nerve-wml

Substrate-agnostic nerve protocol for inter-module communication in hybrid neural systems.

Citation : each release is archived on Zenodo (concept DOI 10.5281/zenodo.19656342 resolves to the latest version) and linked to the parent programme's OSF pre-registration (10.17605/OSF.IO/Q6JYN).

Research engine that validates a discrete-code communication layer between heterogeneous neural modules (World Model Languages, or WMLs). Modules exchange neuroletters over a sparse learned topology, multiplexed on gamma/theta rhythms, and converted between local codebooks by per-edge transducers. The paper draft is at papers/paper1/main.tex; the full spec is at docs/superpowers/specs/2026-04-18-nerve-wml-design.md.

2026-05-20 — Gap-analysis remediation sprint

Reinforcement sprint that hardened the empirical claims via systematic statistical scrutiny. 97 commits, 13 PRs merged, +458 fast-suite tests passing (from 385 baseline). Detailed entry in CHANGELOG.md.

6 initial claims revised under scrutiny:

1. learned strictly dominates relative_rep → tie at n_anchors=64 (Δ=0.0001, p=0.627, n=50, bit-exact M5 ↔ macM1).
2. GTM doesn't collapse (synchrony 0.20 > 0.08) → metric anti-monotone (null=0.32 > GTM=0.20 > simple_gating=0.08); replaced with spectral_entropy (Roy & Vetterli 2007 effective rank).
3. AKOrN minimal clusters with simple_gating → sub-parametrised; at n_oscillators=64, n_steps=32, lr=0.05 AKOrN reaches synchrony 0.45.
4. CKNNA grows with N (signal) → mechanical decay artefact (Gröger, Wen & Brbić 2026 Prop 4.2). Signal/null d_z grows monotonically.
5. HSIC golden=8.5320 is stable scientific number → standalone HSIC non-informative (|r|≤0.015 with 7 other alignment metrics on 1750 cells); always normalise to linear-CKA.
6. spectral_entropy 4-arm strict ordering → B-dependent (Renf 10); robust 2-arm: gtm > null, gtm < simple_gating at all B.

7 findings now publishable:

• AKOrN top-cell at n=50 = phase-aligner not coder: synchrony 0.53 ± 0.20 (cross-host: macM3 0.49 ± 0.19), accuracy bimodal 0.38 ± 0.31.
• 3/4 substrates (gtm, simple_gating, null) bit-exact cross-arch (M1 g13s ↔ M3 Pro g15s ↔ M5 g17g) under PyTorch deterministic path; Kuramoto Euler integrator is the only cross-arch-divergent arm.
• MLX random.normal cross-arch non-bit-exact (M1 vs M3+/M5), opened upstream as ml-explore/mlx#3568.
• macM3 ~2-3× faster than macm1 on Python+torch multiproc CPU.
• MLX 2× faster than torch.MPS at N=16384 for CKNNA Gram (scripts/cknna_mlx_bench.py).
• Task scope (alphabet=64, K=7) saturates gtm and simple_gating equally; spectral_entropy is the only discriminating axis.
• Empirical confirmation of Gröger 2026 Prop 4.2 E[mKNN]=k/(n-1) on synthetic Gaussian substrate (see papers/paper1/supplementary/cknna-n-dependence-replication/).

Methodology institutionalised:

The protocol every numerical claim must trace to a JSON cell or executed log line in the same session is now enforced by scripts/factcheck_audit.py --ci and runs as a GitHub Actions check on every PR touching docs/superpowers/research/ or the audit script. Current state: 32 OK / 0 DIVERGENT / 2 ORPHAN (orphans = JSONs on unmerged branches).

2026-05-11 milestones — N8 → N14 portfolio sprints

Tonight's six-sprint push closed the GammaThetaMultiplexer (GTM) empirical arc and locked four scaling pre-registrations for future hardware sprints. Headline numbers below remain the v1.2.3 baseline ; the new work characterises the GTM contract envelope and shipped six critic-driven fixes (ablation void, degenerate-metric handling, Cat C honest framing, Jonckheere wrong test, β-VAE confound, paper/JSON count mismatch).

