Executable formal frameworks for cognitive AI.

A small research lab running an open-science programme on knowledge consolidation in artificial cognitive systems. Based in Grandris, Beaujolais, Auvergne-Rhône-Alpes, France. 100 % open-source. Code : MIT (most repos) / Apache-2.0 (micro-kiki, inherited from Qwen). Docs : CC-BY-4.0.

The lab builds executable formal frameworks that bridge cognitive-neuroscience theory and machine-learning practice. The flagship programme axiomatises dream-based knowledge consolidation and instantiates it on two structurally divergent substrates — an MLX gradient-based language model (kiki-oniric) and a numpy-LIF thalamocortical spiking network (esnn-thalamocortical). Communication between heterogeneous modules is formalised as a substrate-agnostic Nerve Protocol with a measured polymorphism envelope. A behavioural benchmark (bouba_sens) stress-tests the same machinery on cross-modal plasticity.

Open-science discipline is public-by-default :

• Pre-registered hypotheses on OSF — 10.17605/OSF.IO/Q6JYN (locked 2026-04-19).
• Bit-exact reproducibility contract R1 : every experimental claim resolves to a deterministic run_id keyed on (c_version, profile, seed, commit_sha, benchmark_version).
• Dual-axis versioning (DualVer) : the formal-consistency axis (FC) and the empirical-consistency axis (EC) bump independently.
• No AI co-authorship trailer. Byline credits project contributors ; individual authors are credited in §Acknowledgments per publication.
• Internal critical-validation before external review. Every pre-registered finding passes through null-model controls, bootstrap confidence intervals, and multi-estimator robustness checks before publication. 2026-04-21 proof-of-concept: the bouba_sens v0.3 findings were retracted by the lab (v0.5.0, ADR-0006) when Sprint 7-8 critical tests downgraded all three to null results. See bouba_sens/papers/paper1/main.md for the methodology write-up.

• dream-of-kiki — substrate-agnostic formal framework for dream-based knowledge consolidation. Axioms DR-0..DR-4, executable Conformance Criterion (C1 signature typing, C2 axiom property tests, C3 blocking invariants), 8 typed primitives, 4 canonical operations (replay, downscale, restructure, recombine), 5-tuple Dream-Episode ontology. Two substrates pass all three conformance conditions : MLX kiki-oniric (cycle-1 canonical) and E-SNN thalamocortical (numpy LIF spike-rate skeleton, synthetic substitute — no Loihi-2 hardware). Gates G1, G7, G8, G9 LOCKED (cycle 2 closed) ; G10 deferred to Paper 2 per the 2026-04-20 PLOS CB pivot. Current paper : Paper 1 v0.2 (frozen 2026-04-20, 22 p) — primary submission target PLOS Computational Biology (retargeted today from Nature Human Behaviour). arXiv deposit ready (cs.LG primary, q-bio.NC + cs.AI cross-list), pending only the web-UI walkthrough. OSF DOI mint PENDING (gated on arXiv-ID lock). Paper 2 scaling-law observations (p_max between 1.5B and 7B) are in backlog ; any confirmatory rejection is gated on a 3rd scale point, bootstrap CI on per-seed p-values, numerical-precision audit, and log-probability arithmetic if underflow is detected — same critical-validation discipline that triggered the bouba_sens v0.5.0 null-results retraction (ADR-0006). Python 3.12+, MLX on Apple Silicon. Internal DualVer C-v0.7.0+PARTIAL (git tag v0.6.0, 201 commits, 277 tests / 91.17 % coverage). MIT + CC-BY-4.0.

• kiki-flow-research — Wasserstein-gradient-flow engine for LLM consolidation, grounded in the Levelt-Baddeley model of language production. Four psycholinguistic species (ρ_phono, ρ_lex, ρ_syntax, ρ_sem) evolve under a JKO-regularised flow on commodity Apple Silicon. Three parallel tracks :
  • track1_perf — nightly offline consolidation over the LoRA stack tensor (40 hybrid species, ~30 min on Studio M3 Ultra). Tag track1-v0.1.
  • track2_paper — N-particle Langevin × JKO rigorous proxy, 4 Levelt-Baddeley species, 5-seed paper figures. Tag track2-v0.1, paper draft advanced to paper-v0.7-draft ; phase1-pilot10k-done reached, scale-50k prep (T22–T23) complete.
  • track3_deploy — pure-NumPy streaming surrogate, MiniLM query encoder, advisory-only routing (p50 ≈ 0.04 ms on M5). Tag track3-v0.1 ; v0.3 text-bridge surrogate just merged (PR #6) with query-conditioned free energy + g_JEPA phase-A pre-training + 4-species simplex init fix. 153 commits, 93 tests, 95 % coverage, strict ruff + mypy. Zenodo-GitHub integration live — concept DOI 10.5281/zenodo.19666755, latest-version DOI 10.5281/zenodo.19666756 for paper-v0.8-draft. MIT.

