Kahlus Neural-CASP general plan

Core update

You’re right. I narrowed too fast into Criticality Residual as an implementation bet. That was useful, but the research question should stay wider for one more layer.

Here is the stronger frontier framing:

Kahlus should search for hidden information channels and invariant transition structure in brain-body dynamics, then force each candidate through Neural-CASP gates.

Not:

bigger EEG model.

Not:

seizure wearable.

Not:

generic multimodal fusion.

The big question:

Are there conserved, recoverable state variables across brain, body, and modality that make neural transitions forecastable beyond nuisance baselines?

That is the right “weird but evidence-gated” target.

---

0. Current repo state matters

PR #30 already gives you a real scaffold: Forecastability Trial 0, M0-M3 gates, synthetic known/null tests, Sleep-EDF machinery, CHB-MIT smoke, TUSZ external path, baseline ladders, negative controls, and claim boundaries. It correctly labels M3 as failed/underpowered and explicitly blocks clinical claims.

So the next step is not “invent framework.”

The next step is:

Use this courthouse to try stranger defendants.

---

1. Candidate moonshots, updated

Candidate	Upside	Evidence readiness	Data readiness	Killability	My confidence
Universal criticality / transition order parameter	Very high	Medium	High	High	Medium
Gut-brain information channel, BCI↔GCI	High/weird	Low-medium	Low-medium	Medium	Low-medium
Predictive consciousness complexity index	High	Medium	Medium-high	High	Medium
Provable nuisance-invariant SSL objective	High CS value	Medium	High	High	Medium-high
Cross-modal neural translation law	Highest theory upside	Low	Low-medium	Medium	Low
Neural-CASP benchmark itself	Field-shaping	High	High	High	High

Main change:

The most revolutionary scientific bet is not one model. It is a set of weird candidate laws tested inside one unforgiving benchmark.

---

2. Candidate A: Universal cross-modal neural-translation law

Hypothesis

There exists a latent neural state (Z_t) recoverable from partial observations across modalities, such that EEG, MEG, fMRI, fNIRS, behavior, and physiology are noisy observation operators of the same state.

Formal version:

[ X_t^{(m)} = g_m(Z_t, \eta_t^{(m)}) ]

where (m) is modality.

Goal:

[ X^{(a)}{\le t} \rightarrow Z_t \rightarrow X^{(b)}{t:t+h} ]

Why this is attractive

This is closest to “AlphaFold-like” in shape:

partial observation → hidden object → held-out reconstruction

AlphaFold worked because it had a hard blind benchmark, strong priors, confidence estimates, and generalization. It achieved median domain GDT_TS 92.4 at CASP14, a field-legible win 1, and the Nature paper frames it as solving a 50+ year protein-structure prediction problem with near-experimental accuracy in many cases 2.

Strongest disconfirming evidence

This may be mathematically underidentified.

Without strong assumptions, many latent states (Z) can explain the same observations. For any invertible transform (T):

[ Z’_t = T(Z_t) ]

can be equally valid if observation operators adapt:

[ g’_m = g_m \circ T^{-1} ]

So “recover the true latent brain state” is not identifiable. At best, you recover a predictive equivalence class.

Also, EEG foundation-model work is already warning that bigger models do not reliably generalize better, specialist baselines remain competitive, and benchmark heterogeneity blocks clean claims 3. EEG partitioning can inflate performance when subject/session splits are wrong [4].

What would make it real

Not “EEG predicts fMRI.”

Real result:

A model trained on modality A predicts held-out modality B on unseen subjects/sites/datasets better than all nuisance, persistence, and specialist baselines, with calibrated uncertainty and negative controls failing.

Verdict

Do not start here.

This is too broad and too easy to fake. Use it as the long-term theory umbrella, not first moonshot.

Confidence: Low.

---

3. Candidate B: Gut-brain information channel, BCI↔GCI

Hypothesis

Gastric electrophysiology contains state information about cognition/sleep/neural dynamics that is not contained in EEG/heart/respiration alone.

