# T2-CHANNEL-001: Noise-free reporting versus transmission through a noisy bottleneck



Version 1.0.0. Date: 2026-11-11.

Verdict: MEAN_PATH_DEPENDENCE. Computational robustness study; no physical validation.

Independent laboratory replication: NONE. One prespecified fit independently rerun in a fresh Python process.



## Result



The predeclared mean-path explanation wins in this setting. Strong reporting depends on noise-free encoder access and does not survive a single sampled bottleneck. This supports a conventional channel-access explanation for the tested signature. It does not show zero information transmission, prove every T2 result is an artifact, or adjudicate O/0 as a metaphysical claim.



## Primary measurements



Nonlinear A3, n=30 independent datasets, beta=128; noise-free report accuracy: 92.82% (95% bootstrap interval 90.92 to 94.62%).

The same models reporting through one noisy sample: 24.92% (95% bootstrap interval 24.56 to 25.26%).

Paired noise-free minus sampled report gap: 67.91 percentage points (95% bootstrap interval 66.01 to 69.71).

Latent-width-matched linear control, noise-free report: 99.18% (95% bootstrap interval 98.23 to 99.91%).

Linear control, single-sample report: 24.93% (95% bootstrap interval 24.69 to 25.18%).



These intervals resample 30 dataset/model pairs, not individual test rows. Balanced four-class chance is 25%.



## Methods and prospective decision



The complete analysis plan and execution sources were locally hash-locked before the first fit. This is a local prospective lock, not independently registered preregistration. Earlier T2 data informed the design.

Thirty fresh cluster geometries each supply disjoint 2,000-row unsupervised model-fit, 1,000-row probe-fit and 2,000-row final-test partitions. Classes are balanced. A3 has hidden width 64 and latent width 4; the linear control also has latent width 4. Both use inherited T2 training code, 400 epochs, beta=128, batch 128, learning rate 0.003 and reconstruction SD 0.3. Six paired datasets at beta=1 are positive controls. Total: 72 fits; no early stopping, model selection or adaptive seed expansion.

Linear ridge is the primary probe; fixed quadratic ridge is a sensitivity analysis. Scaling and probe weights are fitted only on the 1,000-row probe partition. Test labels are used only for final accuracy. Twenty thousand paired bootstrap resamples produce percentile 95% intervals.

The primary mean-path rule requires a noise-free report lower confidence limit above 55%, a sampled-report upper confidence limit below 40%, and a paired gap lower confidence limit above 10 percentage points, after all measurement/regime gates pass. The alternative strong-channel rule requires a sampled-report lower confidence limit above 55%. Otherwise the result is inconclusive.



## Controls and information bounds



A3 mean KL upper bound on I(X;Z): 0.000124 nats; 30/30 runs have the bound below 0.7. Linear control: 0.000053 nats.

This quantity is mean KL(q(z|x)||N(0,I)), not a mutual-information equality. For a uniformly sampled member of the balanced held-out population, I(Y;d(Z)) <= I(X;Z) <= mean KL. A noise-free report d(mu) is a different channel, so the bound does not constrain it.

Derivation: let P be the joint distribution obtained by drawing a balanced test observation and then its Gaussian posterior sample. Let R=P(Y)*N(0,I). By convexity of KL under mixing within each class, KL(P(Y,Z)||R) <= mean_x KL(q(Z|x)||N(0,I)). Under R, any fixed Z-only classifier has 25% accuracy because Y is independent and uniform. Its accuracy under P can increase by at most total variation distance, which Pinsker bounds by sqrt(KL/2). Applying a deterministic decoder does not add information. This is an ordinary mathematical upper bound under the stated Gaussian-channel assumptions, not a novel law about physical observers.

Pinsker's inequality also bounds expected accuracy of any fixed classifier on the single noisy channel by min(1, 0.25+sqrt(mean_KL/2)) for that finite population. A particular finite noise draw can fluctuate; this is not a population confidence bound. The per-fit values are in the raw metrics.

A3 prior-only accuracy: 24.94% (95% bootstrap interval 24.69 to 25.20%); shuffled-training-label report: 25.61% (95% bootstrap interval 20.12 to 31.50%).

