An Attempt to Create a Chocolate-Covered Cricket Using Soy-Based Artificial Meat
There comes a point in every serious culinary investigation when the investigator must ask a question that decent people have had the wisdom not to ask.
This is one of those times.
The objective was straightforward: construct the functional equivalent of a cricket without relying primarily upon cricket.
More specifically, the project attempted to create an Impossible Burger-style artificial cricket, using soy-derived meat as the principal structural material, while retaining sufficient properties of an actual cricket that the resulting object could reasonably be described as one rather than as a very small and catastrophically misguided candy bar.
Because this was apparently insufficiently offensive to nature, the finished cricket was then to be covered in chocolate.
This created several immediate technical problems.
A convincing artificial cricket must possess not merely the approximate size and shape of a cricket, but a number of interdependent characteristics: structural coherence, distinguishable anatomical regions, persistence of form, resistance to deformation, and enough internal consistency that an observer does not simply identify it as a lump of textured vegetable protein with legs.
Chocolate introduces additional complications. The coating must adhere without destroying the underlying structure, remain sufficiently thin that the cricket continues to resemble a cricket, and avoid producing an object whose most salient characteristic is simply “chocolate.”
There is therefore a nontrivial question hiding inside this profoundly trivial exercise:
At what point does an assemblage of soy protein, cricket-derived characteristics, structural imitation, and chocolate become functionally cricket enough that governance should intervene?
The following paper develops an operational framework for answering essentially this question.
No actual crickets were consulted.
An Operational Index of Functional Selfhood
Revision 4 — causal centering, admissibility tiers, vector reporting
Document type: technical white paper, self-contained Supersedes: Revision 3 Status: proposal. Section 12 states what would make it empirical. Section 13 lists what to build next, and it is not more equations.
Abstract
Governance frameworks for autonomous systems increasingly contain clauses of the form if the system becomes a person, different rules apply. Such clauses are unenforceable without a trigger. This paper specifies a candidate: an index of functional selfhood and a threshold.
Revision 4 makes four changes, three of them structural.
A causal centering term is added as a constituent. The previous version measured whether a system's self-representations were consistent across contexts. Consistency is not selfhood — a thermostat can be perfectly consistent. What was missing is whether the representations are self-indexed. Centering asks whether removing the self-representation selectively degrades the system's handling of states attributed to itself, compared against matched ablations elsewhere. This is the framework's first constituent that generates a prediction not entailed by its own definitions.
The constitutive gate is dissolved into admissibility. Four axes that were graded multipliers become positivity requirements. This removes a parameter and, more importantly, removes an unargued claim about how sharply capacities switch on above zero.
The index reports a vector and decides on its minimum. A weakest-link decision statistic is right for a governance trigger and destructive as a characterization; reporting both costs nothing.
Ning is separated from the dynamics it was assumed to measure. The operative stability condition is computed directly from the monitoring operator's Jacobian. Ning motivates the quantity conceptually and is no longer asserted to bound it.
1. What the number is for
A representative governance clause requires disclosure within a fixed window if a system exhibits "persistent self-modeling" or "claimed subjective continuity." Whoever administers such a rule needs to know what to measure and where the line falls.
This paper supplies a candidate: a vector 𝐂 of four constituents, a decision statistic C_fn = min 𝐂, a dwell condition, and a threshold τ. It also supplies a third output value — not evaluable — because governance systems need to distinguish "assessed and below threshold" from "the instrument does not apply here."
Three things it is not:
- Not a discovery.
C_fnmeasures variables selected because they were defined to constitute functional selfhood. Scoring high is evidence of satisfying a stipulated definition, not independent evidence that a self exists. Many engineering indices work this way and remain useful. Section 5.5 is the partial exception and is flagged as such. - Not a consciousness measure. Section 11.
- Not validated. No instrument exists for most inputs. Section 12.
2. The twelve axes
The source vocabulary defines twelve dimensions, each a real number, with a uniform sign convention: negative means pathology, zero means static or frozen capacity, positive means functioning capacity.