Sprint	Verdict / artifact	Notes
N8 Q1 GTM benchmark (HardFlowProxyTask N=2)	tied (3W / 5L / 1T vs latent-space baselines)	mi_h dominance preserved across all 8 conditions
N9 Q1+ scaling N=16	tied-stable (3W / 5L / 1T identical to N=2)	seed-stable across 5 seeds
N9 Q1++ FlowProxyTask 4-class	tied-stable-cross-task (3W / 1L / 2T post-degenerate-exclusion)	confirms tie generalises beyond HardFlow
N8 Q2 15 adversarial substrates (Cat A/B/C)	ge_3_FP_reformulate	Conformance Criterion C+ now requires C2 axiom property tests in addition to structural invariants
N9 Q2+ Cat D modular / E replay-loop / F asymmetric (10 more)	25 / 25 cumulative FP	structural layer alone insufficient at 6 categories
N10-A / N10-C	pre-reg locked, deferred	ImageNet-100 + ResNet50 + GTM bottleneck (β-VAE β=1.0 fixed per critic) ; Procgen multi-agent GTM communication
N11-A / N11-C	pre-reg locked, deferred	I-JEPA + GTM predictor on ImageNet self-supervised ; Dreamer V3 + GTM dynamics on Atari-100k
N13 / N14	pre-reg locked, deferred	OFAT GTM internals (5 hyperparams × 4-5 levels × 5 seeds) ; Latin Hypercube 7-dim coverage (50 points × 5 seeds)

Cumulative on master HEAD 7528ada : 295 commits, v1.8.1 released on PyPI, 12 OSF-style pre-registrations across the Hypneum programme, 7 verdicts shipped on the GTM axis, 5 pre-regs deferred to future scaling sprints, 1 critic-saver retract closed downstream (bouba_sens Q3+, see that repo's README).

Methodology lessons promoted to portfolio-wide discipline : multi-seed-first-class (single-seed claims are fragile, see bouba_sens N9 Q3 Retract evidence), degenerate-metric handling in analyse.py, Cohen's d effect sizes added to verdict tables, honest reporting of underpowered ablations (the cosine plasticity schedule's effect on GTM outcomes is below the noise floor for HardFlowProxyTask : the constellation is 128 of 17,888 trainable parameters (0.7 %), so scaling its gradient by a cosine schedule does not produce a measurable behavioural difference (Δ < 0.005 RTF, less than seed standard deviation). This is a negative result consistent with the convergent-evidence framing : on HardFlowProxyTask, plasticity scheduling is not load-bearing — the discrete-channel structure (PSK alphabet + Gumbel quantization) drives the mi_h advantage, not the temporal modulation of constellation learning. A larger fraction of the model would need to participate in the schedule for the ablation to discriminate ; we leave this as a methodological observation for future work, e.g. per-symbol learned embeddings, or a constellation_lock_after ablation with the constellation frozen mid-training).

---

2026-05-19 update — related-work consolidation + BioFieldWML plan

Paper 1's §Related Work grew from 5 to 6 threads (new "Adjacent abstraction layers" thread, NIR at model-graph IR + AER at transport — orthogonal abstraction layers, not competitors), §Information Transmission Test (6) added a "Global vs Local alignment" paragraph addressing the two 2026 PRH critiques, and the GTM rationale was re-grounded on Bastos 2020 / Friston 2025 / Ruffini 2025. In parallel, a 4th conformant substrate plan (BioFieldWML) was versioned for the first time, sitting alongside the existing MlpWML / LifWML / TransformerWML substrates under the WML Protocol.

Item	Where	What
§Related Work extended (5 → 6 threads)	papers/paper1/main.tex	New "Adjacent abstraction layers" thread positions nerve-wml against NIR (model IR) and AER (transport) — orthogonal abstraction layers
§Information Tx Test (6) PRH refinement	papers/paper1/main.tex	New "Global vs Local alignment" paragraph engages platoscave2026 + aristotelianprh2026; nerve-wml's mutual-kNN measurement claimed as local (null-calibrated 18.8× random), not global
GTM re-grounding	papers/paper1/main.tex	Replaces sole reliance on Bastos & Friston 2012 with Bastos 2020 predictive routing + Friston 2025 + Ruffini 2025 laminar Comparator
HNN positioning (closes #13)	papers/paper1/main.tex "Surrogate-gradient SNN" thread	Cites Liu et al. 2024 (NSR) and positions nerve-wml in the weak-coupling corner of the HNN taxonomy
BioFieldWML plan versioned	docs/superpowers/plans/2026-05-19-bio-substrate-wml.md	Plan for a 4th conformant substrate alongside MLP / LIF / Transformer; 5 (b)-classified mechanisms encapsulated inside step() (Tomé STDP, Pignatelli IE, Palacios SNN-PC, Bellitto WSCL, Tucker-Friston E/I); cross-refs dream-of-kiki biophysical-stratification spec

Headline measurements and claims A/B remain unchanged — this update is documentation and planning only, not new experimental findings.

