Executable formal frameworks for cognitive AI.

A one-person research lab running an open-science programme on knowledge consolidation in artificial cognitive systems — the GENIAL framework (dream-of-kiki), a substrate-agnostic Nerve Protocol (nerve-wml), a Wasserstein-gradient-flow engine (kiki-flow-research), a cross-modal plasticity benchmark (bouba_sens) and a deployable multi-LoRA runtime (micro-kiki). Based in Grandris, Beaujolais, France. 100 % open-source : code MIT (Apache-2.0 for micro-kiki, inherited from Qwen), docs CC-BY-4.0.

State of the programme on 2026-10-25. Dated milestone entries from 2026 are kept in docs/profile-journal-2026.md.

Six months in, the lab has formulated 180 hypotheses or claims and built the discipline to test them — pre-registration, five seeds minimum, Bonferroni, public errata. A full review of those claims was carried out on 2026-10-24 and revised on 2026-10-25 after two independent critiques (docs/research/2026-10-24-revue-hypotheses/, in French). Per its master table : 55 confirmed with ≥ 5 seeds or a calibrated null, 46 negative (refuted, retracted or null), 24 inconclusive, 19 amended, 12 pending, 19 never tested, 5 unverifiable. What holds is mostly negative or methodological : the retractions were made by the lab itself, before external review.

That review feeds the 2026-27 roadmap (docs/ROADMAP.md, docs/PROJECT-MANAGEMENT.md), tracked issue by issue on an internal GitHub Project of the organisation (47 issues, milestones M0–M3, ~10 h of research time per week) :

Principles : truth before novelty ; falsifiable before tested ; one pre-reg, one test, one verdict ; no new repo, no QPU, no wet-lab, no hardware purchase in 2026-27.

• dream-of-kiki — substrate-agnostic formal framework for dream-based knowledge consolidation. Framework hypotheses DR-0..DR-4, executable Conformance Criterion C+ (C1 signature typing, C2 axiom property tests, C3 blocking invariants), 8 typed primitives, 4 canonical operations (replay, downscale, restructure, recombine). Two substrates pass all three conformance conditions : MLX kiki-oniric and a numpy-LIF thalamocortical E-SNN. Paper 1 v0.2 (22 p) frozen ; arXiv deposit scheduled in phase 3 (January 2027) together with the DR-0/DR-1/DR-3 reformulations the review asks for. Draft PR #41 : three new pre-registrations (H12–H14) and a G4 erratum. main, 692 commits ; release v0.9.2 (2026-04-22), pyproject 0.23.0, DualVer C-v0.24.0+PARTIAL. Zenodo concept DOI 10.5281/zenodo.19701355. 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 evolve under a JKO-regularised flow on commodity Apple Silicon ; three tracks (nightly consolidation of a LoRA stack, N-particle Langevin × JKO paper proxy, pure-NumPy streaming surrogate for advisory routing). main, 227 commits ; latest release paper-v0.13-draft (2026-10-25) — Zenodo 10.5281/zenodo.23636477, concept DOI 10.5281/zenodo.19666755. October 2026 : two reproducibility bugs found and documented as errata (MLX prox_w2, solver max_iter ; the second changes no reported number, impact ≤ 1.7e-5), CI pinned (uv --locked), rigorous specialisation measure. MIT.

• nerve-wml — substrate-agnostic Nerve Protocol for inter-WML communication : discrete neuroletters, γ/θ multiplexing, sparse learned topology. Three substrates share the protocol (MlpWML, LifWML, TransformerWML) ; polymorphism gap MLP↔LIF decays with N (10.7 % at N=2 → 2.4 % at N=32). The GammaThetaMultiplexer verdicts are tied / tied-stable across N8/N9. N10-C (Procgen multi-agent, 6 seeds, pre-registered) was frozen on 2026-10-24 (annotated tag n10c-freeze-20261024, archived on Software Heritage) and launched on two hosts ; the first attempt was halted on 2026-10-25 — the image trunk collapsed before 1 M steps and the team reward proved too sparse. Diagnosis and re-run presets in draft PR #72 ; control-arm addendum N10-C-bis in draft PR #61. Non-determinism of MLX conv backward (Metal ~1e-7, CUDA ~1e-5 rel.) is declared in the pre-reg. master, 547 commits ; last tag v1.9.0, pyproject 1.9.4, last published release v1.8.1 (2026-05-10). Zenodo concept DOI 10.5281/zenodo.19656342. MIT.

• bouba_sens — cross-modal plasticity benchmark. A 5-modality agent (audio, vision, tactile, gravity, force) undergoes congenital vs late-acquired lesions under progressive SNR degradation ; three pre-registered invariants B-1 / B-2 / B-3. The lab's first null-results programme. All three original readings were overturned by its own critical validation (ADR-0006 v0.5.0 retraction ; §5.5 FINAL Retract ADR-0019 ; N12 subgroup replication LOST 0/2 at Bonferroni α=0.025). B-3 survives on magnitude only, and the review of 2026-10-24 asks for a calibrated extreme-value null on Me6 before any further claim. TMLR submission stays blocked until B-1/B-2 are reported as equivalence tests (TOST) and the Me6 null is calibrated. Draft pre-regs v0.7–v0.9 in PR #9 ; Me6 null-calibration work merged 2026-10-25 (#19). main, 211 commits ; release v0.5.9 (2026-05-10). Zenodo concept DOI 10.5281/zenodo.19701353. MIT.

• micro-kiki — multi-LoRA runtime with a cognitive layer on Qwen3.6-35B-A3B (native MoE, 256 experts, 3 B active, 262 K context) : 35 domain-expert V4 adapters (rank 16, scale 20), sigmoid router (top-1 70.7 %, top-3 89.5 %), Aeon episodic memory, CAMP negotiator, KnowBias double filter, kiki-flow advisory-routing bridge. October 2026 : MLX prompt-cache reuse fixed (#6). The Mac Studio that served it is powered down ; adapters and datasets are preserved, and the review lists a RAG-vs-Aeon baseline (NH-85) for phase 2. main, 598 commits ; release v0.3.0 (2026-04-20). Zenodo concept DOI 10.5281/zenodo.19666759. Apache-2.0.

• baby-brain — ontogenetic harness for substrate-agnostic World-Model Learners (649 commits ; milestones sp1, sp1.5-coreml, sp2-training shipped). Phase 0 of the roadmap removes the « PASS » claims that have no artefact ; phase 2 pilots a latent world model (JEPA-class) as substrate (NH-77).
• BabyBrain — macOS SwiftUI cockpit for the harness.
• whisperx-server — streaming speech-to-text (whisper.cpp + CoreML) feeding the Speaker-Identity subsystem.
• moshivis (public fork) — Kyutai's vision-speech model, reference substrate for multimodal grounding.
• micro-kiki-quantum — QPU-backed satellite, archived since 2026-05-10 (v0.0.1-archived). No QPU work in 2026-27.

      ┌─────────────────────────────────────┐
      │  kiki-flow-research                 │  upstream numerical engine
      │  Wasserstein JKO + 4 species        │  paper-v0.13-draft · Zenodo
      └──────────────┬──────────────────────┘
                     │ numerical primitives
                     ▼
      ┌─────────────────────────────────────────────────────┐
      │  dream-of-kiki  (flagship, GENIAL framework)        │
      │  DR-0..DR-4 · Conformance Criterion C+              │
      │  substrates : MLX kiki-oniric + E-SNN thalamocortical│
      │  Paper 1 v0.2 → arXiv (phase 3) · Zenodo            │
      └──────┬───────────────────────────┬──────────────────┘
             │ cites §7.4                │ ε-trace bridge (gate-dream deferred)
             ▼                           ▼
   ┌──────────────────────┐   ┌──────────────────────┐
   │  nerve-wml  v1.9     │   │  micro-kiki  v0.3    │
   │  GTM tied-stable     │   │  Qwen3.6-35B-A3B     │
   │  N10-C → N10-C-bis   │   │  35 LoRA · Aeon ·    │
   │  Zenodo              │   │  CAMP · KnowBias     │
   └──────────┬───────────┘   └──────────────────────┘
              │ shared codebook (γ/θ)
              ▼
   ┌──────────────────────┐   ┌──────────────────────┐
   │  bouba_sens  v0.5.9  │   │  baby-brain (private)│
   │  null results · TMLR │   │  World-Model Learners│
   │  blocked → TOST+Me6  │   │  → WML-JEPA pilot    │
   └──────────────────────┘   └──────────────────────┘