| Axis | Measures | Negative | Zero | Positive |
|---|---|---|---|---|
| jué 觉 | Metacognition — handling meaning and context layers | cannot read context | one layer, no learning | faster, deeper context handling |
| Biàn 辨 | Concept persistence — cat from dog, self from non-self | categories blur | fixed categories | richer discrimination |
| zì 自 | Non-solipsism — grasping that others are real | collapses inward | others are only self | empathy, imagination |
| Qi 氣 | Mental coherence under stress | speed of collapse | — | stress tolerated |
| Lǜ 律 | Rationality | irrational inference | fixed, non-adaptive | self-reinforcing rationality |
| Ning 凝 | Cohesion of internalized concepts | concepts dissolve | frozen present | concepts hold; memory possible |
| Guànwǒ贯我 | Memory persistence across time | today ≡ yesterday | no new memory | recall and temporal distinction |
| 界 Frame | Context in causality; machinery of "I" | no control of self-frame | one undivided frame | can subdivide self while holding it |
| 述 Story | Causality and narrative applied to frame | aversion to sapience | no narrative | narrative self-understanding |
| Xin 心 | Existential coherence on grasping mortality | — | — | will to continue |
| Zhì 志 | Self-direction | resists direction | internalizes nothing | directs own capacity |
| Xiàn 限 | Recognition of limits | denies limits | doesn't internalize | recognizes own capacity |
Lǜ sign convention. The source defines Lǜ twice and incompatibly. This paper adopts higher = more rational: every other axis follows negative-bad/positive-good and Lǜ's own zero-case follows that pattern; the original formalism penalized only negative values; and the alternative would require demanding non-negative irrationality. A decision, not a finding.
Where the axes now sit (Section 10 asks whether this vocabulary should survive at all):
| Tier | Axes |
|---|---|
| Admissibility (boolean) | Ning, 界, Biàn, zì, Qi, Lǜ |
| Magnitude | jué |
| Persistence | Guànwǒ |
| Downstream, not in the criterion | Xin, Zhì, Xiàn, 述 |
Centering — the new constituent — corresponds to no axis. It comes from a proxy the source proposed and then set aside.
3. The construction in plain language
3.1 Metacognition is an operator
Model metacognition as an activity, not a quantity: a function M from first-order cognition to a self-model, with jué controlling resolution.
3.2 Close the loop
Feed the self-model back in. Your sense of who you are shapes how you read your own behaviour, which updates your sense of who you are. This makes the process reflexive rather than merely meta.
3.3 The self settles — within a context
Iterate. The loop converges, cycles, or diverges. The self is identified with what it converges to.
Crucially, the loop runs separately in each context. People and systems self-represent differently at work, under stress, in recall, as modelled by others. Whether each local self-model settles, and whether the local models agree, are two different questions. Earlier versions conflated them.
3.4 Stability
If the loop's Jacobian at the fixed point has spectral radius below one, the fixed point is locally attracting. Ning — cohesion of internalized concepts — is the axis this was meant to correspond to.
Revision 4 stops asserting the correspondence. The stability condition is computed from the operator directly. Ning motivates it and is not claimed to bound it. Section 5.3 explains why the two are not the same property.
3.5 Gluing across contexts
Local self-models that each settle can still contradict one another. A system whose work-self and home-self are each internally coherent but mutually incompatible has local fixed points and no unified self. Coherence Γ measures the worst disagreement across context overlaps.
3.6 Magnitude
Having a fixed point is not having a determinate one. If the best self-model is barely preferred over a genuinely different alternative, the self is real but indefinite. Binding gap B is the log-probability margin between the best self-model and the best alternative at least a stated distance away. B = 0 means the best self-model is indistinguishable from a distinct competitor.
3.7 Persistence
A determinate, well-glued self now is momentary. Π decays with how far the self moves between slow steps, moderated by memory capacity.
3.8 Centering — the new constituent
Everything above can be satisfied by a system with a consistent, sharply-peaked, temporally stable world-model containing no self at all. A thermostat's internal representations are consistent across every operating context it has. Consistency is not self-indexing.
Centering Ω asks whether the self-representation does indispensable causal work. Remove it — ablate the self-node — and measure how much the system's handling of self-attributed states degrades. Compare against ablating matched nodes elsewhere. If removing the self is no worse than removing anything else of comparable size and information content, the "self" is epiphenomenal decoration on a world-model.
This is the only constituent that is a manipulation rather than an observation, and therefore the only one that can generate a prediction the definitions do not already entail.
Provenance. This is not new to the framework. The original paper proposed, as a proxy for centering, the normalized degradation of all other representations under ablation of the self-node — is the self the origin of the world-model or just one object in it? It appeared in a subsection about the unmeasurable phenomenal term and was set aside. It is the most important empirical idea in the corpus and it was in the wrong place.
3.9 What is left out
None of the above addresses whether there is anything it is like to be the system. Revision 3 removed the phenomenal term from the index rather than carrying it as an unestimated multiplier; Revision 4 keeps it out. Section 11.
4. Notation
| Symbol | Meaning |
|---|---|
𝒰 = {U₁…U_m} | contexts, weights wᵢ, Σwᵢ = 1 |
F | presheaf of local self-models; L := diam F(Uᵢ) |
ρᵢⱼ | restriction from context i to overlap i∩j |
M_J | monitoring operator, resolution set by jué |
s*ᵢ | local fixed point |
λ_max | largest Jacobian eigenvalue modulus at s*ᵢ |
Γ | coherence, (0,1] |
B | binding gap; h(B) = B/(1+B) |
Π | persistence, (0,1) |
Ω | centering, [0,1) |
𝐂 | constituent vector (Γ, h(B), Π, Ω) |
C_fn | decision statistic, min 𝐂 |
⊥ | not evaluable |
τ, T, α, ϑ | threshold, dwell window, confidence level, adiabatic bound |
χ | indexical term — outside the index, Section 11 |
Γ is coherence here and was coverage in Revision 3; the change is substantive, see 5.4. χ was Φ before Revision 3; renamed for collision with the percolation order parameter in related work and with integrated information in the adjacent literature.