Formal target:

[ I(Y_{\text{future}}; EGG_{\le t} \mid EEG_{\le t}, HRV_{\le t}, Resp_{\le t}, B_t) > 0 ]

Practical metric:

[ RFS_{\text{gut}} = \frac{NLL(EEG + nuisance) - NLL(EEG + EGG + nuisance)}{\log 2} ]

Evidence for

This is real enough to investigate.

A simultaneous EEG+EGG study in 14 young adults found EGG-EEG phase-amplitude coupling changed with satiety and cognitive state, and coupling strength correlated with working-memory behavior [5]. A 2024 method paper proposes a mutual-information PAC estimator and demonstrates its use on simultaneous gut-brain datasets [6]. A 2025 bioRxiv preprint reports simultaneous high-density EEG+EGG across 60 participants and three nights, finding gastric infraslow dynamics aligned with cortical sigma power during sleep and predicted subjective sleep quality beyond PSG/cardiac measures [7].

Why this is weird/high-upside

If true, it expands the neural forecasting problem:

brain-only state → brain-body state

That is more original than another EEG model. It also fits a frugal hardware future: EEG + EGG + ECG are scalable sensors.

Strongest disconfirming evidence

Data is the bottleneck.

The best simultaneous EEG+EGG evidence is small, early, or not fully established as public benchmark infrastructure. The 60-participant sleep result is a preprint, not settled consensus [7]. Gut-brain coupling may reflect common autonomic/arousal confounds rather than directed information.

Major nuisance variables:

respiration
heart rate / HRV
posture
meal timing
satiety
motion
sleep stage
circadian timing
skin/electrode artifacts

If EGG loses after conditioning on these, BCI↔GCI collapses.

Best experiment

Interoceptive Residual Forecastability Track

Question:

Does EGG improve sleep-stage transition prediction or subjective sleep quality prediction beyond EEG+ECG+respiration+nuisance baselines?

Use DML-RPSM:

baseline: sleep stage + EEG + HRV + respiration + time
residual: EGG features / directed coupling / PAC / transfer entropy
endpoint: RFS_gut > 0

Verdict

Most original “weird” track.

Not first if public data is unavailable. First scout task: find/download real EEG+EGG dataset or reproduce from public available materials.

Confidence: Low-medium, but high novelty.

---

4. Candidate C: predictive consciousness complexity measure

Hypothesis

A theory-neutral complexity/state-space measure predicts held-out level-of-consciousness labels across wake, sleep, anesthesia, seizure, and disorders-of-consciousness data better than existing complexity measures.

Do not say “consciousness solved.”

Say:

predictive complexity index for brain-state labels.

Evidence for

Criticality and complexity measures are plausible across altered states. A scoping review of criticality and altered states included sleep, anesthesia, epilepsy, psychedelics, and disorders of consciousness, and found preliminary evidence of deviations from criticality across altered states, although heterogeneous and often directionally unclear [8].

There is also established conceptual precedent from PCI and Lempel-Ziv-style complexity, though TMS-EEG PCI requires perturbation, not passive EEG. Passive EEG complexity is easier but weaker.

Strongest disconfirming evidence

This track is claim-dangerous.

It can become:

sleep/anesthesia classifier wearing consciousness language

Also, many measures track arousal, spectral power, muscle artifact, signal amplitude, or drug protocol rather than consciousness itself.

Best experiment

Predictive Complexity Benchmark

Train no giant model. Compare scalar/vector indices:

Lempel-Ziv complexity
permutation entropy
multiscale entropy
spectral slope
metastability
criticality distance
state-space occupancy

Task:

held-out subject classification:
wake vs NREM vs REM vs anesthesia depth/emergence

Primary endpoint:

RFS / NLL improvement over spectral-power and sleep-stage baselines

Gate:

measure must transfer across datasets and not collapse under artifact controls

Verdict

Good benchmark track. Bad branding track.

