Refractory virtualization

When the carrier changes but the visual content does not

Applies to: dpsh

Neurons cannot fire continuously. After a spike, refractoriness temporarily makes a neuron unavailable. In a recurrent network, later processing may therefore have to take a different route even when it must preserve the same percept or decision.

This research asked whether that rerouting can be more than ordinary backup redundancy. The operational proposal was called refractory virtualization: a functional state is carried by a time-varying equivalence class of causal realizations rather than by one permanently available route.

The final experiment produced a bounded positive result. Across 40 reserved seeds, refractoriness switched the causal route while a spatial colour percept remained unchanged. The result establishes neither consciousness nor qualia. It establishes a narrower mechanism—and documents the many attractive but incorrect explanations that had to be removed first.

The hypothesis and its falsifier

The functional hypothesis predicts all of the following at once:

  1. a simple route works while it is available and loses causal influence when refractoriness blocks it;
  2. an alternative route gains causal influence in that refractory condition;
  3. the represented visual content remains distinguishable and supports the same delayed decision;
  4. a feed-forward null does not show the same effect;
  5. a structurally different alternative route remains measurably different from an isomorphic duplicate;
  6. matched perturbations change continuous observations across seeds without changing the preregistered conclusion.

The claim would be rejected if the route did not change, if visual content changed with it, if added activity or a feed-forward path explained the result, or if the apparent replication consisted only of identical runs carrying different seed labels.

Why vision was the test case

The study deliberately converged on spatial colour rather than treating all sensory channels as interchangeable. Colour supplies a strong structured target: it depends on nonlinear channel ratios and on spatial context. A single isolated activation is not an adequate model of a visual percept.

The visual apparatus established four prerequisites:

  • nonlinear receptor mixing and opponent normalization produced distinct colour identities;
  • neighbourhood integration suppressed isolated disagreement while preserving real boundaries;
  • four colours remained distinguishable at multiple grid locations and after translation;
  • swapping two colours moved the local identities even though the global class histogram—and in one test the carrier spike raster—remained unchanged.

That last control produced an important conceptual result: the carrier is not the content. The same temporal spike skeleton can accompany different structured values in continuous network state. Comparing spike counts or rasters alone is therefore insufficient evidence that two percepts are equal.

Repository evidence: Docs/refractory-virtualization/results/TRV-C1-spatial-colour-percept-v3-2026-10-25.md and Docs/refractory-virtualization/results/TRV-C3-multicolour-spatial-identity-v1-2026-10-25.md.

What failed—and what each failure taught us

The positive result was not obtained by repeatedly tuning one model until it passed. Each failed branch closed a specific explanation or exposed an invalid instrument.

A local trace did not discover its own return

The earlier TIRD-R line tested passive traces, prediction error, locally tagged perturbations and randomized spike lock-in. Even after the intervention became measurably active, the true code did not beat shifted or permuted controls. The local signal mixed return, refractoriness, spontaneous activity and the consequences of previous probes. That family was stopped rather than promoted.

This rejected a local return-trace mechanism; it did not test or reject route switching with preserved content.

Repository evidence: Docs/Tird-theory/results/T7-randomized-spike-lockin-development-v2-2026-10-24.md.

Spatial colour was not a pointwise threshold

Pointwise nonlinear gates, a median rule, independent pixel detectors, hard neighbour voting, linear spatial mixing and narrow bilateral kernels each failed for different reasons. Some erased edges, some blurred weaker colours, and some amplified the background. The surviving apparatus needed both nonlinear colour formation and an explicitly spatial relation among neighbours.

Persistence was not yet a memory

Delayed colour decisions exposed a formation limit followed by decay. Excitatory recurrence caused ignition; balanced biphasic recurrence suppressed ignition but did not create memory. Figure-ground, neighbour-support, centre-surround and coincidence layers often changed activity without preserving colour-specific recall. Read-only probes confirmed that several intermediate layers simply did not contain a hidden usable colour code.