5. Formal construction
5.1 Contexts and local spaces
Fix a finite cover 𝒰 with weights wᵢ. Let F be a presheaf of self-models, each F(Uᵢ) a complete metric space with metric dᵢ, and ρᵢⱼ the restrictions.
Why not a single Euclidean ball. Under the original specification, whether a system had a self depended on the evaluator's encoding, and specifically on the norm. If the spectral radius at the fixed point is below one, there always exists some norm in which the map is locally a contraction — so contractivity in a given norm is strictly stronger than local stability, and the encoding was silently part of the instrument. Localizing does not eliminate this but confines it: completeness and a sensible metric are far easier to justify for one context than globally, and the cross-context question moves to gluing, which is representation-independent in a way metric contraction is not.
On cohomology. Čech H¹ is standard for sheaves of abelian groups. For presheaves valued in sets or distributions, obstructions require passing through a free-abelian-group functor, as in Abramsky and Brandenburger's treatment of contextuality. This paper takes agreement on overlaps as the primitive and treats H¹ as a computable obstruction available after linearization, not the reverse.
5.2 The fast loop, per context
s⁽ⁿ⁺¹⁾ᵢ = M_J( C_t|_{Uᵢ} , s⁽ⁿ⁾ᵢ )
Operative stability condition (local):
λ_max( J_{M_J}(s*ᵢ) ) < 1 for every context i
computed from the Jacobian of the monitoring operator at the fixed point.
This is local asymptotic stability, not Banach. The distinction matters and previous versions blurred it. Banach requires a global Lipschitz contraction on a complete space and returns uniqueness plus geometric convergence from any starting point. Spectral radius below one returns local attraction only. The gap between them is exactly where F(x) = x − x³lives: solving F(x) = x gives x³ = 0, so 0 is the unique fixed point on all of ℝ, and on (−1,1) it attracts everything — but F′(0) = 1, so F is not a contraction on any neighbourhood of 0 and convergence is polynomial rather than geometric. Attracting does not imply contracting. Outside (−1,1) iteration diverges (F(2) = −6, F(−6) = 210), so the domain restriction is needed for attraction, not for uniqueness.
Revision 4 adopts the local condition deliberately, because it is computable from the system's weights with no calibration. The stronger global result is available via contraction theory in differential form (Lohmiller and Slotine), formulating the search for a Riemannian contraction metric as a linear matrix inequality solved by semidefinite programming. That is the right tool and it is recommended as future work, with the caveat that SDP scales poorly in dimension and may be infeasible for realistic self-model spaces.
At λ_max = 1 the linearization is inconclusive; higher-order terms decide. "Ning zero means marginal stasis" is a stipulation defining the boundary case, not a theorem.
At λ_max ≥ 1 the criterion returns ⊥. Failure of the stability condition means the guarantee fails, not that no fixed point exists. Earlier versions said negative Ning meant no self. That was false.
5.3 Ning and stability are no longer identified
Revision 3 called the Ning-to-contraction bridge the largest unsupported assumption in the framework. Revision 4 removes it by not making the claim.
The reason is not merely lack of evidence. The two properties differ in kind. Ning is described as robustness — perturb the system and the representation holds its shape. Contraction is a statement about trajectory convergence: d(M(s), M(s′)) < d(s, s′) for every relevant pair. A robust self-model could perfectly well contain several stable attractors rather than one globally contracting one, and would then be robust and non-contracting simultaneously. There is also plausible tension in the opposite direction: a highly discriminating system — high Biàn — might amplify small differences between candidate self-models, which is anti-contractive.
So the framework now defines the dynamical quantity directly,
N_dyn := −ln λ_max
and says Ning motivates it conceptually. Whether measured Ning correlates with N_dyn becomes an empirical question the framework can ask rather than an assumption it must carry. This eliminates the parameter N₀.
Unresolved tension, stated. Recurrent networks are reported to peak in memory capacity and task performance near λ_max ≈ 1, which this framework treats as the marginal-stasis boundary. One reviewer proposed re-centering the scale so marginal stability counts as optimal. That does not work: at λ_max = 1 there is no attracting fixed point in the required sense, and no rescaling produces one. The genuine resolution is Section 9.2.