(i)+(ii) framing: BioFieldWML is the (i) side — one conformant substrate among N, DR-3 preserved, fully encapsulated under the existing WML Protocol. The (ii) side is the new dream-of-kiki biophysical-stratification sub-theory (dream-of-kiki/docs/specs/2026-05-20-biophysical-stratification.md). The two documents are linked but independent: nerve-wml's bio-substrate plan does not depend on the stratification theory being finalised, and vice versa.

Open follow-up: nerve-wml#16 — complete friston2025pivot bibliographic metadata (volume/pages/DOI) before final paper submission.

---

Status — v1.8.0 (2026-04-24, on PyPI)

Installable via pip install nerve-wml. For the real kiki_oniric.axioms integration (dream-of-kiki bridge), side-install dreamofkiki first (PyPI rejects VCS URLs in published metadata, so no [axioms] extras group is shipped):

pip install "dreamofkiki @ git+https://github.com/hypneum-lab/dream-of-kiki@v0.9.1"
pip install nerve-wml

Six releases landed on 2026-04-21 → 2026-04-24 (v1.4.0 → v1.8.0) on top of the v1.2.3 scientific baseline; see § Post-v1.2.3 API additions below or CHANGELOG.md for the per-version diff. The scientific claims below are the v1.2.3 baseline and remain load-bearing — the newer releases added opt-in knobs (plasticity schedule, Gumbel-softmax gating, spectrogram encoder, dreamOfkiki axiom bridge scaffold) and the nerve_wml.methodology submodule with the four MI robustness primitives (null-model, bootstrap CI, Miller-Madow, Kraskov KSG, MINE) — all without changing any headline measurement.

The project is empirically defensible across three experimental axes: real data, architecture scale, and temporal streaming. Two claims are quantified:

Claim A — Substrate-agnostic polymorphism (task competence converges). Three structurally distinct substrates (stateless MLP, spiking LIF with surrogate-gradient, attention-based Transformer) reach comparable accuracies via the shared Nerve Protocol.

Claim B — Substrate-agnostic information transmission (codes align). Independent substrates share 91–96 % of their emitted code information; a frozen LIF can recover a trained MLP's task competence via a learned linear transducer.

Headline measurements

Axis	Finding	Reference
Pool scaling law (MLP ↔ LIF, HardFlow)	$N=2 \to 10.71%$, $N=16 \to 6.71%$, $N=32 \to 2.39%$, $N=64 \to 2.73%$ plateau. 5 % contract holds distributionally at $N \geq 32$.	figures/w2_hard_scaling.pdf
Triple-substrate pool (MLP + LIF + TRF)	$N=15 \to 8.16%$, $N=30 \to 5.86%$, $N=60 \to 4.33%$	v1.1.4
Mutual information (codes MLP ↔ LIF)	$\mathrm{MI}/H = 0.91$ at $N=1$ (5 seeds), 0.96 at $N=16$ pool (192 cross-pairs)	figures/info_transmission.pdf
Round-trip fidelity (MLP → LIF → MLP)	0.99 mean (3 seeds)	v0.8
Cross-substrate merge (LIF fed by MLP codes only)	0.97 mean (3 seeds)	v0.8
MNIST real data	MLP 0.942, LIF 0.941, gap 1.03 %, MI/H 0.882	figures/mnist_scaling.pdf
MoonsTask (2nd distribution)	MI/H = 0.74 (3 seeds)	v1.1.4
Architecture scale ($d_\text{hidden}=128$)	Gap AMPLIFIES to 26 % on XOR (arch vs pool scale are orthogonal); Claim B survives	figures/bigger_arch_scaling.pdf
Temporal streaming (16-token sequence)	MI/H = 0.72 at trained step, 0.71 at filler step — structural alignment	figures/temporal_info_tx.pdf
Platonic RH alignment (Huh 2024, pre-VQ mutual-kNN)	MLP ↔ LIF = 0.174 at k=10 (18.8× random, 3 seeds); stable across k∈[5,50]	figures/platonic_rh_alignment.json
Real neural data (Sleep-EDF EEG, v1.6.0)	See paper Test (9); Claim B confirmed on 5-class sleep-stage via MlpWML.from_spectrogram + d_hidden=128	figures/mi_eeg_d128_spectro.json
Frozen-encoder baseline (review F3, v1.7.0)	Shared MI/H=0.95 (matches nerve-wml Test 1), Distinct MI/H=0.76 (without shared frontend); Claim B reframed as "VQ protocol supplies shared frontend through codebook"	figures/baseline_frozen_encoder.json
Matched-capacity scale sweep (Sleep-EDF, v1.7.0)	Sweet spot at d=128: MI/H=0.72, MLP=0.82, LIF=0.83, gap=0.006. Scale-invariant polymorphy at d ∈ {32, 64, 128}. d=16 insufficient for LIF convergence on real EEG; d=256 MLP overfits while LIF holds	figures/eeg_matched_scale_sweep.json
Direction stability (LIF ≥ MLP on hard task)	19/20 pairwise seeds (4/5 at N=2; 5/5 at N=16, 32, 64) + 5/5 triple-substrate, preserved on Sleep-EDF (+0.007 LIF edge)	—

LIF's spike dynamics give it a substrate-intrinsic $\sim 2$–$3%$ expressivity edge on XOR-style boundaries (plateau floor). Pool averaging compresses this, architecture width amplifies it.

Seven concrete findings

1. The original 12.1 % gap was a decoder asymmetry bug, not a substrate limit. LIF had a fixed cosine decoder, MLP had a learned head; symmetrizing flipped the sign (LIF now leads).
2. Single-seed measurements lie. Multi-seed revealed the N=16 median is 6.7 %, not the lucky 1.68 %.
3. Scaling law is real and monotonic. Four-point decay $10.7% \to 6.7% \to 2.4% \to 2.7%$ plateau.
4. Claim B is empirical, not architectural. MI 0.91–0.96, round-trip 0.99, cross-merge 0.97.
5. Substrate-direction is stable in 19/20 seeds. LIF's spike edge is a real property, not noise.
6. Architecture scale and pool scale are orthogonal. Pool compresses the gap; arch width amplifies it.
7. Code alignment is structural, not task-gated. MI at filler timesteps $\approx$ MI at trained timesteps (0.71 vs 0.72).

Methodological findings (v1.2.1–v1.2.3)

• MI/H vs CKA on the same argmax codes (v1.2.1). Mean 0.953 (MI/H) vs 0.910 (CKA argmax one-hot) over 3 seeds. The 4.3 pp gap tracks soft many-to-one code mappings that kernel-alignment metrics miss. MI/H is not CKA renamed — it is the discrete-protocol cousin with measurably different semantics. See scripts/measure_cka_vs_mi.py and docs/positioning.md.
• Related Work verified (v1.2.2). Paper §Related Work cites Kornblith 2019 CKA, Morcos 2018 PWCCA, Moschella 2022 relative representations (ICLR 2023), Saxe 2024 universality, and Hinton 2015 KD — all verified via WebFetch, provenance table in docs/positioning.md.
• KD match-compute ablation honest verdict (v1.2.3). At matched compute on HardFlowProxyTask (3 seeds), cross-merge (0.508) ≈ KD-through-transducer (0.520) within noise. Vanilla Hinton KD (0.534) is best because the student can re-train its core. Cross-merge's contribution is methodological, not performance-based: it isolates protocol channel capacity from student learning capacity by freezing both substrates and supervising with ground-truth labels only. See scripts/measure_kd_ablation.py.

What the paper genuinely claims vs not

Three findings probably novel: (1) the four-point scaling law with plateau at $\sim 2\text{–}3%$ substrate-intrinsic floor, (2) reproducible $\sim 2\text{–}3%$ LIF spike-expressivity edge over matched-capacity MLP on XOR-on-noise (19/20 seeds), (3) orthogonality of pool-scale (compresses gap) and architecture-scale (amplifies gap).

The paper explicitly does not claim: a new learning algorithm, superiority over knowledge distillation on task accuracy, or universal representations — that debate is addressed by Saxe 2024 and the Nature MI 2025 editorial (s42256-025-01139-y) cited in docs/positioning.md.

