted boundary IDs and use exact decoded UTF-8 byte denominators. Report bits/byte, shared-tokenizer perplexity and per-article losses for every run. Word-token published WikiText perplexities are not directly comparable. Primary topology effects are `measured BPB - rewired BPB`, individually for all six graph/training-seed combinations; negative favors measured wiring. Report each graph's mean across its two training seeds, each training seed's mean across the three graphs, and the grand mean. Show all values and ranges; six contrasts sharing two measured references are not six independent replications. No binary significance claim or topology-superiority headline is licensed by three null graphs and two initializations alone. Primary slow-state effects are `measured BPB - retrained no_slow BPB`, separately for seeds 42/43 and as their mean. Do not substitute acute inference-time disabling for this retrained comparison. The design does not estimate a topology-by-slow-state interaction: null graphs without slow state would need a separately declared follow-up. For each individual paired contrast, retain the existing 10,000-replicate paired article bootstrap (seed 31415). Its 95% interval is conditional on the two trained checkpoints; it does not represent graph, training or dataset uncertainty. Do not pool articles across repeated seeds as extra independent documents. Validation curves at identical updates are secondary descriptive learning-efficiency evidence. Partial validation snapshots must say that test comparisons are pending and never select a preferred null graph. A positive, null or negative result is publishable. A null result means no advantage demonstrated under this specific subset, parameterization, data and budget; it is not proof that connectomes never help. Cue/context toy tasks, body animation and physical command execution cannot answer this language question. The transformer and GRU comparisons remain visible as separate architecture benchmarks. ## Scientific context - Maslov & Sneppen (2002), [Specificity and stability in topology of protein networks](https://doi.org/10.1126/science.1065103): motivates degree-preserving edge-swap nulls; our sign, weight-slot and self-edge constraints are explicit additions for this model. - Lappalainen et al. (2024), [Connectome-constrained networks predict neural activity across the fly visual system](https://doi.org/10.1038/s41586-024-07939-3): evidence in a visual-system activity-prediction setting, not language. - Dhiman (2026), [Topological Sensitivity in Connectome-Constrained Neural Networks](https://arxiv.org/abs/2604.04033v1), preprint: a short-horizon flyvis study reports that shared from-scratch initialization and degree-preserving controls weaken apparent topology advantages. Its task and training horizon differ from FLM; this is motivation for controls, not a predicted FLM result.