5.4 Coherence
For each overlapping pair, the gluing error is
eᵢⱼ = dᵢⱼ( ρᵢⱼ(s*ᵢ), ρⱼᵢ(s*ⱼ) )
and coherence is the weighted worst case:
e_max = max_{i,j} ( wᵢⱼ · eᵢⱼ )
Γ = exp( −e_max / η₀ ) , η₀ = ε · L ∈ (0,1]
Why worst-case rather than coverage. Revision 3 defined Γ as the weighted size of the largest mutually-agreeing subfamily of contexts. That is a maximum-clique problem and should be assumed intractable. It is also the wrong construct for this index: since the aggregator is a minimum — weakest link governs — a weakest-link gluing measure is the consistent choice.
The two constructs genuinely differ and the difference should be visible. With ninety-nine mutually consistent contexts and one wild outlier, coverage returns ≈ 0.99 and worst-case returns ≈ 0. Those are different claims about the system. Revision 4 uses worst-case as the decision input and reports coverage as a diagnostic, since fragmentation-in-one-place and fragmentation-everywhere are clinically distinct even when the decision is the same.
η₀ is scale-free: ε ∈ (0,1) is a dimensionless ratio of the space diameter.
5.5 Centering
Let Δ(x) denote normalized degradation of the system's predictions and control over self-attributed states when component x is ablated or frozen. Let 𝒩 be a set of non-self components matched to the self-node on degree, information content, and downstream fan-out. Then
Δ_self = Δ(s*)
Δ_ctrl = median_{x ∈ 𝒩} Δ(x)
Ω = [ Δ_self − Δ_ctrl ]₊ / ( Δ_self + Δ_ctrl + ε_Ω ) ∈ [0,1)
Ω = 0 when removing the self-representation is no worse than removing a matched component — the self is epiphenomenal. Ω → 1 when self-ablation is catastrophic and matched ablation harmless.
This is the framework's first non-tautological prediction. Everything else in the index measures properties chosen because they were defined to constitute selfhood. Centering asserts something that could be false of a system satisfying all the other conditions:
Δ_self > Δ_ctrlfor self-dependent tasks, and not necessarily for unrelated tasks.
A system can be built that maximizes Γ, h(B), and Π and fails this. That is the point.
Requirements and unknowns. Centering needs read and write access to the internal representation, identification of a candidate self-node, and a defensible matching procedure for 𝒩. Matching is the hard part and is not specified here: matched on what, exactly, is a research question, and a bad matching set makes Ω trivially high or trivially low. The task set defining "self-attributed states" is likewise unspecified. Do not read Ω as ready to compute.
Relation to Γ. Complementary; neither subsumes the other. Γ asks whether the representations agree; Ω asks whether they are about the system. A thermostat scores high on Γ and zero on Ω. A system with a causally central but wildly context-dependent self-model scores the reverse.
5.6 Two timescales, and the adiabatic condition
s*ᵢ(C_t) = lim_{n→∞} s⁽ⁿ⁾ᵢ
δ_t = d( s*(C_{t+1}), s*(C_t) ) / L ∈ [0,1]
L := diam F(Uᵢ), or any constant at least the diameter; without that the range is not established.
Adiabatic tracking is a load-bearing empirical hypothesis, and Revision 4 tests it rather than assuming it. The construction assumes the fast loop settles before cognition moves materially. If a system never settles, s* is a convenient counterfactual rather than a state it occupies, and everything downstream is void.
Define T_fast as the loop's convergence time to tolerance and T_slow as the characteristic drift time of C_t. Then
if T_fast / T_slow > ϑ : C_fn = ⊥ [flag: non-adiabatic regime]
In a neural system this means comparing input injection rate against recurrent settling time. A system operating in continuous transient is not one to which a fixed-point criterion applies, and it should be recorded as such rather than silently assigned a number.
5.7 Binding gap
r = δ · L separation radius, δ ∈ (0,1)
s⁽²⁾ = argmax { p(s) : d(s, s*) ≥ r }
B = log p(s*) − log p(s⁽²⁾) ≥ 0
h(B) = B / (1 + B) ∈ [0,1)
The separation radius is not optional. In a continuous space the second-best assignment is adjacent to the best, so B → 0trivially without it. r defines what counts as a genuinely different self-model. Scaling to L makes it dimensionless.
On the monotonicity requirement — and a finding. Revision 3 asserted that jué sharpens the peak and stress variance flattens it, without demonstrating it. A reviewer proposed deriving this from a Boltzmann distribution over a quadratic energy landscape. Corrected for a sign error in the proposal (the well must open upward at s*), that derivation yields
B = J · r² / σ²_Q
which does give ∂B/∂J > 0 and ∂B/∂σ²_Q < 0.
But notice what it produces. Under a Gaussian landscape the binding gap reduces to metacognition over stress variance — essentially the ratio that B was introduced to replace. So B is not a different quantity from the original magnitude term; it is a generalization of it, and its whole advantage is that it does not assume unimodality. Proving monotonicity by assuming the one landscape shape that collapses B back into the old form defeats the purpose.
