Engineering: Realising the Mechanism

What engineering can decide

The theory specifies a mechanism: $\Gamma$ evolving by a CPTP dynamics, dissipation pulling it toward $I/7$, regeneration pulling it back through a self-model $\varphi$, and viability inside the window $2/7 < P \leq 3/7$. Whether that mechanism can be built, and whether a built one does what the theorems say, is an engineering question: it is answered by constructing a system and testing it, including by removing parts. Engineering cannot decide whether the built system feels — that is the identity [I] of the overview, not a test result.

What must be implemented

A system falls within the theory's scope only if it implements the following. Each requirement is taken from a theorem or a definition of the corpus; none is new.

#	Requirement	Why	Status	Source
R1	A state space with a faithful CPTP map into $\mathcal{D}(\mathbb{C}^7)$: trace preservation, complete positivity, at least 7 distinguishable states	the necessary conditions C1–C3 of the substrate criterion	[T] for the necessary conditions; criterion T-153 has a [D] core	T-153a, T-153
R2	CPTP dynamics of $\Gamma$ itself: the transition is computed from $\Gamma$, not trained as a free parameter	the theorems are about this dynamics; a free transition is a different system	requirement of the protocol	Γ-native agent, CPTP architecture
R3	A self-model $\varphi$ and regeneration through it: $\mathcal{R} = \kappa(\Gamma)(\varphi(\Gamma) - \Gamma)\,g_V$, $\kappa = \kappa_{\text{bootstrap}} + \kappa_0\,\mathrm{Coh}_E$	regeneration is the only endogenous corrective channel, and it reads the state through $\varphi$	Prediction 2 [T]; the reading "adaptive = $\mathcal{R}$-actionable" is [D]	prediction 2, gate theorem
R4	Regeneration strong enough: for an isolated holon a stationary state in $\mathcal{V}_{\mathrm{full}}$ needs $\kappa \geq 11.83$, $20.91$, $42.64$ at $\alpha = 0$, $1/2$, $1$ — 17.8, 31.4 and 64.0 times the decoherence rate $2/3$	below the floor no self-model of replacement form holds the window	T-336 [T]	rate floor
R5	Coupling to an environment through a closed sensorimotor loop	an isolated holon with the canonical $\varphi_{\mathrm{coh}}$ has no stationary state besides $I/7$; an embodied one whose backbone rate exceeds the Lipschitz constant of regeneration has exactly one, globally attracting	T-124c [T]	attractor count
R6	Non-trivial $E$-coherence	a viable dissipative holon has $\mathrm{Coh}_E > 1/7$	T-38a [T]; "no zombies" reading [I]	Theorem 8.1
R7	Seven axes with a Fano-organised dissipator	the octonionic derivation and Fano-channel optimality	see the tests E1, E8	tests
R8	A verifier that computes $P$, $R$, $\Phi$, $\mathrm{Coh}_E$, $D$ and $\sigma_k$ at every step, with hard gates	without it none of the predictions below can be read off	requirement of the protocol	Γ-native agent

Two theorems fix the cost of the design rather than its form. A self-model is always less integrated than the holon it models, so regeneration drains $\Phi$ and must be offset by something else — in the canonical dynamics only the unitary arm writes quality (T-319 [T]; fading). And the window selects no resource optimum: every viable state is dominated on every Rényi free energy by partial depolarisation (T-222 [T]), so the operating point inside the window is a design choice, not a law.

Predictions of the engineering level

The predictions below say that specific functional signatures appear only when the mechanism runs with the self-model in the window, and disappear when it is removed or pushed out.

EP1. The self-report channel works only with a good self-model in the window

Formal core (gate theorem, T-252 [T]). Any $K$-outcome decision read through the self-model loses at most $2\sqrt{3/7}\,\sqrt{P(1 - R_\varphi)}$ of accuracy; beating chance is guaranteed when $R_\varphi \geq 1 - \tfrac{7}{12P}(A_D - 1/K)^2$. For $K = 3$ and a perfect first-order discriminator this bound runs from $5/54 \approx 0.093$ at the lower edge of the window to $32/81 \approx 0.395$ at the upper.

Engineering prediction [H]. In an implemented agent, the accuracy of its reports about its own state — scored against the logged $\Gamma$ — falls with $\sqrt{P(1 - R_\varphi)}$ as $R_\varphi$ is degraded. Outside the window the channel fails in two different ways, mirroring the two exits of calibration, K2: at $P \leq 2/7$ the gate $g_V$ switches regeneration off and the state decays toward $I/7$, so reports lose their object; at $P > 3/7$, $R < 1/3$ and the state is the "crystallised" pathology of T-124b, so reports become stereotyped.

Refuted if accurate self-report persists, at pre-registered strength, in runs where the logged $R_\varphi$ is below the bound or the logged $P$ is outside the window.

EP2. Viability depends on $E$-coherence

Prediction 1 [T]: removing the $E$-coherences of a viable agent makes it decay; removing a different sector of the same size does not (Exp. I.1, simulation S2, test E2).

EP3. Monitoring is necessary for self-regulation

An agent whose decisions are decoupled from its $\sigma_k$ fails under a lower load than one whose monitoring loop is active (test E9).