• nerve-wml — substrate-agnostic Nerve Protocol for inter-WML communication. Discrete neuroletters, γ/θ multiplexing, sparse learned topology. 10 gates passed (protocol, polymorphism, merge, scaling to N=32, interpretation, neuromorphic INT8 export, adaptive alphabet, LLM advisor, realism, temporal MI ; gate-dream is partial — awaits kiki-oniric v0.5+). Three substrates share the Protocol : MlpWML (dense), LifWML (surrogate-gradient SNN), and TransformerWML (attention-based, added in v1.2). GammaThetaMultiplexer finalised with Q1–Q5 arbitrated (issue #1 closed). v1.2 "realism" axis delivered : HardFlow MLP↔LIF polymorphism gap decays N=2 10.71 % → N=16 6.71 % → N=32 2.39 % → N=64 2.73 % plateau, MNIST gap 1.03 %, 16-token temporal streaming MI/H = 0.72, mutual information MI($c_\mathrm{MLP}$;$c_\mathrm{LIF}$)/H($c_\mathrm{MLP}$) = 0.91–0.96 across seeds. The 12.1 % XOR-projected gap previously reported was traced to a decoder asymmetry bug, not a substrate limit. Latest tag v1.2.3 (2026-04-20), 184 commits on master. Paper v0.9 draft (paper-v0.9-draft, built PDF at papers/paper1/main.pdf) cited in flagship Paper 1 §7.4 as independent corroboration of substrate-agnosticism. Zenodo concept DOI 10.5281/zenodo.19656342, next-version DOI 10.5281/zenodo.19666405 reserved for v1.3.0. MIT.

• micro-kiki — deployable 34-LoRA + cognitive-layer runtime on Qwen3.6-35B-A3B (Apache 2.0 base, 262 K context, native MoE, 256 experts, 3 B active). Router emits 35 sigmoid outputs (34 domain niches + base slot ; domains co-activate, not mutually exclusive). 10 SFT adapters trained (kicad-dsl, spice-sim, emc, stm32, embedded, freecad, platformio, power, dsp, electronics ; final losses 0.38–0.55) ; V4 SOTA training in progress (32L, rank=α=16, chat-fr val_loss 0.849, reasoning 0.638 at iter 100). Sequential per-domain MLX training on Mac Studio M3 Ultra (peak ~107 GB, BF16) ; Q4_K_M inference on RTX 4090. Cognitive layer : Aeon memory (SIMD-vector Atlas + neuro-symbolic Trace graph), CAMP negotiator with Catfish dissent, KnowBias double-filter (RBD + DeFrame). New kiki-flow serving bridge merged at HEAD (v0.3 advisory-routing integration). Adjacent research artefact — triple-hybrid Quantum VQC + SNN + Classical routing documented in docs/paper-outline-triple-hybrid.md (Paper B, ICONS / Frontiers target). Validated as PoC, not in the production runtime. Latest tag v0.3-rc1, 437 commits on bridge-surrogate-v0.3, 971 test functions across 126 test files. Zenodo concept DOI 10.5281/zenodo.19666759, v0.3.0 DOI 10.5281/zenodo.19666760. Apache-2.0 (inherited). Training datasets + adapter-mixing experiments live in the shared fine-tuning infrastructure L-electron-Rare/KIKI-Mac_tunner — a joint MLX pipeline (Mac Studio M3 Ultra / M4 Pro) operated from the L'Électron Rare organisation and used to train both FineFab industrial models and Hypneum research substrates ; models and hyperparameter search are versioned separately per programme.