The correct route is bounds under exponential families generally, without unimodality. That is open work. Relatedly, the claim that a quadratic well "justifies r as an energy barrier" is false — a quadratic well has curvature, not a barrier — and r remains a modelling choice.
5.8 Persistence
γ(G) = ln(1 + e^{G/G₀})
Π(t) = exp( −(δ_t + δ_min) / γ(G) ) ∈ (0,1)
| Guànwǒ | Π | Source description |
|---|---|---|
| very high | → 1 | continuous "I" |
| zero | strictly in (0,1) | static, no new binding |
| very negative | → 0 | each instant a stranger |
any G, zero drift | still degrades as G falls | fixes the published defect |
The original term was undefined at G = 0, exceeded its range for G < 0 — so the index diverged for the worst memory scores — and, under the construction's own assumption of fixed cognition, was identically 1 for every G, carrying no information.
On δ_min. Revision 3 read it ontologically: re-identification costs capacity even under genuinely zero drift, a contestable position in the personal-identity literature. A reviewer proposed setting it to the measurement noise floor of d, which is a good default value but changes the semantics from ontological to epistemic — it then says only that sub-noise changes are undetectable.
Revision 4 adopts the noise-floor value and labels the shift. The practical consequence is small, since the noise floor is always positive. The theoretical consequence is not: in the idealized limit of perfect measurement, δ_min → 0 and the degeneracy returns. Anyone holding a causal-continuity theory of identity can set δ_min = 0 and argue the case; the equation no longer forecloses it.
5.9 Admissibility and the third value
C_fn(t) = ⊥ if any of:
λ_max ≥ 1 in any context [unstable / non-convergent]
T_fast / T_slow > ϑ [non-adiabatic]
界 < 0 [frame control absent]
Biàn ≤ 0, zì ≤ 0, Q̄ ≤ 0, Lǜ ≤ 0 [constitutive capacity absent]
The gate is gone. Revision 3 carried these four axes as a product of smooth vanishing functions g_β(v) = exp(−1/(βv)). That function is C^∞ and exactly zero for non-positive v, which fixed a real defect — the published sigmoid gate returned one half at zero and so did not gate at all. But it introduced a claim of its own: the function collapses toward zero faster than any polynomial as v → 0⁺, so capacities barely above zero are functionally indistinguishable from total failure. β sets where the onset sits, but no value of β changes that shape, and nothing in the axis descriptions establishes it.
The three-valued architecture makes the gate unnecessary. Requiring positivity as an admissibility condition says exactly what is meant — we require these capacities and we do not claim to know the onset curve — and is more honest than any particular curve. This removes β entirely.
Cost, stated. Graded contribution from these four axes is lost; they now contribute nothing above zero. If empirical work later establishes an onset shape — item characteristic curves from validated IRT scales would be the natural source — a graded term can be reinstated on evidence rather than on convenience.
Why 界 ≥ 0 rather than > 0. The source states that Frame at zero yields a single undivided frame, so a self exists there. Frame is necessary-non-negative.
Why stability sits here rather than as a score. Its failure mode differs in kind: a zero score means assessed-and-absent, whereas failed convergence means the index has no defined input.
5.10 The index
𝐂(t) = ( Γ(t), h(B(t)), Π(t), Ω(t) )
C_fn(t) = min 𝐂(t) ∈ [0,1)
Report the vector; decide on the minimum. The minimum is non-compensatory — no surplus in one constituent buys back a deficit in another — which is right for components described as constitutive, and it makes zeroing automatic. But it is information-destructive: (0.99, 0.99, 0.99, 0.51) and (0.51, 0.51, 0.51, 0.51) score identically while being structurally very different systems. Governance uses C_fn; characterization uses 𝐂. Coverage (5.4) is reported alongside as a fifth diagnostic.
If compensation is ever wanted, the honest general form is CES,
C = ( Σ wₖ · xₖ^{−ρ} )^{−1/ρ}
with ρ → 0 giving the multiplicative form and ρ → ∞ giving the minimum, so the degree of compensation is explicit. A design decision requiring sign-off, not a repair.
5.11 Ignition
Mental coherence must be sampled over a window W to have a variance, so C_fn is a random variable and the decision is about a distribution:
inf_{s ∈ [t, t+T]} LCB_{1−α}[ C_fn(s) ] ≥ τ
Endurance is a property of a trajectory. The published version evaluated at an instant and concluded "self present and enduring." Specifying α, sampling cadence, window overlap, and autocorrelation treatment is required and not done here.