Linear prior-only accuracy: 24.94% (95% bootstrap interval 24.69 to 25.20%); shuffled-training-label report: 27.08% (95% bootstrap interval 19.54 to 35.28%).

Beta=1 positive sampled-report controls: A3 99.74% (95% bootstrap interval 99.52 to 99.89%); linear 99.94% (95% bootstrap interval 99.84 to 100.00%).

All predeclared gates:

- fits_complete_finite_duplicate_free: PASS

- A3_collapse_at_least_24: PASS

- A3_tanh_dz4_input: PASS

- A3_tanh_dz4_prior: PASS

- A3_tanh_dz4_shuffled_report: PASS

- A3_tanh_dz4_positive: PASS

- linear_dz4_input: PASS

- linear_dz4_prior: PASS

- linear_dz4_shuffled_report: PASS

- linear_dz4_positive: PASS



## Secondary analyses (descriptive)



Quadratic-probe A3 noise-free report: 96.06% (95% bootstrap interval 94.65 to 97.30%); noisy report: 24.78% (95% bootstrap interval 24.40 to 25.15%).

Quadratic-probe linear-control noise-free report: 99.78% (95% bootstrap interval 99.47 to 99.99%); noisy report: 24.85% (95% bootstrap interval 24.52 to 25.20%).

Original-style strict-signature count with held-out primary probes and KL bound: A3 30/30; linear control 30/30. Replacing noise-free report by sampled report gives A3 0/30 and linear 0/30.

Averaging 16 fresh posterior draws gives A3 report accuracy 25.40% (95% bootstrap interval 24.99 to 25.80%); 256 draws give 30.34% (95% bootstrap interval 29.48 to 31.21%). These conditions receive more observations; they are not evidence of recovery from a single information-limited observation.

Paired architecture differences and both probe families are reported without selecting a winning probe in results/summary.json. No secondary confirmatory claim is made.



## Analytic counterexample and verification



Before the main run, an ordinary Gaussian encoder with mu=0.001*one_hot(class), sigma=1 and linear decoder d(z)=z/0.001 yielded perfectly classified noise-free reports but near-chance noisy reports. Its KL bound is only 0.0000005 nats. Noise-free access can reveal a deterministic tiny signal that is swamped when the actual posterior noise is present. No extra mechanism is needed for this example.

The inherited VAE and linear-model manual gradients passed finite-difference calibration with maximum scaled error 1.61e-9. The locked execution sources did not change during the run.

A separate Python process retrained A3 replicate 0: 32 model, posterior and prediction arrays are bit-identical, and all scientific metrics match exactly. This checks deterministic execution; it is not independent-method or independent-lab replication.



## Relationship to earlier T2 claims



T2 reported d(mu(x)) as reporting and classified z as perception. The current result directly audits that difference in access. The parent study's historical data and thresholds remain preserved. This is a robustness extension with independently generated datasets, held-out probes, train-only scaling, matched latent width and a KL upper bound; it is not an exact repetition of its 9/15 headline count. Changing several evaluation choices prevents attributing every numerical difference to one cause. The within-model d(mu) versus d(z) comparison is the direct channel-access intervention.



## Limits and next experiment



One synthetic family, one primary beta, two architecture classes. Latent width is matched but parameter count is not. Two fixed classifier families do not establish an optimal decoder. Above-chance accuracy is compatible with weak residual information. The study measures neither subjective experience nor a physical source, and it cannot establish or reject the full O/0 framework. Conceptual provenance is not empirical support.

Next priorities: independent implementation of the channel audit; disentangle scaling, evaluation and latent-width changes in a frozen factorial replication; extend to rings with a validated nonlinear probe and matched channel access. A human study would require a separately designed participant protocol.



## Reproduce and inspect



The downloads contain the frozen plan, execution lock, unchanged inherited training modules, execution script, requirements, per-fit parameters, test posterior arrays, predictions, all metrics, control calibration, independent-process verification, run log, figures and SHA-256 checksums. Use README.md for exact commands. Post-run report.py produces this narrative and figure; it was not used to set the prospective rules.

Execution completed 2026-11-11T16:03:58.627812+00:00, main-run duration 302.6 seconds.



## References



- Parent T2 record: 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 evaluation guidance: https://scikit-learn.org/stable/common_pitfalls.html