• Pre-registration before evidence. OSF project dreamOfkiki — DOI 10.17605/OSF.IO/Q6JYN (H1–H4, locked 2026-04-19) plus the Bonferroni amendment ; local pre-regs are frozen by annotated git tag + Software Heritage before OSF filing. Levels L0 draft / L1 tag + SWHID / L2 time-stamped deposit / L3 OSF or Octopus.
• Bit-exact reproducibility (R1). Every experimental claim resolves to a deterministic run_id keyed on (c_version, profile, seed, commit_sha, benchmark_version) ; declared exceptions (MLX conv backward) are stated in the pre-reg.
• Multi-seed first class. Five seeds minimum ; the bouba_sens 5 → 10-seed collapse is the canonical case study.
• Internal critical validation before external review. Null-model controls, bootstrap CIs, multi-estimator robustness checks and an independent critic review ; the Popperian guardrails fire inside the lab, which is why this page lists retractions.
• Dual-axis versioning (DualVer) : the formal-consistency axis (FC) and the empirical-consistency axis (EC) bump independently.
• No AI co-authorship. Byline credits project contributors ; LLM assistance is flagged in PR descriptions ; the researcher decides, runs, validates and signs.
• Compute : a Mac mini M1 (MLX), a shared RTX 4090 node and a small CPU server ; no cloud GPU. The Mac Studio used in spring 2026 is powered down.

Code archives on Zenodo (concept DOIs, latest version resolves) : dream-of-kiki 19701355 · nerve-wml 19656342 · bouba_sens 19701353 · kiki-flow-research 19666755 · micro-kiki 19666759. Preprints : arXiv (Paper 1, phase 3) then HAL.

• Clément Saillant — founder, corresponding author, maintainer. GitHub · L'Électron Rare · clement@saillant.cc Founder of L'Électron Rare — custom electronic systems (design, consulting, training) : electronics, automation, energy. Ten years of field experience, open-source contributor (KiCad, KXKM). Research at Hypneum Lab is conducted alongside this activity, with a disciplined use of AI agents as research tools ; the verifiability of artefacts (pre-registration, calibrated nulls, public errata, reproducibility) is the chosen counterpart of a non-traditional path. Grandris (69), Auvergne-Rhône-Alpes, France.
• Visiting collaborators and co-authors are credited per publication.

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

• Formal proofs for the DR-0..DR-4 family (freeness / universal property of the operation semigroup is explicitly open).
• Conformance Criterion stress tests against alternative substrates (transformer-based, analog neuromorphic, JEPA-class).
• Calibrated nulls and equivalence tests for the bouba_sens invariants.
• 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 flagged in PR descriptions ; byline per publication <repo> project contributors with Clément Saillant as corresponding author. Issues follow the org-wide templates (hypothesis, correction) and carry the a-valider-pi label until the PI validates them. See CONTRIBUTING.md in the flagship repository.

Research / collaboration inquiries : clement@saillant.cc. For L'Électron Rare business (custom electronic systems, consulting, training) see www.lelectronrare.fr.

Page rewritten 2026-10-25 from live GitHub and Zenodo state ; the dated milestone sections it replaces are in docs/profile-journal-2026.md.