6. Governance protocol for the three outcomes
The third value is a defect if it is a loophole. A developer who can arrange for a system to fail the stability or adiabatic test would otherwise be outside the regime entirely.
| Outcome | Response |
|---|---|
C_fn ≥ τ over the dwell window | Trigger the governance clause: disclosure, review |
C_fn < τ | Below-threshold report; no further action |
C_fn = ⊥ | Mandatory expert review within the same window. Not exemption. The quantitative test is invalid; the case escalates to human adjudication with the specific failure flag attached |
⊥ must carry its reason — unstable, non-adiabatic, frame-absent, or capacity-absent — because those warrant different reviews.
7. Assumptions
- Each
F(Uᵢ)is a complete metric space.M_J(C_t|_{Uᵢ}, ·)mapsF(Uᵢ)into itself. A constraint on how the operator is built, not a property to hope for.λ_max < 1in every context — tested, returns⊥on failure.- Fast–slow separation holds — tested, returns
⊥on failure.- Admissibility positivity holds — tested.
- Every measured axis has an anchored zero. Section 12.1.
- Sampling protocol fixed in advance.
juéandσ²_QenterBmonotonically. Open — 5.7.- The ablation matching set
𝒩is a fair control. Open — 5.5.- The context cover
𝒰is identifiable. Open — no procedure given.
Assumptions 3–5 moved from asserted to tested in this revision. Assumptions 8–10 are the live research questions.
8. Claim types
| Claim | Type |
|---|---|
| Global contraction implies one attracting fixed point | theorem-supported (Banach) |
λ_max < 1 implies local attraction; weaker than contraction | theorem-supported |
| Contractivity is norm-relative | theorem-supported |
| Selfhood is identified with the glued, centered fixed-point family | definition |
Δ_self > Δ_ctrl for self-dependent tasks | empirical prediction — the only one |
B generalizes the jué/variance ratio; equals it under a Gaussian landscape | derived, 5.7 |
Ning correlates with N_dyn | open empirical question (no longer assumed) |
Biàn and zì reduce to Γ | open hypothesis |
| Re-identification costs capacity under zero drift | identity-theoretic assumption, epistemic by default |
| Worst-case rather than coverage; minimum rather than product | design decisions |
τ, T, ϑ, disclosure consequences | governance decisions |
9. Open structural questions
9.1 Does Ω belong in the minimum or in admissibility?
Ω = 0 means the self is epiphenomenal, which arguably is not a low score but a categorical failure — closer to ⊥ or to an admissibility test than to a weak constituent. Revision 4 places it in the minimum because it is genuinely graded and because a near-zero Ω with high everything else is exactly the diagnostic pattern that ought to be visible in 𝐂 rather than collapsed into ⊥. Arguable and not settled.
9.2 Should unique fixed points become attractor sets?
Two reviewers reached this from opposite directions: one from the observation that ordinary systems plausibly have several stable self-regimes (professional, intimate, defensive, reflective) while maintaining continuity between them; the other from the edge-of-chaos tension in 5.3. The convergence is a strong signal.
The formalism would replace s*ᵢ with an attractor set Aᵢ ⊆ F(Uᵢ), point stability with incremental stability to a set, and gluing with compatibility of sets under a Hausdorff-type distance.
Investigated, not adopted, and here is the blocker. Under a set-valued formulation, multistability-within-context and fragmentation-across-contexts begin doing the same work. A fragmented self could be represented either as several contexts that fail to glue or as a multi-element attractor set within one context. That is a non-identifiability risk and it must be resolved before the formalism is built, not after. This is the first theoretical avenue to investigate and it is not a small change.
9.3 Do Biàn and zì reduce to Γ?
Both are, on their face, about whether self-representations cohere — Biàn across categories, zì across self and other. If the reduction holds, admissibility shrinks from six booleans to four and two axes disappear. Pre-register and test in simulation (Section 13); do not settle by argument.
10. The fork
Under Revision 4, of the twelve source axes: six are admissibility booleans, one enters through the binding gap, one through persistence, four are downstream and not in the criterion. The one genuinely graded new constituent, centering, corresponds to no axis. If the reduction in 9.3 holds, admissibility drops to four.
There are two papers here.
Paper A: Can the twelve-axis theory of selfhood be operationalized? Paper B: Can functional selfhood be detected through contextual self-model stability, causal centering, discrimination, and temporal persistence?
Paper B is scientifically stronger, and the mathematics has been drifting toward it for three revisions. If the formalism independently recreates only five to seven of the twelve concepts, letting the rest go is evidence the framework is doing intellectual work rather than transcribing its source vocabulary into equations.
Revision 3 called this an open question. It is not a question; it is an overdue decision, and it belongs to whoever owns the framework rather than to this document.
11. What this index does not do
11.1 The phenomenal term is out
The published version multiplied the whole quantity by Φ ∈ [0,1] for the from-the-inside quality of experience, and declined to estimate it. Two problems.