• bouba_sens — cross-modal plasticity benchmark. A 5-modality agent (audio, vision, tactile, gravity, force) undergoes a two-arm lesion protocol — congenital (T1, pre-training) vs late-acquired (T2, post-convergence) — combined with progressive SNR degradation. Three pre-registered invariants evaluated against fixed thresholds on every grid run :
  • B-1 Congenital gap — Me7 > 0.05
  • B-2 MI migration — Me3_delta > 0.10 bit
  • B-3 Perceptive/proprioceptive asymmetry — Me6 max-abs off-diag > 0.02 Built on top of nerve-wml (shared neuroletter codebook, γ/θ multiplexing) with adaptive gating, adaptive codebook expansion, and learnable cross-modal transducers. State (2026-04-21, v0.5.0) — Sprint 8 closed. All three pre-registered findings have failed critical validation — the programme is now a methodology-first contribution with an active paper draft. F1 B-3 is a 3+2-partition tautology (pre-reg at 12.5th percentile of an n=8 null-model distribution, docs/adr/0006-critical-validation.md). F2 B-1 sign flip is sampling noise (bootstrap 95 % CIs straddle 0 across all three worlds). F3 B-2 decay is a Kraskov-k-NN artefact at n=16 probe batch (binning + MINE both return median 0). Paper 1 v0.1 draft (papers/paper1/main.md, ~9 pages) reframes the programme around three methodological recommendations: probe batch size ≥ 128 for MI estimators, bootstrap CIs mandatory for sub-threshold effects, null-model partition controls for group-asymmetry claims. The Popperian guardrails fired before external review — this is the lab's first null-results publication, tagged v0.5.0. Target venue TMLR or NeurIPS D&B.

• micro-kiki-quantum — QPU-backed R&D satellite of micro-kiki. Three research tracks : Quantum-PEFT (Pauli parameterisation, arxiv 2503.05431), QMoE interference-based routing (arxiv 2507.05190 / 2512.22296), Hybrid Quantum Transformer (HyQuT) (arxiv 2511.10653). Hardware targets : IonQ Aria 25 q via AWS Braket (primary), IBM Eagle 127 q via IBM Quantum Network (secondary), Qiskit Aer simulation offline. Hard budget cap $250/experiment with $200 hard-abort checkpoint, total envelope ~$2.5–5 k. Scaffold stage (3 commits, no tag, first experiments pending). Apache-2.0. Results back-port to micro-kiki/quantum-inspired only if classical simulation validates them.

      ┌────────────────────────────────────┐
      │  kiki-flow-research                │  upstream numerical engine
      │  Wasserstein JKO + species         │  (ρ_phono, ρ_lex, ρ_syntax, ρ_sem)
      │  153 commits · Zenodo DOI live     │  paper-v0.7-draft + pilot10k-done
      └──────────────┬─────────────────────┘
                     │
                     ▼  numerical primitives
      ┌────────────────────────────────────────────────────────┐
      │  dream-of-kiki  (flagship, formal framework)           │
      │  Axioms DR-0..DR-4, Conformance Criterion              │
      │  Substrates : MLX kiki-oniric + E-SNN thalamocortical  │
      │  DualVer C-v0.7.0+PARTIAL · 277 tests / 91 % cov       │
      │  Paper 1 v0.2 → PLOS Comp Bio    Paper 2 → NeurIPS     │
      └──────┬──────────────────────────┬──────────────────────┘
             │                          │
   cites §7.4│                          │ε-trace bridge (gate-dream partial)
             ▼                          ▼
   ┌──────────────────────┐   ┌──────────────────────┐
   │  nerve-wml  v1.2.3   │   │  micro-kiki          │
   │  10 gates · 184 cmt  │   │  34 niches + base    │
   │  MlpWML · LifWML ·   │   │  Qwen3.6-35B-A3B     │
   │  TransformerWML      │   │  971 tests · Aeon    │
   │  Zenodo auto-DOI     │   │  Zenodo auto-DOI     │
   └──────────┬───────────┘   └──────────┬───────────┘
              │ shared codebook           │ gate-llm-advisor-passed
              ▼ (γ/θ multiplexer)         │ (NerveWmlAdvisor)
   ┌──────────────────────┐               │
   │  bouba_sens v0.5.0   │               │
   │  5-modality lesion   │               │
   │  B-3 PASS 3/3 worlds │               │
   │  ADRs 0003→0005      │               │
   └──────────────────────┘               │
                                          ▼
                               (integration back into micro-kiki runtime)

See each repo's docs/ARCHITECTURE.md / design spec for the detailed interface contracts.