These negative results separate “the network remains active” from “the network preserves this content.”

More copies helped speed, but did not explain virtualization

Increasing independent routes from four to eight reduced the shortest passing rate-readout window from 1,280 ms to 640 ms. Sixteen routes did not improve it further. No tested configuration passed in the biologically interesting 80 ms window, although accuracy increased with population size.

Redundancy can trade neurons for sampling time. It did not by itself establish an alternative causal realization, and it did not explain fast biological vision.

Repository evidence: Docs/refractory-virtualization/results/TRV-C14-population-redundancy-speed-v1-result-2026-10-25.md.

Early timing carried information, but not a stable solution

First-spike latency sometimes outperformed rate in short windows. Mean onset, median onset, onset order, multi-seed calibration and survival curves did not produce a robust colour recall across all controls. The experiments ruled out the convenient assumption that one simple early-time statistic was the missing code.

Correct normalization was necessary but not sufficient

Analytic divisive normalization, c_i / (epsilon + sum(c)), exactly recovered colour under spatially varying common illumination, while absolute channels failed badly. A fixed subtractive or shunting approximation was not equivalent. Placing the correct ratio before spike encoding still did not guarantee an 80 ms downstream readout, because the encoder's operating point and sampling remained limiting.

Repository evidence: Docs/refractory-virtualization/results/TRV-C26-divisive-normalization-known-truth-v1-result-2026-10-25.md.

Several apparent causal results were instrument errors

Early path probes observed outside the causal light cone. Later controls were contaminated by autonomous baseline activity or by the accumulated effects of previous probes. A valid observer had to fork active and sham branches from the same pre-probe snapshot and measure a difference of differences.

This work also found and repaired runtime defects: production WindowEngine did not originally evaluate deterministic threshold firing after synaptic delivery, and causal fan-out events were missing from the observer trace. The experiments improved the engine even when their scientific hypothesis failed.

Repository evidence: Docs/refractory-virtualization/results/TRV-C20-coincidence-neighbour-layer-v4-result-2026-10-25.md, Docs/refractory-virtualization/results/TRV-C31-causally-reachable-path-probe-v1-result-2026-10-25.md and Docs/refractory-virtualization/results/TRV-C34-evoked-path-difference-of-differences-v1-result-2026-10-25.md.

Forty labels were not forty replicates

TRV-V1 consumed 40 calibration seeds, but every observed matrix was identical. The deterministic network had no seed-dependent measured quantity, so a bootstrap over those copies would have been pseudoreplication. V1 was declared invalid and its development and confirmation partitions remained closed.

V1 nevertheless revealed the correct contrast: not “all non-feed-forward routes have positive influence,” but an interaction between architecture and refractoriness. The simple route should lose influence while duplicate and non-isomorphic alternatives gain it.

Repository evidence: Docs/refractory-virtualization/results/TRV-V1-family-wise-calibration-v1-2026-10-25.md.

The final TRV-V2 experiment

TRV-V2 froze four architecture families:

  • SP — a single path;
  • DP — duplicate isomorphic paths;
  • NP — non-isomorphic alternative paths;
  • FF — a feed-forward null.

Each was measured without the refractory switch (R0) and with a 40 ms refractory condition (R1) across five visual scenarios. Matched perturbation draws were shared by counterfactual branches. Path influence came from actual ablations; route identity combined hop count, latency and a causal lag profile; and the colour readout corrected only the input jitter that the protocol itself introduced.

The run used a clean build from commit 4194f3e16334, MSVC 19.44.35228.0, design hash 13987969641573228549 and the reserved seeds 42001..42040. Every seed was written atomically, then the complete partition was reloaded and validated before summarization.

Result

All 40 reserved seeds produced the same preregistered primary contrasts:

Measurement Result
SP causal-coverage loss from R0 to R1 1.0
DP causal-coverage gain from R0 to R1 1.0
NP causal-coverage gain from R0 to R1 1.0
maximum FF causal coverage 0.0
minimum NP route-signature distance 0.5
minimum percept accuracy 1.0
maximum percept drift 0.0

The simple path carried the effect before the refractory switch and lost it afterwards. Both the duplicate and non-isomorphic alternatives gained causal influence under refractoriness. The feed-forward control remained zero. The decoded spatial colour percept and delayed decision remained unchanged.

The continuous ablation-impact diagnostic varied across seeds, producing 40 distinct observation hashes. Over 40 seeds and five visual scenarios, DP impact at 40 ms ranged from 14.67 to 15.92 (mean 15.23); NP ranged from 10.75 to 11.91 (mean 11.27). SP, FF and all R0 diagnostic impacts were zero.

Evidence status: ACCEPTED WITHIN THE PREREGISTERED SYNTHETIC DPSH SCOPE.

Repository evidence: Docs/refractory-virtualization/preregistration/TRV-V2-route-switch-and-robustness-v1.md and Docs/refractory-virtualization/results/TRV-V2-route-switch-reserved-v1-2026-10-26.md.

What was discovered

  1. A stable visual content can change causal carrier. The measured output did not depend on one permanently available route.
  2. Refractoriness can control that switch. The architecture-by-refractory interaction, rather than route count alone, was the decisive signal.
  3. Redundancy and virtualization are different measurements. More copies can improve sampling speed, but virtualized realization requires observed transfer of causal influence while content remains fixed.
  4. Structural difference and causal strength are separate axes. NP was structurally different (distance=0.5) yet had lower ablation impact than isomorphic DP. “Deeper” representation cannot be identified with either route diversity or effect magnitude alone.
  5. Content cannot be inferred from the spike carrier alone. Spatial state and counterfactual consequences must be measured alongside event timing.
  6. Negative experiments were productive. They identified invalid observers, pseudoreplication, unsuitable colour operators, unstable memory mechanisms and two production-runtime defects.

What this result does not show

  • It does not show that the simulated system is conscious.
  • It does not identify a neural correlate of subjective colour experience.
  • It does not show that every loop, redundant network or refractory neuron virtualizes a state.
  • It does not prove that non-isomorphic paths are better than duplicates; in this apparatus their measured causal impact was lower.
  • It does not explain human-scale visual latency. The earlier rate apparatus remained much slower than biological vision.
  • It does not yet establish a general counterfactual state Q(x) containing rich predictions, transitions and responses across natural inputs.

The phenomenal proposal remains a hypothesis: subjective experience might be the intrinsic side of such a dynamically realized structure. A simulation can motivate that idea, but cannot confirm it. Biological experiments would have to separate perception from input, attention, report, memory and arousal while selectively disrupting the proposed organization.

Reproduction

The public command descriptor, parser, text help, JSON help and generated CLI reference share one source of truth. A clean committed build is required before reserved seeds can run.

.\make.ps1 target cognia_trv_v2
.\cmake-build-debug\cognia_trv_v2.exe deterministic

foreach ($seed in 42001..42040) {
  .\cmake-build-debug\cognia_trv_v2.exe calibrate temp\trv-v2-route-switch-v2-4194f3e --seed $seed
}

.\cmake-build-debug\cognia_trv_v2.exe summarize temp\trv-v2-route-switch-v2-4194f3e

CLI reference: Docs/agents/reference/cli/cognia_trv_v2.md.

The next falsifiable question

TRV-V2 establishes route switching with preserved visual output. It does not establish a scalar “depth of percept.” The next experiment should independently vary route diversity, causal impact and counterfactual richness while matching performance, activity and capacity. If one of those variables predicts generalization, completion or robustness beyond the controls, it may become a component of representational depth. If not, the proposed connection should be rejected rather than folded into the successful TRV-V2 claim.