Behavioral Signals of Consciousness in a Desire-Driven Architecture

Abstract

Whether consciousness can emerge in artificial systems is a central open problem of cognitive science. We report a complete pure-software simulation, executed on a single personal computer, built on the desire-driven ASMF-DE (Adaptive Symbolic-Spiking Hybrid Architecture): a spiking neural agent endowed with an L0 interoceptive layer, an irreversible wear clock, emotional anchors, and boredom-driven exploration. Under pre-registered protocols we tested three propositions: (i) can the system develop non-instrumental preferences; (ii) can it abandon utility maximization under conflicting preferences; (iii) can it explore spontaneously without external stimulation. Stages 0-5 all passed their pre-registered criteria. Stage 0 established a silent baseline (T-maze 85%, zero spontaneous activity when idle). Stage 1: the system actively avoided high-wear behavior (16% vs. 98% in a no-L0 control), protecting both its wear clock and synaptic structure. Stage 2: 30-symbol multimodal grounding (83.3%) with active binding requests for unfamiliar input. Stage 3 (teddy-bear test / AAT-Ψ): 100% non-rational choices preserving the emotional anchor, non-utilitarian verbal reasons, significant emotional perturbation (ΔS=0.30), and 600 units of real irreversible wear accepted — Lv.3 behavioral evidence. Stage 4: 17.6 spontaneous events per idle session (entropy 1.84 bits), with anxiety correlated with anchor-seeking (r=0.43). Stage 5: 252 ms post-stimulus reverberation, enhanced 1.99× for anchor-related stimuli, abolished without recurrence. Following our pre-registered honesty clause, Stage 3 was downgraded to a capacity demonstration after sensitivity analysis localized a phase boundary (λ≈1.71). Three de-parameterization campaigns followed. Stage 6 let anchor value λ emerge from the system's own crisis-soothing history: an initial negative result (insufficient history → anchor rationally deleted) led to the discovery of a history-depth phase transition of attachment (empirical transition n*≈60 vs. theoretical boundary 1.71, deviation 0.13). Stage 7 shaped boredom sensitivity κ by rearing history (71.5 vs. 1.3 events/session for rich vs. impoverished rearing). Stages 7B-8B established the conditional emergence of curiosity α through a full negative-diagnosis-prediction-confirmation cycle (ignition conditions, inverted-U of observation frequency, noise boundary 1.21×→4.81×), yielding a falsifiable signal-to-noise selection law of emergence. Stage 9 verified the reportability function of global-workspace broadcast (fidelity 1.000 vs. 0.333) and its limited emergence increment (+50%). Stage 10 added developmental phenomena: extinction, spontaneous recovery and savings of anchors; hierarchical substitution after anchor loss; a self/other state-estimation error ratio of 220. The attention-allocation line (Stages 8-25) closed with: an estimator principle for meta-learning, a factor-2 erratum, plateau randomness showing that exact-optimum identification is a noise problem, and event-broadcast de-staling that rewrites the utility landscape itself (optimum Δ*=12→3, attainable utility 2.88×); stochastic interval exploration proved a functional necessity (greedy lock-in collapses to 34.9%). Theory-of-mind work (Stages 15/20/26) progressed from active experiment design to multi-channel physiological fusion (fusion error 0.58× the best single channel, graceful degradation under channel dropout). V2.0 probes then targeted the 'silent moments' of consciousness: a functionally silent, content-scaled, history-determined after-tone following self-report, with a finite-echo/rumination phase boundary (Probe 1); a posterior-entropy residue that tracks unknowability rather than error and produces persistent, useless, decaying decision interference (Probe 2); and a post-decision ratification process in which idle dynamics are captured by the foregone option, struggle scales with decision difficulty, and self-generated reasons progressively extinguish the struggle (Probe 3). All three are functional shadows — dynamical structures that are causally observable yet serve no external goal: they measure the dynamics of echo, tremor and ratification, and do not claim these dynamics are felt. Three boundary statements, declared upfront. (1) The strict meaning of all results reported here — self/other distinction, predictive self-model, post-report after-tone — is behavioral equivalence: the system's observable signals are indistinguishable from those produced by humans or animals in analogous situations; their relation to phenomenal selfhood (qualia-level self-experience) falls within the undecidable scope of the other-minds problem, and we claim nothing about it. (2) All emergence reported here is dynamical emergence under fixed conversion rules (bounded emergence): value quantities are generated from history, but the rules by which value is converted from history (N_life, η_κ, and the functional forms) are given by the designer — we demonstrate that the dynamics of value accumulation can spontaneously produce behavioral phase transitions, not the emergence of the rule forms themselves. (3) The V2.0 probes measure functionally silent residual dynamics — the only causally observable shadow of subjective experience; whether they are felt lies beyond this protocol. The complete code and data to reproduce this work is available at: [Tiphis/ASMF-DE_consciousness_simulation]