Skip to content

Repository files navigation

bouba_sens

A benchmark for cross-modal plasticity in artificial neural systems. Hypneum Lab — 2026.

bouba_sens studies how a 5-modality agent (audio, vision, tactile, gravity, force) reorganises itself when one sensory channel is lost or degraded — inspired by the cross-modal cortical recruitment observed in congenital and late blindness (Amedi 2007, Merabet 2010, Heimler 2020).

Three testable invariants

  • B-1 — Congenital gap: lesion pre-training yields better adaptation than lesion post-convergence.
  • B-2 — MI migration: mutual information between surviving-modality neuroletters and the target label rises post-lesion.
  • B-3 — Perceptive/proprioceptive asymmetry: losing vision/audio/tactile produces a quantitatively different plastic response than losing gravity/force.

2026-05-11 update — Q3+ FINAL Retract, TMLR submission BLOCKED

§5.5 B-1 dose-response signature requires reformulation. The N9 Q3+ 10-seed re-run with the correct test (quadratic regression over LOCK_AFTER) returned sign p = 0.62, paired t-test p = 0.28 — verdict FINAL Retract per the Q3 pre-registered decision criteria. The earlier Q3 INTERIM verdict had used a Jonckheere-Terpstra test which the critic v2 review correctly flagged as the wrong test for a non-monotone (peaked) hypothesis ; fixing that and adding 5 more seeds collapsed the qualitative +0.0125 Sprint 10 peak to noise.

TMLR benchmarks-track submission is BLOCKED until §5.5 is reformulated. Three reformulation drafts (A / B / C, conditional on N12 outcome) are at docs/paper/§5.5-reformulation-draft.md. Decision-of-record : ADR-0019.

N12 targeted subgroup replication (RUNNING). Pooled-v2 analysis surfaced two candidate subgroups for replication — tactile-floor and force-plus10 — at sub-threshold but consistent positive shifts. N12 sweep (N=20 seeds, focused on those subgroups) is running tonight on root@kx6tm-23, ETA ~07:30 morning. The N12 result picks which §5.5 reformulation draft (A / B / C) becomes the paper text.

Multi-seed-first-class lesson promoted portfolio-wide. The Q3 collapse from a 5-seed qualitative claim to a 10-seed null is the canonical evidence for the methodology rule that single-seed (or under-seeded) claims are fragile. Adopted across nerve-wml, dream-of-kiki, and this repo as a hard pre-publication gate.

No new positive finding is announced ; this update is an honest scope statement on what tonight's verdicts removed and what the running sweep may add back.

Axis Status
B-3 Me6 perceptive/proprioceptive asymmetry PASS (5–6× threshold, architectural invariant — unchanged)
B-1 Me7 congenital gap (§5.5) FINAL Retract (Q3+ 10-seed, ADR-0019) — reformulation in flight
B-2 Me3 MI migration Null (estimator-resolved, unchanged from v0.5.5)
TMLR submission BLOCKED until §5.5 reformulation lands
N12 subgroup replication RUNNING (kx6tm-23, ~07:30 ETA)

Cumulative on main HEAD 3fbeee7 : 156 commits, v0.5.9, 19 ADRs (0004–0019), 185+ tests, pooled-v2 analysis + N12 pre-reg locked. Honest scope ; no overselling.


Status

v0.5.5 (2026-04-23) — Paper v0.1 submit-ready. Sprints 0–17 closed, 14 ADRs (0004–0017), 185+ tests. 9 grids (150 cells each) on the 4.5-modal real Studyforrest Phase-2 bridge + 5 worlds synthetic.

  • Paper draft : docs/paper/paper-v0.1-draft.md
  • OSF registration : 10.17605/OSF.IO/Q6JYN, amendment v0.6
  • Target venue : TMLR benchmarks track (primary), NeurIPS D&B (fallback)

Headline findings (v0.5.5)

Thresholds locked by OSF before the campaign : B-1 ≥ 0.05, B-2 ≥ 0.10, B-3 ≥ 0.02 ; no threshold changes across all 14 ADRs.