Use if framed as:

state-predictive complexity, not consciousness.

Confidence: Medium.

---

5. Candidate D: brain-as-critical-system order parameter

Hypothesis

Approaching neural transitions share measurable loss-of-resilience features: rising autocorrelation, variance, spectral slowing, entropy shifts, or distance-to-criticality changes.

This is still the best “test now” moonshot.

Evidence for

Maturana et al. found critical-slowing signatures in long-term intracranial EEG from 14 focal-epilepsy patients, with seizure risk associated with those signatures plus epileptiform discharges [9]. A 2023 altered-states criticality review found preliminary evidence that NREM sleep, epileptic seizures, anesthesia, and psychedelic states deviate from criticality, though methods are heterogeneous [8].

Strongest disconfirming evidence

Criticality is not a magic universal law.

Dablander et al. warn that early-warning signals are not truly model-agnostic, depend strongly on system specifics, and are noise-sensitive [10]. Helmich et al. argue evidence is weak for using critical-slowing signals to forecast psychological/clinical symptom transitions, and emphasize that real clinical prediction must specify both whether and when transitions occur [11].

Mechanism problem:

seizure = excitability / epileptogenic networks
sleep = thalamocortical + homeostatic/circadian systems
anesthesia = pharmacologic perturbation

Shared signatures may be generic math folklore, not biology.

Best experiment

Universal Criticality Residual

Primary metric:

[ \Delta RFS_{\text{crit}} = RFS(B + standard\ EEG + criticality)

RFS(B + standard\ EEG) ]

Win condition:

ΔRFS_crit > 0 with CI excluding 0
on at least two domains
under held-out subject/dataset splits
controls collapse

Domains:

Sleep-EDF first
CHB-MIT/TUSZ second
anesthesia third

Verdict

Best immediate moonshot.

It is weird enough, testable now, and falsifiable.

Confidence: Medium.

---

6. Candidate E: provable nuisance-invariant self-supervised objective

Hypothesis

Build a representation (Z=f(X)) that provably discards nuisance variables (N) while retaining target-relevant information (Y).

Desired:

[ I(Z;N \mid Y) \le \epsilon ]

and:

[ I(Z;Y) \text{ remains high} ]

Why it matters

This would turn “leakage-proof” from a checklist into a theorem.

Current PR #30 makes claim gates executable, but it is still evaluation-first: split manifests, baselines, controls, and reports.

A model-level guarantee would be a real CS contribution.

Evidence base

DML gives one rigorous route: cross-fitting, nuisance estimation, and orthogonalization reduce sensitivity to nuisance error. DoubleML’s framework explicitly emphasizes Neyman orthogonality, high-quality nuisance estimation, and sample splitting [12].

Invariant representation learning is attractive but fragile. IRM was proposed to learn invariant predictors across environments, but later work shows IRM can fail to capture natural invariances or fail catastrophically under certain settings. (arXiv) (arXiv) (arXiv)

Strongest disconfirming evidence

“Probe at chance” does not prove no nuisance information.

It only proves a chosen probe class cannot extract nuisance.

A stronger statement needs:

probe class specification
finite-sample bounds
held-out nuisance labels
worst-case adversary or information bound

Otherwise, you get false comfort.

Best experiment

Certified Leak Score

For every representation (Z), report:

nuisance probe performance
mutual-information upper-bound estimate
domain adversary strength
target performance
RFS

Define:

[ CLS = RFS_Y - \lambda \cdot LeakScore_N ]

Do not use it as the primary science metric yet. Use it as a representation audit.

Verdict

Important infrastructure. Not the standalone biological revolution.

Confidence: Medium-high for usefulness, low for proving true nuisance-free representation.

---

7. Candidate F: Neural-CASP benchmark

Hypothesis

The field’s biggest missing object is a blind, leakage-proof benchmark that defines what counts as evidence for neural forecasting/translation.