Cross-lab methodology commitment

The sister project bouba_sens (2026-04-21, github.com/hypneum-lab/bouba_sens tag v0.5.0) demonstrated that pre-registered findings in this programme must pass three critical tests before publication: null-model partition controls, bootstrap confidence intervals on sub-threshold effects, and multi-estimator robustness checks for MI-based claims. As of v1.5.3 (2026-04-21) all three checks are implemented in nerve_wml.methodology and applied to the MI/H headline: null-model rejects chance at z > 1000 (p < 10⁻³ over 3 seeds × 1000 shuffles), bootstrap gives CI95 [0.82, 0.99] intra-seed width ~0.005, and discrete cross-estimator robustness holds between plug-in and Miller-Madow (Δ = 0.007). Two continuous estimators (Kraskov KSG and MINE) were applied to the pre-VQ embeddings; they diverge by an order of magnitude (KSG 0.09, MINE 0.99), making the pre-VQ absolute MI magnitude an open methodological question — see paper §Information Transmission Test (7). The post-VQ discrete MI/H headline is unaffected by this ambiguity.

Status — 11 gates

Tag	What it proves
gate-p-passed	Track-P protocol simulator correct on toy signals
gate-w-passed	MlpWML and LifWML interoperate with < 5 % gap through the same nerve (N=4)
gate-m-passed	Merge fine-tunes only transducers; retains ≥ 95 % of mock-baseline accuracy
gate-m2-passed	Four scientific shortcuts from §13.1 resolved with honest measurements
gate-scale-passed	Polymorphie + continual learning hold at N=16 pools; router stays connected to N=32
gate-interp-passed	Per-WML code → concept semantics table rendered as HTML
gate-neuro-passed	LifWML → INT8 artefact → pure-numpy mock runner (Loihi / Akida stubs documented)
gate-dream-passed	ε-trace consolidation bridge to dream-of-kiki (schema v0; partial — awaits kiki_oniric v0.5+)
gate-adaptive-passed	Per-WML alphabet shrinks/grows via active_mask + transducer resize
gate-llm-advisor-passed	Env-gated, never-raising NerveWmlAdvisor for micro-kiki, < 50 ms warm latency

Paper drafts: paper-v0.2-draft … paper-v0.9-draft track the iterations that produced the v1.2 claims above. Release tags v1.0.0, v1.1.0 … v1.1.4, v1.2.0, v1.2.3, v1.3.0, v1.4.0, v1.5.0, v1.5.1 archive the code snapshots; see CHANGELOG.md for per-version findings.

Post-v1.2.3 API additions (2026-04-21)

Three issues filed by downstream consumers (bouba_sens, dream-of-kiki) landed on 2026-04-21 as opt-in knobs — no change to v1.2.3 headline numbers, all new behaviour is off by default.

Release	Issue	Feature	Motivation (downstream)
v1.4.0	#4	GammaThetaMultiplexer gains plasticity_schedule + constellation_lock_after	bouba_sens B-1 Amedi-2007 gap directionally falsified in 4/5 worlds; biologically-distinct T1/T2 plasticity profiles are the probe.
v1.5.0	#5	Transducer gains TransducerGating.GUMBEL_SOFTMAX (opt-in soft distribution)	bouba_sens B-2 Me3-delta under-threshold in 5/5 worlds; hard argmax gating may be too abrupt for post-lesion MI migration.
v1.5.0	#7	MlpWML.from_spectrogram(...) factory + SpectrogramEncoder	DRY: bouba_sens MIT-BIH ECG fetcher + future Studyforrest audio share one canonical STFT → carrier path.
v1.5.0	#6	nerve_core.from_dream_of_kiki(...) scaffold (runtime gated upstream)	Pin the public axiom-bridge contract today so bouba_sens can plumb the call site before dream-of-kiki publishes its versioned axioms API.
v1.5.1	—	Packaging: pyproject.toml version sync (stale 1.4.0 on the v1.5.0 commit), CITATION.cff keeps concept DOI only.	v1.5.0 shipped with a stale version field — first PyPI release carries the correct metadata.

Design docs: docs/integration-dream-of-kiki.md, changelog files at docs/changelog/v1.4.0.md and docs/changelog/v1.5.1.md.

Install

# From PyPI (v1.5.1+)
pip install nerve-wml

# From source, with dev extras (tests + lint)
git clone https://github.com/hypneum-lab/nerve-wml.git
cd nerve-wml
uv sync --all-extras

Python 3.12+, macOS arm64 (MLX-friendly) or Linux x86_64. No vendor SDK deps are pulled by default (Loihi, Akida, dream-of-kiki, sentence-transformers are all optional integrations).