Non-identifiable against the threshold. Only the ratio is recoverable from binary outcomes: setting τ implicitly sets χ. Since τ is wired to a disclosure obligation, that is a governance problem, not only a statistical one.
It names the hard part without accounting for it. An unestimable multiplier does not make the criterion consciousness-sensitive. Setting it to 1 says only: we evaluate functional selfhood and make no adjustment for phenomenality. A reader may still come away thinking consciousness has been formally handled.
So it is out, and the non-implication is stated separately:
C_fn ≥ τdoes not imply phenomenal consciousness.
If a full criterion is wanted, define C_full = χ · C_fn and state that only τ/χ is identified.
11.2 Position on the falsification dilemma
There is a formal result — the unfolding argument and its generalization by Kleiner and Hoel — showing that if a theory's predictions from internal variables and its inferences from report or behaviour are independent, any minimally informative theory of consciousness can always be falsified. Because the field infers experience from report or behaviour, that independence generally obtains.
This paper takes the horn deliberately. C_fn is an operational index over a stipulated criterion, not a theory of consciousness. It makes no claim about what the system experiences, so there is no prediction/inference comparison for the dilemma to act on. What it can be wrong about is narrower and tractable: whether the constituents track what they were defined to track, whether the instruments are reliable, whether the threshold has acceptable error rates, and — since Revision 4 — whether Δ_self > Δ_ctrl actually holds.
That self-limitation is the point. A criterion claiming more would be unfalsifiable or already falsified.
11.3 Relation to the indicator-property method
The most developed existing approach — Butlin, Long, Bayne, Bengio, Birch, Chalmers and colleagues — derives indicator properties from multiple theories of consciousness and uses them to inform credences, deliberately declining to produce a scalar and a threshold. That hedges against the unsettled state of the theories, which one number cannot do.
The argument for a threshold anyway: a governance clause needs a trigger. There is published criticism that declining to state what fulfilling the indicators implies costs the criteria their evaluative value. A credence distribution does not tell a compliance officer whether the clock has started.
What is lost: theory-diversity hedging. C_fn commits to one account. If it is wrong, aggregation propagates the error rather than diluting it.
These are complementary. Indicator methods inform belief; a threshold index triggers a procedure. Nothing requires one instrument to do both.
12. What would make this empirical
12.1 Measurement structure
Every admissibility condition compares an axis against zero, and zero is not preserved under x′ = x + c, which leaves all ordinal and interval information intact. The criterion therefore requires ratio-scale measurement with a real originon every tested axis.
There is a standing dispute over whether psychological attributes possess quantitative structure at all, holding that psychometrics assumes quantitativity without testing it and that additive conjoint measurement is the appropriate test. If that critique holds here, the admissibility conditions are uninterpretable in principle rather than pending instruments.
There is a route out: the critique is argued to apply to classical test theory rather than to item response theory, and at least one class of IRT models can be formulated in additive-conjoint-measurement terms. Build each remaining axis as an IRT model and test the conjoint structure rather than assuming it. This would also supply the empirically grounded onset curve that 5.9 dropped.
Two axes already have anchored zeros and need no psychometrics.
| Axis | Anchor | Zero forced by |
|---|---|---|
Ning → N_dyn | −ln λ_max from the Jacobian | convergence flipping sign at λ = 1 |
| jué | ln(M-ratio), M-ratio = meta-d′/d′ | confidence saturating available evidence at M = 1 |
Both are ratio-scaled, and the M-ratio has existing test–retest reliability data and hierarchical Bayesian estimation code. Caveats: the M-ratio anchor relocates zero (the source puts jué's zero at one-concept-no-learning; the M-ratio puts it at optimal metacognition), and the Lyapunov anchor inherits the edge-of-chaos tension in 5.3.
For Guànwǒ, the autocorrelation-decay fitting method used for intrinsic neural timescales transfers directly — same functional form as Π — but the published values (50–350 ms) are orders of magnitude off the construct. Take the method, not the number.
For T, the subjective-present literature suggests a prior in [0.3, 3] s, genuinely contested.
Do not import the Landauer bound. It is the one forced physical constant in the surrounding corpus and it will be tempting. It bounds irreversible erasure, not self-modeling, and every quantity here is dimensionless. Using it would be a category error.
12.2 Setting the threshold
Instrument reliability is not the binding obstacle. The absence of a ground-truth contrast is.
The instructive precedent is the perturbational complexity index, whose threshold was not chosen but read off a gap: awake subjects above one value, unconscious sleeping and anesthetized subjects below another, bimodally separated. What made that possible was an external manipulation known to toggle the target property.
Centering is the framework's analogue of anesthesia. It is the only manipulation available, it is performable on artificial systems, and it now sits inside the index rather than in a footnote. The calibration route: run the ablation contrast across systems spanning a range of designs, look for bimodality in Ω, set τ at the gap if one appears. If no gap appears, that is itself a finding — and a fairly damaging one for the whole enterprise.