This is the safest field-shaping move.

Evidence for

EEG foundation models are fragmented: inconsistent preprocessing, objectives, and evaluation protocols; specialist baselines remain competitive; larger models do not reliably generalize 3. EEG data partitioning can change results drastically and create leakage/overestimated performance [4].

This is exactly where PR #30 already points: evidence artifacts, split manifests, baseline ladders, gates, and failed-underpowered M3 rather than overclaiming.

Strongest disconfirming evidence

Benchmarks alone do not become revolutionary unless a model or result demonstrates that the benchmark matters.

CASP mattered more when AlphaFold won it. ImageNet mattered because models climbed it.

A courthouse with only convictions gets ignored.

So Neural-CASP needs:

one positive F3 result
or one clean negative result that changes consensus

Verdict

Most realistic field-shaping contribution.

Confidence: High.

---

8. The new synthesis

The strongest research program is not one of these alone.

It is:

Neural-CASP benchmark
+ Universal Criticality Residual
+ Interoceptive Gut-Brain scout
+ Predictive Complexity index
+ DML / nuisance-invariance audit

But staged.

---

9. Updated Kahlus moonshot ladder

Track 0: Neural-CASP Forecastability

Purpose:

scientific standard of proof

Status:

already started by PR #30

Next:

finish powered M2/M3, external validation, DML-RPSM V0

Track A: Universal Criticality Residual

Purpose:

test shared transition signature

Best first moonshot.

Track B: BCI↔GCI Interoceptive Channel

Purpose:

test whether stomach/gut carries residual state information beyond EEG

Highest originality, but data bottleneck.

Track C: Predictive Complexity Index

Purpose:

benchmark passive complexity measures across wake/sleep/anesthesia

Good if carefully de-consciousnessed.

Track D: Certified Nuisance-Invariant Representation

Purpose:

CS/theory upgrade

Important support layer.

Track E: Cross-modal Translation Law

Purpose:

long-term theory

Not first.

---

10. What I would research/build next

Next research question

Do criticality and interoceptive features add residual forecastability beyond standard EEG and nuisance baselines across sleep and seizure transition tasks?

This combines the two weirdest credible bets:

criticality = universal transition dynamics
gut/EGG = hidden brain-body information channel

But split into two experiments.

---

11. Experiment 1: Criticality Residual

Use current repo.

Add features:

lag-1 autocorrelation
variance
coefficient of variation
spectral exponent / 1/f slope
multiscale entropy
approximate entropy
Hurst / DFA exponent
recovery-time proxy
change-point proximity
criticality composite

Ablations:

B only
B + standard EEG
B + criticality
B + standard EEG + criticality
B + shuffled criticality
B + time-shifted criticality

Primary endpoint:

[ \Delta RFS_{\text{crit}}

RFS(B+EEG+C)-RFS(B+EEG) ]

Claim only if:

CI excludes 0
control margin CI excludes 0
held-out subject/domain

---

12. Experiment 2: BCI↔GCI scout

Only if data exists.

Search target:

simultaneous EEG + EGG
sleep or cognitive task
raw accessible
subject labels
state labels
ECG/respiration ideally

Primary endpoint:

[ \Delta RFS_{\text{EGG}}

RFS(B+EEG+EGG)-RFS(B+EEG) ]

Controls:

phase-shifted EGG
subject-shuffled EGG
time-shifted EGG
respiration/HRV-conditioned baseline
meal/satiety/time controls

Claim:

EGG carries residual state information.

Not:

gut controls consciousness.

---

13. Experiment 3: Predictive complexity index

Use:

Sleep-EDF
PhysioNet GABA anesthesia if accessible
seizure datasets

Features:

LZ complexity
permutation entropy
multiscale entropy
metastability
spectral slope
criticality distance

Primary endpoint:

held-out NLL / RFS for state labels or transitions
over spectral-power baseline

Claim:

complexity features predict state labels under held-out validation.

Not:

consciousness solved.

---

14. Hard theorem-style insight

The universal law you want is not:

every transition shares the same latent state.

That is likely false.

The better law is:

Forecastability decomposes into baseline-cycle information, subject-specific information, sensor-specific residual information, and cross-domain transition information.

Formally:

[ I(Y;X,B,S,D)

I(Y;B) + I(Y;S \mid B) + I(Y;X \mid B,S) + I(Y;D \mid X,B,S) ]

Where:

Y = future transition
B = nuisance baseline / cycles / history
S = subject / site / patient structure
X = sensor stream
D = domain / transition type

Kahlus estimates pieces by held-out NLL gains:

baseline forecastability
subject-specific forecastability
residual neural forecastability
cross-domain forecastability

This is more credible than “brain transition law.”

It gives you a mathematical atlas.

---

15. The actual “revolutionary” claim if it works

Not:

Kahlus predicts brain states.

The better claim:

Brain-state predictability can be decomposed into measurable information channels, and many claimed neural forecasts vanish once nuisance, subject identity, and cycle information are removed.

If positive:

Some transition features survive this decomposition across domains, suggesting conserved loss-of-resilience dynamics.

That is field-changing.

---

16. Honest recommendation

Run this sequence:

1. Finish Neural-CASP Forecastability Track from PR #30.
2. Add Criticality Residual as Track A.
3. Scout EEG+EGG public data for BCI↔GCI.
4. Add Predictive Complexity Track only after sleep/anesthesia data is wired.
5. Leave cross-modal translation law as long-term theory.
6. Treat nuisance-invariance theorem as infrastructure, not science headline.

Most promising weird bet now:

Criticality Residual.

Most original dark-horse:

Gut-brain information channel.

Most realistic field-shaper:

Neural-CASP benchmark.

Highest-theory but lowest readiness:

Cross-modal neural translation law.

---

Bottom line

You wanted “higher and weirder.” Here it is:

Kahlus should become an information atlas of brain-state forecastability, testing whether neural transitions share residual criticality signals and whether non-brain physiology like gastric rhythm carries hidden state information.

That is way more interesting than ordinary neural prediction.

Still evidence-gated. Still killable. Still no fake throne.

---

Key references

1 Applying and improving AlphaFold at CASP14 — Jumper et al., 2021, Proteins, citation count: 348. 2 Highly accurate protein structure prediction with AlphaFold — Jumper et al., 2021, Nature, citation count: 36679. 3 EEG Foundation Models: Progresses, Benchmarking, and Open Problems — Liu et al., 2026, ArXiv, citation count: 8. [4] The role of data partitioning on EEG deep learning performance — Del Pup et al., 2025, Computers in Biology and Medicine, citation count: 24. [5] Critical slowing down as a biomarker for seizure susceptibility — Maturana et al., 2020, Nature Communications, citation count: 216. [6] A scoping review for building a criticality-based conceptual framework of altered states of consciousness — Gervais et al., 2023, Frontiers in Systems Neuroscience, citation count: 22. [7] Anticipating critical transitions in psychological systems — Dablander et al., 2020, Psychological Methods, citation count: 53. [8] Slow down and be critical before using early warning signals in psychopathology — Helmich et al., 2024, Nature Reviews Psychology, citation count: 25. [9] Simultaneous Gut-Brain Electrophysiology Shows Cognition and Satiety Specific Coupling — Balasubramani et al., 2022, Sensors, citation count: 19. [10] A mutual information measure of phase-amplitude coupling using gamma generalized linear models — Perley & Coleman, 2024, Frontiers in Computational Neuroscience, citation count: 4. [11] Simultaneous stomach-brain electrophysiology reveals dynamic coupling in human sleep — Rao et al., 2025, bioRxiv, citation count: 2. [12] DoubleML — Bach et al., 2021, ArXiv, citation count: 62.