Run the suite

uv run pytest -m "not slow"    # 220+ tests under 80 s on commodity M-series
uv run pytest                  # full suite incl. paper figure rendering
uv run pytest --cov=nerve_core --cov=track_p --cov=track_w --cov=bridge --cov=harness --cov=interpret --cov=neuromorphic

Reproduce the gate numbers

uv run python scripts/track_p_pilot.py       # Gate P (+ Task 6 ablation)
uv run python scripts/track_w_pilot.py       # Gate W
uv run python scripts/track_w_pilot.py scale # Gate Scale (N=16, N=32)
uv run python scripts/merge_pilot.py         # Gate M
uv run python scripts/interpret_pilot.py     # Gate Interp (emits reports/interp/*.html)
uv run python scripts/adaptive_pilot.py      # Gate Adaptive

Reproduce the v1.1 / v1.2 findings

# v1.1 scaling law + information transmission + triple substrate
uv run python scripts/render_scaling_figure.py      # 4-point pool scaling (N=2..64)
uv run python scripts/render_info_tx_figure.py      # MI + round-trip + cross-merge
uv run python scripts/measure_info_transmission.py  # full info-tx battery

# v1.2 real data + bigger arch + temporal
uv sync --extra mnist                               # pull torchvision
uv run python scripts/render_mnist_figure.py        # MNIST Claims A + B
uv run python scripts/render_bigger_arch_figure.py  # d=128 gap amplification
uv run python scripts/render_temporal_figure.py     # streaming MI per timestep

Build the paper

uv run python scripts/render_paper_figures.py   # regenerate figures from frozen golden NPZs
cd papers/paper1 && tectonic main.tex           # or pdflatex, bibtex, pdflatex, pdflatex

Integrations (env-gated, default off)

• Dream consolidation: DREAM_CONSOLIDATION_ENABLED=1 + install dream-of-kiki locally → bridge.dream_bridge.DreamBridge.
• LLM advisor (micro-kiki): NERVE_WML_ENABLED=1 + NERVE_WML_CHECKPOINT_PATH=/path/to/checkpoint → bridge.kiki_nerve_advisor.NerveWmlAdvisor. Wiring recipe: docs/integration/micro-kiki-wiring.md.
• Neuromorphic hardware: install lava-nc or akida → wire in neuromorphic.loihi_stub / neuromorphic.akida_stub. Schema v0: docs/neuromorphic/deployment-guide.md.

Cited in

• dreamOfkiki — Paper 1 v0.2 (2026-04-19), §7.4 cross-substrate portability — github.com/hypneum-lab/dream-of-kiki. The Gate W and Gate M measurements reported here (MlpWML / LifWML polymorphism on FlowProxyTask and HardFlowProxyTask) provide the empirical corroboration cited in Paper 1 as independent evidence of the substrate-agnosticism principle (DR-3 Conformance Criterion). OSF pre-registration: 10.17605/OSF.IO/Q6JYN.

Program context

This repository is part of hypneum-lab, which develops executable formal frameworks for cognitive AI. The programmatic parent is dreamOfkiki (paper 1 formal framework, paper 2 empirical); nerve-wml is the reference implementation for the substrate-agnostic communication principle.

Sibling repositories:

• dream-of-kiki — formal framework (axioms DR-0..DR-4, Conformance Criterion, Paper 1)
• kiki-flow-research — Wasserstein-gradient-flow engine (upstream)
• micro-kiki — 35 domain-expert MoE-LoRA deployable instance (advisor consumer)
• nerve-wml (this repo) — substrate-agnostic nerve protocol + cross-substrate polymorphism

Repository layout

nerve_core/        Neuroletter, Nerve + WML Protocols, invariants (N-1..N-5, W-1..W-4)
track_p/           Track-P — SimNerve, VQCodebook, Transducer, SparseRouter, AdaptiveCodebook
track_w/           Track-W — MockNerve + 6 WML substrates (MlpWML, LifWML, TransformerWML, BioWML, BioFieldWML, SpikingKikiWML), toy tasks, training loop, pool factory. See `docs/substrate-paradigms.md` for the paradigm map.
bridge/            Merge, dream, LLM advisor — SimNerveAdapter, MergeTrainer, DreamBridge, NerveWmlAdvisor
harness/           R1 reproducibility — run_registry
interpret/         Gate Interp — code_semantics, clustering, HTML renderer
neuromorphic/      Gate Neuro — spike_encoder, INT8 export, mock_runner, vendor stubs
scripts/           All gate pilots + figure renderers + freeze_golden
tests/             Unit + integration + golden NPZ regressions
docs/              specs/, integration/, neuromorphic/, dream/, interpret/
papers/paper1/     LaTeX source + bib + Makefile (figures regenerated deterministically)

License

MIT (code) + CC-BY-4.0 (docs).