# T2-CHANNEL-001: Does reportable information survive the sampled bottleneck?

Version 1.0.0; dated 2026-11-11. Follow-up to T2-NETI-NETI v1.2.0.

This is a locally frozen, prospective analysis plan. The lock JSON records UTC
time and SHA-256 hashes before the first experimental fit. It is not an
independent registry or third-party timestamp. Earlier T2 results informed this
design; none of the new dataset or initialization seeds has been inspected.

## Question and competing explanations

T2 calls d(mu(x)) a report while calling z = mu(x) + sigma(x)*epsilon perception.
The report therefore has noise-free encoder access. Its original strict
signature also uses ridge classifiers fitted and scored on the same labels.
R2/R2a add held-out probes, but retain the original strict-signature metric.

H_CHANNEL_RETAINS: strong class reportability survives when the report actually
passes through the single-sample stochastic bottleneck on unseen observations.

H_MEAN_PATH: noise-free mean access produces substantially stronger reporting
than the stochastic channel; ordinary channel geometry explains the contrast.

Neither hypothesis establishes awareness, mystical experience, source identity,
or the physical truth of O/0. Conceptual provenance is not empirical support.

## Fixed design

- Thirty independent four-cluster datasets in R^8. For replicate i=0..29, seed
  11110000+i generates four random unit directions scaled to radius 3, with
  independent isotropic Gaussian observation noise SD 0.3. Exactly balanced
  classes in each independently drawn partition: 2,000 unsupervised model-fit
  observations, 1,000 supervised probe-fit observations, 2,000 final test
  observations. No partition is used for another role.
- One A3_tanh_dz4 beta-VAE (hidden 64, latent 4) and one linear Gaussian VAE
  with latent 4 on every dataset, paired initialization seed 11120000+i.
  The latter matches latent dimension, not total parameter count.
- Primary beta=128, 400 epochs, batch 128, Adam LR 0.003, reconstruction SD
  0.3, float64, inherited T2 training code copied without changes.
- Sensitivity/positive controls: both models at beta=1 for i=0..5. Total 72
  fits. No architecture search, beta sweep, early stopping, or adaptive seeds.
- Channels: input x; noise-free mu; noise-free d(mu); single posterior draw z;
  d(z); independent N(0,I) prior noise; and means of 16 or 256 posterior draws.
  The last two have more observations and are secondary, not single-sample
  evidence. Sample means use the exact Gaussian law mu+sigma*epsilon/sqrt(m).
- Primary classifier: ridge (sum-squared loss + 1.0*sum(weight^2)), intercept,
  per-column scaling fitted on probe-fit data only. Secondary: fixed quadratic
  ridge with all pairwise products, then scaling. Each probe uses only its own
  channel. No test-label tuning, feature selection, or best-of-probe selection.
- One training-label permutation control on d(mu), and prior-only probes.
  Both are evaluated against unchanged, unseen test labels.
- Save model parameters, test posterior arrays, per-observation predictions,
  per-fit metrics, seeds, runtime environment, source hashes and run log.

## Information measure

Replace the Monte Carlo mutual-information estimate by the exact Gaussian
mean KL(q(z|x)||N(0,I)) on the balanced test population. This is an upper
bound on I(X;Z), not an equality or a new MI estimator. KL<0.7 is a sufficient
collapse condition. For the finite, uniformly sampled test population:

I(Y;d(Z)) <= I(X;Z) <= mean KL.

Pinsker additionally gives expected classification accuracy <=
min(1, 0.25 + sqrt(mean_KL/2)) under fresh posterior noise and a fixed classifier.
The recorded single-noise empirical accuracy is a sample, and may fluctuate
around that bound. No claim of a finite-sample population confidence bound is
made. The bound does not constrain d(mu), which is a different channel.

## Primary analysis and decisions

The unit is an independent dataset plus initialization (n=30), never a test
row or a posterior draw. Use 20,000 paired bootstrap resamples of these units,
seed 11140000, percentile 95% intervals. Report means, intervals, mean paired
gaps, and all individual results. Intervals describe this simulated generator
and training procedure. No p-values or population metaphysical inference.

Apply gates before hypothesis decisions:

1. All fits finite; no duplicate observations across roles; 72/72 fits present.
2. In each primary model family, mean input accuracy >=0.95, and mean prior
   and mean permuted-label report accuracies within [0.20,0.30].
3. In each beta=1 control family, mean sampled-report accuracy >=0.55.
4. At least 24/30 A3 primary runs have mean KL <0.7.

If any gate fails, verdict MEASUREMENT_OR_REGIME_FAILURE; still publish all data.

Using A3 primary linear-probe results only:

- CHANNEL_RETAINS if sampled-report mean accuracy has 95% lower bound >0.55.
- MEAN_PATH_DEPENDENCE if the mean noise-free-report minus sampled-report gap
  has 95% lower bound >0.10, sampled-report mean accuracy has 95% upper bound
  <0.40, and noise-free-report mean accuracy has 95% lower bound >0.55.
- Otherwise INCONCLUSIVE.

These rules do not declare the channel information-free. Above-chance weak
transmission may coexist with failure of the strong reportability criterion.

Secondary: report both classifiers separately; compare A3 with matched-latent
linear control via paired differences; report 16/256-draw recovery; count the
original-style strict signature and its channel-respecting version using
held-out accuracy and KL<0.7. These are descriptive, with no additional
confirmatory declarations. Original-style: mean report >0.55, z accuracy
<0.40, mu accuracy >0.55. Channel-respecting: replace mean report by d(z).

## Calibration and reproducibility

Before real fits, finite-difference-check inherited gradients and test an
analytic conventional channel with mu=0.001*one_hot(class), sigma=1 and
d(z)=z/0.001. Its noise-free report is perfectly classifiable while its
stochastic report is near chance and KL=0.0000005 nats. This demonstrates
the signature's reachability without a special mechanism.

Lock all execution sources and this plan after calibration, before main fits.
Re-run one prespecified primary fit (A3, replicate 0) in a separate Python
process and require identical model arrays, posterior arrays and predictions.
Do not alter locked sources during the run. Post-run report/figure scripts
are separately identified. Preserve failed or interrupted outputs. Any
deviation must be dated and reported; do not silently revise thresholds.

## Limitations and next steps

Single synthetic data family, two architecture classes, one primary beta;
quadratic probes are not universal decoders. The linear control matches
latent width but not parameter count or nonlinear capacity. Changes relative
to T2 include balanced fresh datasets, standardization, ridge penalty,
held-out evaluation and KL bound, so this is a robustness extension, not an
exact replication. Independent laboratory replication remains necessary.
A human phenomenology experiment would need separate consent, operational
definitions and an appropriate participant protocol; none is run here.

## References

- T2-NETI-NETI v1.2.0: https://ozonearchive.online/science/T2-NETI-NETI.html
- Alemi et al., Deep Variational Information Bottleneck, ICLR 2017:
  https://arxiv.org/abs/1612.00410
- Doersch, Tutorial on Variational Autoencoders:
  https://arxiv.org/abs/1606.05908
- scikit-learn, Common pitfalls (independent test sets; train-only scaling):
  https://scikit-learn.org/stable/common_pitfalls.html