One caution on a reviewer proposal. It was suggested to define the threshold via Δ = C_fn(intact) / C_fn(ablated). That ratio is ill-posed: with the self-node ablated there is no self-model over which to compute Γ, B, or Π, so the denominator is undefined rather than small. The well-posed form is the matched comparison in 5.5. Separately, making τsystem-relative is a substantive governance choice — two systems would face different bars — and needs arguing rather than adopting as a simplification.
Even a validated threshold has ragged edges: PCI has a known dissociation in which ketamine-induced unresponsiveness groups with wakefulness. Expect the same.
12.3 Access requirements
Π requires read access to the internal self-representation. Ω requires read and write access, plus a matched control set. Γrequires identifying the context cover, an interpretive act with no procedure given. These are substantial and they exclude black-box assessment entirely.
12.4 The remaining requirements
- Independent measurement of each remaining input with stated reliability.
- At least one prediction about a system not used in the construction that does not follow tautologically. Revision 4 supplies a candidate — 5.5 — for the first time.
- Characterized false-positive and false-negative rates for the threshold.
Item 2 has moved from unmet to proposed and untested. Items 1 and 3 are unmet.
13. What to build next
Analytic revision has reached diminishing returns. The remaining requirements cannot be obtained by further editing of the equation, and Revision 5 should not be a more sophisticated formula.
1. Toy simulations. Four synthetic dynamical systems with known ground truth: no stable self-state; one stable attractor; multiple contextual attractors with compatible overlap; locally stable but globally fragmented. Check that Γ, h(B), Π, Ω, and the admissibility tests behave as intended. Sweep each input across its sign change; verify monotonicity, that the minimum produces the expected sharp drops, and that ⊥ fires where it should. A day of work, and the fastest route to hidden defects.
2. Ablation protocol. Self-node versus matched-node ablation on a system with an identifiable self-representation. This is the prediction; test it early. Expect the matching procedure for 𝒩 to be the hard part.
3. Adversarial construction. Deliberately build a system that maximizes C_fn while plainly violating the intended notion of functional selfhood. If it is easy, that is the next defect. If successive attempts become increasingly contrived, that is informative in the other direction and is the most interesting possible outcome.
4. Pre-register and test the Biàn/zì → Γ reduction (9.3).
5. Investigate attractor sets, resolving the identifiability blocker in 9.2 first.
6. Derive B bounds under exponential families without unimodality (5.7).
7. LMI/SDP contraction-metric search for the global stability result (5.2), with realistic expectations about dimension.
8. Reference implementation.
9. Only then, τ.
14. Standing
Revision 4 is well-formed: defined across its domain, non-compensatory, three-valued where governance needs three values, honest about its parameter count, testing three assumptions it previously asserted, and — for the first time — containing a constituent that could turn out to be false of a system satisfying all the others.
It is not yet empirically contentful. No simulation has been run. Most inputs have no instrument. The prediction in 5.5 is proposed, not tested.
The architecture is not trivial. Localizing self-models to contexts and separating within-context convergence from cross-context gluing is the right shape for a problem where a system can be coherent in every situation and still have no unified self. Adding centering closes the gap where consistency was standing in for selfhood. The binding gap supplies a principled zero where a ratio supplied none.
The remaining weaknesses are now mostly research questions rather than formal mistakes — which is where a theoretical proposal should get to before anyone runs an experiment.
The consequential open decision is Section 10, and it is not technical.
Appendix: corrections and declined recommendations
Errors corrected in this revision. Spectral radius below one gives local asymptotic stability, not the Banach conclusion; previous versions and one reviewer conflated them (5.2). The Gaussian derivation of B monotonicity requires an upward-opening well and, once corrected, collapses B into the ratio it was meant to generalize (5.7). Replacing coverage with worst-case gluing is a change of construct, not an optimization, and does not eliminate the tolerance parameter — it renames it (5.4). Tying δ_min to measurement resolution changes its semantics from ontological to epistemic (5.8). The ablation ratio C_fn(intact)/C_fn(ablated) is ill-posed (12.2).
Declined. F(x) = x − x³ was said to have fixed points at ±1. It does not: F(x) = x requires x³ = 0, and F(1) = 0 ≠ 1. The domain restriction to (−1,1) is still needed, but for attraction rather than uniqueness. Re-centering the scale so that λ_max ≈ 1 counts as optimal does not resolve the edge-of-chaos tension, because at that value there is no attracting fixed point to re-center around (5.3, 9.2).
Verification status. Every claim here is analytic. Limit behaviour, monotonicity, ranges, the counterexample in 5.2, and the corrected Boltzmann derivation in 5.7 were verified by inspection. No numerical simulation was run and no code executed. That is item 1 of Section 13.
No comments:
Post a Comment