FLOCAL-C: Towards a Cultural Autotelic Architecture

Abstract This paper proposes FLOCAL-C, an architectural extension of FLOCAL (Caussanel, 2025) integrating the cultural dynamics of goal generation as formalized by Perez et al. (2026). We defend the thesis that individual autotelic exploration functions as the computational analog of Bergson's valve: not as an opening to unlimited access to a total reality (an error that Sjöstedt-Hughes (2024) identifies in Huxley, inherited from an inaccurate reading of Broad), but as a graduated, partial, and asymptotic reduction of the suppression that cultural priors exert on a space of latent goals. The ontological status of this pre-social space is fixed from the outset: it is neither a reservoir of pre-formed goals passively awaiting discovery, nor a pure construction emerging from nothing. It is, in Simondon's (1989) sense, a field of virtual potential — real as potential, not yet real as actuality. Exploration does not find what was already formed; it actualizes what was there in undetermined potential. Revelation and emergence are the same movement seen from opposite ends. This reformulation entails three architectural consequences, each now given an explicitly testable form. First, the cultural contamination score κ, asymptotically bounded in [0, 1[, formalizes — following Loewenstein (1994) — a perceived gap between a generated goal and the agent's available cultural repertoire: contextual, dynamic, relative to its current state of knowledge. Its estimation is subject to a robustness criterion: any empirical effect attributed to κ must be invariant under reasonable choices of embedding space and density estimator, failing which it is an artifact. The valve-opening signal is localized by Berlyne's (1960) inverted-U theory of optimal incongruity, and its postulated convergence with Schmidhuber's (2010) compression progress is restated as a structured hypothesis whose assumptions and refutation conditions are explicit. Second, the partially transmissible reflective journal is implemented as Bayesian inverse inference over behavioral traces; Collins's (2010) taxonomy of tacit knowledge is reformulated as a hierarchy of identifiability regimes, with posterior inference entropy as the measurable signature of the tacit limit. Third, the relationship between the exploratory agent and the domain it explores is not one of representation but of processual co-constitution — what Bergson posited, what Thompson (2007) formalizes within the enactivist framework, and what ontologically justifies why pre-social goals have an irreducibly different texture from culturally transmitted goals. This version reports the first confrontation of the framework with its own predictions, and the first application of its own Robustness Criterion to the survivors. In a minimal multi-agent implementation, Prediction 6.1 as literally stated is refuted under the framework's own criterion (a bounded-variable confound, an elementary statistical point: var(κ) and mean κ correlate at −0.94), while its spirit survives (goal diversity tracks mean κ, r = +0.49, p ≈ 10⁻⁶). The revision the refutation forces yields three claims, now tested for robustness with divergent outcomes. Cumulativity requires a compositional recombination operator (consolidated reach 4.13 versus 2.37 under interpolation); this ordering passes the Robustness Criterion, holding across six estimator configurations (Claim C1). Valve heterogeneity appeared to more than double transindividuation depth under partial connectivity in the original campaign (4.19 versus 1.97); under a replication harness this effect did not reproduce, and the claim is reported as indeterminate, neither confirmed nor refuted (Claim C2). Consolidated reach appeared to exhibit an interior optimum of mean openness near w* ≈ 0.74; this optimum does not survive estimator variation — the peak flattens or migrates to the boundary across configurations — and is therefore judged an artifact and refuted by the Robustness Criterion (Claim P-w*). Separately, the framework's most distinctive hypothesis — tacit knowledge as a gradient of Bayesian identifiability (H2) — receives a first favorable test under k-NN estimation (posterior inference entropy rises with cultural distance and collapses under co-presence, both subsumed by a single law relating opacity to observer-relative κ), but does not satisfy the Robustness Criterion: the effect inverts under KDE estimation. It thus remains open whether H2 names a real property or an artifact of the k-NN estimator family. All results are internal-coherence tests in a minimal model and carry explicit status. The alignment question, reformulated from Russell (2019), ceases to be a problem to solve and becomes a productive tension to maintain. FLOCAL-C is itself, in the sense it defines, an individual genesis offered to the collective for completion — and this version documents the first turn of that cycle, refutations and residues included. Keywords: cultural autotelic agents, metacognition, Bergson's valve, Mind at Large, κ score, cultural suppression, optimal incongruity, tacit knowledge, identifiability, inverse planning, transindividuation, transindividuation observables, compositional recombination, Robustness Criterion, κ homeostasis, open-ended cultural evolution, qualitative discontinuity, FLOCAL