Invariant Threshold Best grid-median Seed-stability Verdict
B-3 Me6 perceptive/proprio asymmetry 0.02 0.109–0.125 (5.5–6.3×) 5/5 pass PASS every configuration, architectural invariant
B-1 Me7 congenital gap 0.05 +0.0125 (25 % of thr) 3+/5 positive Qualitative — only LOCK=100 + HARD on the real bridge (§5.5 Amedi dose-response) ; N8-Q3 5-seed interim verdict Retract 2026-05-11, see CHANGELOG — TMLR blocked, Q3+ 10-seed pending
B-2 Me3 MI migration 0.10 mean +0.033 ±0.047 bits 0/5 cross threshold Null — two independent estimators (Kraskov + MINE) agree within 0.1 bit

F1 — B-3 is an architectural invariant. The perceptive / proprioceptive asymmetry passes at 5–6× threshold across every grid, every world, every seed, and is insensitive to every plasticity control tested (LOCK_AFTER, transducer gating mode, Gumbel tau, codebook freeze, HARD → GUMBEL phase transition). B-3 is anchored as the empirical foundation of dream-of-kiki/docs/specs/2026-05-20-biophysical-stratification.md §4 (embodied sensorimotor grounding stratum). The B-1 Retract (ADR-0019) is honoured — Stratum 4 explicitly anchors on B-3, never B-1.

F2 — B-1 is qualitatively reproduced in exactly one configuration. On the 4.5-modal real biological bridge, constellation_lock_after=100 with a hard-binary transducer gate produces a seed-stable +0.0125 Me7 (Sprint 10 peak, ADR-0013). Every compound attempted since has preserved, weakened, or destroyed this peak — never amplified it. The paper's central empirical claim is this narrow : a frozen mux with hard transducer gating is the minimal configuration that qualitatively reproduces the Amedi T1/T2 asymmetry.

F3 — B-2 is below threshold, estimator-resolved. The earlier "bimodal positive B-2 at tau=0.30" claim (ADR-0016) was retracted after per-seed re-aggregation (ADR-0017 §14a). Sprint 17 replicates the §6.3 Kraskov limit with MINE (Donsker-Varadhan) on the Sprint 10 peak grid : both estimators agree within 0.1 bit that B-2 does not exceed threshold on any seed. The remaining ambiguity concerns probe shape (1-D mean-pool vs full (B, d_fused)), not estimator choice — declared follow-up for paper v0.2.

Full trail : docs/adr/0004 through 0017. Every verdict, retraction, and scope change is recorded inline with its grid commit.

Quickstart

uv sync --all-extras
uv run bouba-sens version          # bouba_sens 0.5.5
uv run pytest                      # 185+ items, all green
# Single cell
uv run bouba-sens lesion --world gaussian --seed 0 --modality audio \
    --timing T1 --snr-init 0 --snr-floor -20
# Single grid (≈20 min on M3 Ultra, 5-way seed concurrency)
WORLD=studyforrest LOCK_AFTER=100 STEPS_TRAIN=200 STEPS_LESION=100 \
    OUT_ROOT=runs/sprint10_peak METRICS="Me1,Me2,Me3" \
    bash scripts/run_grid.sh
# One-command reproduction of every grid cited in the paper (~8 h)
bash scripts/reproduce_paper_v01.sh
# Verify released artefacts byte-for-byte
uv run python scripts/sha256_manifest.py
diff <(jq -S . reports/MANIFEST.json) <(jq -S . reports/MANIFEST.json.released)

Dependencies

  • Python 3.14
  • PyTorch ≥ 2.5
  • nerve-wml v1.5.3+ (protocol + methodology: Kraskov, MINE, bootstrap CI, null model permutation — Hypneum Lab, DOI 10.5281/zenodo.19666405)

Priority references

  1. Amedi et al. 2007 — Shape conveyed by visual-to-auditory sensory substitution activates LOC (Nat Neurosci).
  2. Heimler & Amedi 2020 — Revisiting adaptive and maladaptive effects of crossmodal plasticity (Neuroscience).
  3. Röder et al. 2021 — Sensitive periods for functional specialization (PNAS).
  4. Alper & Averbuch-Elor 2023 — Kiki or Bouba? Sound symbolism in VLMs (NeurIPS).
  5. Girdhar et al. 2023 — ImageBind: One Embedding Space (CVPR).
  6. Ma et al. 2022 — Are Multimodal Transformers Robust to Missing Modality? (CVPR).
  7. Liang et al. 2022 — MultiBench (NeurIPS D&B).
  8. Keller & Mrsic-Flogel 2018 — Predictive Processing: A Canonical Cortical Computation (Neuron).

License

MIT. See LICENSE.

Citation

See CITATION.cff.

About

Cross-modal plasticity benchmark — Hypneum Lab

Topics

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages