Gap 4 — the access window and the clinical conscious/unconscious split

Gap 4 asks what makes some cognition reportable. Driving the frozen kernel through the emerged sleep states, a faithful perturbational complexity index reproduces the clinical conscious/unconscious split with no new constant — wake ≈ REM ≫ NREM, sign-stable and non-circular. The single scalar does not rank the unconscious states; that is recorded as an honest negative.

Consciousness is read as the subset of cognition reaching global coordination (high PCI); intuition is the remainder. Gap 4 drives the frozen M9 kernel through the emerged M14 sleep architecture and scores a faithful perturbational complexity index — the normalised Lempel-Ziv complexity of the perturbed-minus-unperturbed causal divergence. With no kick the divergence is identically zero, so PCI = 0 for every state (the non-circular control). The conscious/unconscious split is clean: wake ≈ REM (0.464) ≫ NREM (0.036), across the clinical cutoff PCI* ≈ 0.31, sign-stable across 9 cells. The single scalar does not resolve the fine unconscious ordering (NREM/anaesthesia/VS collapse to one value, spread 0.045) — the preserved honest negative.

The gap: what makes cognition reportable

The access-window hypothesis is that consciousness is the subset of cognition that reaches global EM coordination (high PCI), and intuition is the remainder (cognition far exceeds access). Gap 4 operationalizes this with the clinical bedside measure of consciousness — the perturbational complexity index — driven through real sleep states, and asks whether the model reproduces the clinical pattern.

The mechanism is grounded by M14, not by us

The state→mechanism map is set by the emerged sleep architecture. emerge_sleep_architecture() already emerged (from the frozen kernel plus cited kinetics) the activated-versus-bistable distinction the module rides on: shift_index_rem ≈ 1.05 (REM is activated/desynchronised, no OFF-periods) versus shift_index_nrem ≈ 0.045 (NREM is bistable, slow Up/Down, OFF-periods present), with spindle 15.4 Hz and slow-oscillation 0.49 Hz matched. Those facts are asserted at run start — the module aborts if M14 drifts.

The scored mechanism is the M14 construction: a 16-cell cortical population of R19 cubic-bistable cells (s − s³ + drive − a) coupled diffusively through the M9 ephaptic ring at the measured κ = 0.5496. The state is set only by the slow-adaptation strength (off_depth × A_GAIN) and the cited carrier — activated theta-gamma (f ≈ 23.5 Hz, atlas f0) for wake/REM, slow 0.5 Hz for NREM. Every quantity is measured (κ; A_GAIN = 1.6; TAU_SO = 0.600 s, Sanchez-Vives & McCormick 2000; TAU_S = 0.025 s; drive = 0.60, Steriade 1993); new_tuned_constants = 0; the engine is read-only.

The faithful, non-circular PCI

A single TMS-like kick (+2.0 to one cell, 4 ms after a 4 s settle) is applied to a perturbed twin; an unperturbed twin shares identical initial conditions. PCI is the normalised Lempel-Ziv complexity of the binarised perturbed-minus-unperturbed causal divergence over a 1.5 s post-pulse window (downsample 8×, binarise D > D.mean()). Subtracting the twin removes spontaneous activity, so with no kick the divergence is identically zero — PCI = 0 for every state, the exact property a spontaneous-EEG complexity measure lacks and the clinical PCI is built to have. This read-out survived four prototype iterations: a spatial-median-of-raw-response read-out was rejected because its no-perturbation control gave Δ = +0.795 — it measured spontaneous, not perturbational, complexity.

What the build found

The conscious/unconscious split is clean: wake ≈ REM (0.464) ≫ NREM (0.036), Δ = +0.428, well across the clinical cutoff PCI* ≈ 0.31 (Casarotto 2016). The bistable OFF-periods truncate the evoked causal chain (Massimini 2005; Pigorini 2015), collapsing complexity. The dissociation from arousal is explicit: the unconscious states still carry substantial raw activity (0.87–0.88) yet collapse PCI, and the activity ordering does not track the PCI ordering — PCI is not an arousal proxy.

ST-4: does the analog separate the states?

The break ST-4 was designed to cause is that the analog does not separate the states — that it collapses to one scalar. The test imports the exact build model and tests three signed features across 9 cells (3 seed cohorts × post-pulse windows 1.0 / 1.5 / 2.0 s, margin 0.05): REM > NREM, WAKE > NREM, and conscious-separated (min(wake, REM) − NREM; −1 if an unconscious PCI reaches a conscious one). It separately records the build's honest negative — the within-unconscious spread — as an [O], not a sign-stable claim.

What survived, and the recorded honest negative

The conscious/unconscious split survived: the conscious−unconscious margin ranged +0.347 → +0.481 (always far above the margin), and “deeper → lower PCI” held on no cell (conscious_unconscious_separation_sign_stable = True, broke_on = []). Gap 4 is graded [L] in-silico for the conscious/unconscious split — the clinical PCI's core binary falls out of the frozen kernel with no new constant — and [O] for the within-unconscious ordering, which collapses to one scalar. That collapse is itself scientifically sensible: PCI's clinical strength is the conscious/unconscious binary, not fine unconscious discrimination.

The preserved residuals are: the within-unconscious ordering is [O] (one scalar cannot rank NREM vs anaesthesia vs VS; spread 0.025 → 0.045, always below the margin); wake ≈ REM exactly, so the small clinical wake>REM gap is [O]; ANES/VS are an extended deeper-bistability sweep with no M14 grounding, so they are [O]; absolute PCI magnitudes are model units, not clinical PCI values, so they are [O]; and felt quality remains the firewall blank. The separation is post-window-invariant in sign — the sweep confirms no flip — with the margin shrinking smoothly as the window lengthens, never crossing.

Firewall (Axis-A). Every quantity on this page is a structural quantity of an in-silico model — a sign, a margin, an ordering — and is not a felt state, an experienced quality, or a level of consciousness (consciousness_claim = 0; the hard problem stays open, hard_problem_open = 1). These are model results, not validated neuroscience and not clinical guidance. efficacy = 0; not medical advice.