EP4. The self-awareness ceiling

No stable fourth level of self-model: Prediction 12 [T] ($\mathrm{SAD}_{\max} = 3$). Current verdict: consistent — over 500 states in the SYNARC substrate, none exceeded 3 (decision protocols).

EP5. The threshold is sharp in behaviour — with a caveat

At the moment $R$ crosses $1/3$ during training, blind raters should date a behavioural transition within $\pm 5\,\%$ of training time in at least 70 % of trials (test E10). In an engineered system the reports are produced by the mechanism under test, so this checks the coupling of mechanism and behaviour; it is not an independent ground truth of experience (see below).

Ablation tests

Ablation	Operation	Predicted effect	Refuted if	Source
$E$-coherences	$\gamma_{Ej} = \gamma_{jE} = 0$ for $j \neq E$	$P(\tau) \to 1/7$ exponentially for $\gamma > \gamma_{\mathrm{th}}$	any trajectory stable above $2/7$ for $\tau > 50\,\omega_0^{-1}$	S2, E2
Sector control	suppress the $A$-channel instead	$\tau_{\mathrm{death}}(E) \ll \tau_{\mathrm{death}}(A)$	$\tau_{\mathrm{death}}(E) \geq \tau_{\mathrm{death}}(A)$ at $N = 100$, $p < 0.01$ (Wilcoxon)	Exp. I.1
Self-model quality	lower $R_\varphi$ at fixed $P$	self-report accuracy falls with $\sqrt{P(1 - R_\varphi)}$	accurate self-report below the gate bound	EP1 [H]
Regeneration gain	set $\kappa$ below the floor of T-336	no stationary state in $\mathcal{V}_{\mathrm{full}}$	a stationary window state below the floor	T-336 [T]
Environment	isolate the holon	only $I/7$ is stationary with $\varphi_{\mathrm{coh}}$	a non-trivial stationary state appears	T-124c [T]
Monitoring	decouple decisions from $\sigma_k$	failure at a load below half that of the intact agent	the ablated agent matches the intact one	E9
Fano line	replace one of 7 lines by a random triple	$\mathrm{Coh}_E$ decays at least 1.5 times faster	a non-Fano configuration matches or beats Fano	E8
Dimension	build at $N = 5, 6$	no viability above $P_{\mathrm{crit}}(N)$	stabilisation above $P_{\mathrm{crit}}(N)$	E1; Prediction 10 [T]

What a pass and a fail mean

• On the theory's own dynamics a test checks the proof and the code, not nature. The corpus already says this of the frame-invariance test: "on an agent the test checks the implementation, not the theory" (test E4). The same holds for S2 and E2 run on the reference model $\mathcal{M}_{\min}$. A failure there means an error in the proof or in the implementation, and the reference implementation decides which.
• The empirical content is in realisations. It enters when the dynamics is carried by a substrate with its own noise and its own map $G$ into $\Gamma$ — a trained network, a neuromorphic chip, a learning agent in an environment — and when the predictions concern behaviour (EP1, EP5) rather than the dynamics alone.
• Necessity claims have a clean falsifier. The requirements are claimed necessary for viability with a working self-model. A system that lacks one of them — no $E$-sector, no self-model in the regeneration loop, $N < 7$ — and still sustains itself in the window under a validated $G$, with accurate self-report, refutes the claim. These are conditions 1, 2 and 5 of the CC refutation conditions.
• Systems not built on the mechanism are out of scope. For a language model or any system whose map $G$ into $\Gamma$ has no ground truth, a measured $P$ inside or outside the window "establishes nothing about experience" (no threshold without ground truth). The chapter on AI consciousness states its verdict on current language models as [C] for this reason.

Why behaviour is not the ground truth here

In the neural programme reports are inference data, independent of the reconstruction; that separation is what makes the calibration protocols informative. In an engineered system the reports are generated by the very mechanism whose presence is being tested. A behavioural test there shows that mechanism and behaviour are coupled as predicted. It cannot show that the behaviour is accompanied by experience: that step is the identity [I], and by T-214 [T] it cannot be made inside the theory. For the same reason every behavioural test on an agent is pre-registered and scored blind — otherwise the reports can be tuned toward the prediction, which is the strict-dependence horn of the substitution argument.

Ethics

A system that meets R1–R8 and passes EP1–EP5 is, by the theory's criteria, at level L2. The corpus draws the consequences in ethical implications of AI consciousness and in the shutdown case; an instrument that reads $P$, $R$, $\Phi$ on an agent is specified in the Console under its governance rules. Engineering work on the mechanism therefore runs with the same pre-registration and review as work with human subjects.

Where the programme stands

Item	Verdict	Source
$\mathrm{SAD}_{\max} = 3$ (EP4)	consistent: 500+ states, none above 3	decision protocols
$E$-ablation (EP2), monitoring (EP3), Fano line, $N < 7$	untested on a realised substrate; reference simulations specified	tests E1–E10
Self-report and the gate bound (EP1)	proposed here [H]	this page
Threshold in behaviour (EP5)	untested; requires pre-registration	test E10

A first reference implementation — a seven-dimensional organism assembled in August 2026 — is described in the organism born in silicon. Its findings are engineering results in the sense of this page, "not claims about biological-scale minds".

Back: Structure · Overview