• Pre-registration — OSF project dreamOfkiki · DOI 10.17605/OSF.IO/Q6JYN (locked) · Wiki page Amendments with filing procedure.

• Code archives — Zenodo-GitHub integration live on three repos :

  • nerve-wml — concept 10.5281/zenodo.19656342
  • kiki-flow-research — concept 10.5281/zenodo.19666755
  • micro-kiki — concept 10.5281/zenodo.19666759

  dream-of-kiki DOI mint is pending the arXiv-ID lock ; bouba_sens mints when mirrored on OSF.

• Preprints — arXiv (Paper 1 v0.2 bundle ready, awaiting web-UI deposit ; cs.LG primary + q-bio.NC / cs.AI cross-list) · HAL FR (mirror post-arXiv-announcement).

• Models — HuggingFace clemsail/kiki-oniric-* (posted when S8 PUBLICATION-READY gate is achieved).

• Data — Zenodo DOIs for raw runs at milestones S20 – S22 (experiments start S5).

• Clément Saillant — founder, corresponding author, maintainer. GitHub · L'Électron Rare · Blog · clement@saillant.cc Founder of L'Électron Rare — custom electronic systems (design, consulting, training): electronics, automation, energy. 10 years of field experience. Open-source contributor (KiCad, KXKM). Research at Hypneum Lab is conducted alongside this consulting activity, guided by intuition and a disciplined use of AI as a scientific research tool; the verifiability of artefacts (pre-registration, null-milestone, reproducibility) is the chosen counterpart of a non-traditional path. Grandris (69), Auvergne-Rhône-Alpes, France.
• Future visiting collaborators and co-authors would be credited per publication.

Research-first discipline. Pull requests welcome, especially on :

• Formal proofs for the DR-0..DR-4 axiom family (freeness / universal property of the operation semigroup is explicitly open).
• Conformance Criterion stress tests against alternative substrates (transformer-based, analog neuromorphic, JEPA-class).
• Reproducibility tooling — shared run_registry, invariants harness, benchmark integrity.
• Translations and accessibility improvements.

Policy :

• Every experimental claim must resolve to a registered run_id or a proof file (contract R1).
• No AI attribution in author lists, acknowledgments, or bibliographies. LLM assistance is flagged in PR descriptions.
• Byline per publication : <repo> project contributors with Clément Saillant as corresponding author.

See CONTRIBUTING.md in the flagship repository and each repo's CONTRIBUTORS.md for detailed policies.

Research / collaboration inquiries : clement@saillant.cc

For L'Électron Rare business (custom electronic systems design and consulting — electronics, automation, energy · training · FineFab platform · Kill_LIFE / Mascarade products · AURA regional partnerships) see www.lelectronrare.fr. Business codebase and shared fine-tuning infrastructure (KIKI-Mac_tunner) live on the separate L-electron-Rare GitHub organisation ; the MLX toolkit is used to train both industrial and Hypneum research models, with separate data and run_registry keys per programme.

Last updated : 2026-04-20 21:05 CEST (deep-refresh from live repo state, post Sprint 5 + Sprint 6 opening). Snapshot : dream-of-kiki v0.6.0 / DualVer C-v0.7.0+PARTIAL (201 commits, 277 tests / 91 % cov, Paper 1 v0.2 PLOS CB pivot, scaling-law p_max 1.5B→7B observation flagged for critical-validation gating per bouba_sens ADR-0006), kiki-flow-research 153 commits + paper-v0.7-draft + phase1-pilot10k-done (Zenodo DOI live), nerve-wml v1.2.3 (184 commits, 10 gates, 3 substrates MlpWML+LifWML+TransformerWML, MNIST 1.03 % gap, Zenodo DOI live), micro-kiki v0.3-rc1 HEAD 44b1a01 (437 commits, 971 tests, kiki-flow bridge merged, Zenodo DOI live), bouba_sens v0.5.0 (Sprint 8 closed, 6 ADRs, paper v0.1 draft) — 3/3 pre-registered findings DOWNGRADED via null-model + bootstrap + multi-estimator validation. The programme's Popperian guardrails fired before external review ; paper reframes as methodology-first with three reusable recommendations (probe batch, bootstrap CIs, partition controls). Null results are the scientific contribution.