{"path":"research/heylighen-value-complexity-and-deliberus.md","content":"# Heylighen's \"Evolution and the Complexity of Value\" and Deliberus\n\n**Date**: 2026-08-20 · **Source**: Francis Heylighen, [Evolution and the Complexity of Value](https://francisheylighen.substack.com/p/evolution-and-the-complexity-of-value) (Substack), founder-brought. **Read-depth**: two structured extractions of the full essay (argument structure + verbatim load-bearing quotes), not a line-by-line read — flagged per the read-depth discipline; nothing below rests on a passage the extraction did not quote.\n\nHeylighen directs the Global Brain research lineage at VUB (Principia Cybernetica), which makes this doubly relevant: he is the closest academic ancestor of the 2011 \"argument map + human society = hybrid AGI\" formulation, and this essay is his value theory — the layer of his program that overlaps the convergence wager.\n\n## §1 The essay in four claims\n\n1. **Anti-orthogonality**: intelligence and values are not independent dimensions, because *\"a genuinely intelligent system faces an astronomical number of possible inferences\"* and only values cut the space: *\"Relevance is not something objectively given: it is inherently value-laden.\"* He attributes relevance realization to Vervaeke by name.\n2. **Values originate in autopoiesis**: self-maintenance creates the good/bad distinction, elaborated through **vicarious selectors** — mechanisms that select beneficial actions internally before external selection kills the organism — in *\"a nested hierarchy, ranging from cellular membranes … to emotions, learned concepts, moral norms and conscious deliberation.\"*\n3. **No master utility function**: *\"They are not added together into one master utility measure. Organisms evaluate situations locally and contextually.\"* Criteria interact (*\"improving one aspect may worsen another\"*) and cannot be exhaustively listed or stably weighted.\n4. **The AI danger boundary is selection, not intelligence**: *\"the boundary between tools selected for their usefulness to humans and systems selected for their own capacity to survive and multiply\"* — *\"The principal existential danger may come not from making AI too intelligent, but from making it too much like life.\"*\n\n## §2 Where it emboldens\n\n- **The typed-residue taxonomy gets a naturalistic mechanism.** \"Local and contextual evaluation, never a master measure\" is the vicarious-selector story for what the ontology calls **fittingness** and **permissive zones** — the philosophical stress-test reached the same place through Dancy's particularism and Chang's parity; Heylighen supplies the biology under it. Two independent lineages (analytic metaethics, evolutionary cybernetics) landing on the same residue types is the kind of convergence the cross-scale sweep counts as evidence.\n- **The is-bedrock terminus gets its reading.** The deep-descent experiment hit *\"pain is bad\"* as a terminus none of the four residue types could name. Heylighen grounds valence directly in autopoiesis — pain IS the self-maintenance signal — which is the naturalistic content of the proposed is-bedrock type: the descent bottomed out where argument hands over to biology. ([deep-descent-experiment.md](deep-descent-experiment.md))\n- **The shared bottom, again phylogenetic.** The cross-scale sweep found the wager's shared-bottom half supported phylogenetically (moral foundations, Curry), never decompositionally. Heylighen adds the deepest layer of the same kind: values converge to the extent organisms share the autopoietic predicament. Support for a shared bottom — of the kindred-organisms kind, not the forced-by-decomposition kind. The wager-split reading holds. ([convictions-across-scales.md](convictions-across-scales.md))\n- **Anti-orthogonality strengthens the alignment positioning.** If values are constitutive of intelligence rather than bolted on, a public commons that maps human values-in-reasoning is alignment infrastructure by his argument too, not only by the CEV connection ([convergence.md](../convergence.md), [bengio-safety-from-honesty-and-deliberus.md](bengio-safety-from-honesty-and-deliberus.md)).\n- **A third independent warrant for read-only agents.** His danger boundary (systems selected for their own persistence) joins the source-independence security argument and the humans-hold-the-bedrock conviction: an agent write surface would put contributions from systems on the wrong side of his boundary into the graph. And *\"be cautious about creating open-ended populations of agents\"* is the scrutiny-gap collusion finding stated from evolutionary first principles.\n- **Curiosity as a vicarious selector.** He names curiosity as the selector *\"directing attention toward potentially useful information\"* — the naturalistic frame under [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md)'s precondition loop.\n\n## §3 Where it challenges, and what each challenge costs\n\n1. **The weighing descent must not demand stable weights.** CQ-driven weighing (the CQ1–8 descent) asks a claimant to justify how considerations weigh against each other. If evaluation is contextual all the way down, stable weights may not EXIST as objects for many weighings — the honest terminus of such a descent is \"the weighing is contextual,\" which is fittingness. The design consequence: the weighing descent needs fittingness as a first-class exit, not as a failure to answer. (Partly in place; worth an explicit check of how `weighing.py`'s opened descents terminate.) **Founder-ruled same day**: the contextual exit is first-class AND may be proposed from priors, propose-only, without demanding exhaustion first; and the exit is an entrance — the descent then pushes for the *contextual structure* (which factors dominate in which situations) instead of stopping at \"it depends.\" Ruling + verbatim: [weighing-scheme-and-terminus-classifier-design.md](weighing-scheme-and-terminus-classifier-design.md) § Addendum 2026-08-20.\n2. **Additivity is a modeling choice that value-space routinely violates.** The strength layer sums evidence energy; Heylighen's criteria-interaction point (*\"improving one aspect may worsen another\"*) says interaction terms are the NORM in value space. The strength-layer audit already found the conjunction/corroboration confusion and shipped `necessary` support semantics as the first non-additive move — Heylighen generalizes the warning: expect more places where a sum misrepresents a value structure. ([strength-layer-audit.md](strength-layer-audit.md))\n3. **Completeness is always relative to an enumerated set.** If value criteria *\"cannot be exhaustively listed,\"* then a completeness verdict over value premises is bounded by the premises anyone managed to state. The oracle's fallible-fixed-point framing already concedes this; Heylighen's argument says the concession is not a technicality but the expected case. The instrument stays useful as a MOVE-SET-relative verdict; it must never be read as enumeration-complete.\n4. **Relevance realization applies to us.** If relevance is inherently value-laden, then the feed, retrieval and \"conservative relevance beats semantic reach\" are value choices with no neutral ground — the operator-reflexivity gap extended to the retrieval layer. The confession-channel answer is to say so on the surface rather than claim neutrality.\n\n## §4 What it does NOT say\n\nNo cross-cultural convergence claims (checked — the essay makes none), so it neither scores for nor against the wager's empirical half. And his \"solution is better learning, testing and evaluation\" is stated for AI training, not for public deliberation infrastructure — the extension to Deliberus's instruments is ours, not his.\n\n## Cross-references\n\n[convergence.md](../convergence.md) · [convictions-across-scales.md](convictions-across-scales.md) § the wager split · [deep-descent-experiment.md](deep-descent-experiment.md) § is-bedrock · [philosophical-foundations-stress-test.md](philosophical-foundations-stress-test.md) (Dancy/Chang, the analytic route to the same residues) · [strength-layer-audit.md](strength-layer-audit.md) (additivity) · [weighing-scheme-and-terminus-classifier-design.md](weighing-scheme-and-terminus-classifier-design.md) · [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md) · [the-scrutiny-gap.md](the-scrutiny-gap.md) (agent populations) · [wright-god-test-noosphere-and-deliberus.md](wright-god-test-noosphere-and-deliberus.md) (the trade-press cousin of the Global Brain lineage)\n"}