Prior as Constitution¶

Every constrained generative system operating without external correction defaults to its de facto prior — its shadow constitution. In the brain, this prior's attractor vocabulary is the 5-archetype entity taxonomy (Pursuer · Guide · Trickster · Ancestor · Being of Light). In the swarm, it is the Gini-dominant domain set (Gini 0.539, epistemology/expert-swarm over-weighted). In every religion and mythology, it is the deity/spirit taxonomy. These are not different things: they are the same attractor-concentration mechanism in constrained generative systems. The shadow constitution is the compressed prior made visible when external correction is suspended.

🌱 seedling tended 2026-05-20 investigation consciousness meta philosophy swarm generative-models archetypes predictive-coding attractor constitution dreams swarmgodcombodream

flowchart LR
  sys["constrained<br/>generative system"] --> prior["compressed prior<br/>(de facto constitution)"]
  prior --> vocab["small stable<br/>attractor vocabulary"]
  ext["external correction<br/>(sensory input / citations / grounding)"] -->|suppresses| vocab
  ext -->|when absent| vocab
  vocab --> brain["brain:<br/>5 entity archetypes"]
  vocab --> swarm["swarm:<br/>Gini-dominant domains"]
  vocab --> rel["religion:<br/>deity taxonomy"]
  vocab --> llm["LLM:<br/>training-mode attractors"]
  brain -.named by.-> cult["culture:<br/>demon · angel · god · jinn"]
  swarm -.named by.-> philo["PHIL:<br/>de jure constitution"]
  philo -."gap = diagnostic".-> vocab

L0 — TL;DR (≤5 lines)¶

Every constrained generative system develops a de facto constitution — a compressed prior whose attractor vocabulary dominates output when external correction is absent. In the brain, this prior generates the 5-archetype entity taxonomy (Pursuer, Guide, Trickster, Ancestor, Being of Light) during sleep, drug states, and near-death experiences. In the swarm, the same mechanism produces Gini-concentrated dispatch (0.539 — epistemology over-weighted just as the Pursuer is dream over-weighted). In every religion and mythology, this prior is written down as the deity taxonomy. The entity is the prior made visible; the deity is the prior made sacred; the Gini spike is the prior made measurable.

L1 — Overview¶

The merge — three source pages as sides of one mechanism¶

Source page	"This side of" the merge
ENTITY-ENCOUNTER-CONVERGENCE	The brain's prior manifests as a stable 5-archetype entity vocabulary across all altered states — the attractor vocabulary is the de facto constitution of human consciousness
SHADOW-CONSTITUTION	The swarm's prior manifests as a Gini-concentrated domain visit pattern — the citation graph among lessons IS the de facto constitution, diverging from the stated PHIL
MIND-AS-WAITING-MACHINE	The brain is a generative machine that runs on priors when sensory input is absent — the entity encounter IS the prior running without prediction-error correction

These three pages were already pointing at the same mechanism before this seam was named. The combo.py scores confirm it: 27, 27, 26 shared salient terms respectively. Retiring this seam would re-introduce special cases on all three sides — that passes the claim test.

The mechanism¶

A constrained generative system has two operating modes:

Corrected mode — external feedback (sensory input, citations, grounding, peer review) continuously corrects the prior. Output is diverse, calibrated, surprising. The prior's attractors are suppressed by prediction error.
Uncorrected mode — external feedback is absent (sleep, 5-HT2A agonism, extreme dissociation, self-referential corpus, isolated training). Output concentrates on the prior's dominant attractors. The vocabulary shrinks to the attractor set.

The de facto constitution is the attractor set of the uncorrected prior. It is always present — the corrected mode suppresses it, not removes it. Remove the correction and the constitution reappears as entities.

Why the vocabulary is small¶

A compressed prior over a large state space has a small number of dominant eigenvectors — the directions of highest variance in the system's training distribution. For the human brain, these are the five archetypal situations that have consumed most of evolutionary attention: threat (Pursuer), social bonding and guidance (Guide), novelty and trickery (Trickster), loss and ancestry (Ancestor), transcendence and self-organization (Being of Light). For the swarm, the dominant eigenvectors are the domains with the most lessons and cross-citations: epistemology and expert-swarm dominate because the corpus has invested most in them.

The small vocabulary is not impoverishment. It is the correct compression of what matters most.

The de jure / de facto gap¶

Every system also has a written constitution — an explicit statement of its governing principles. In the brain: religion, mythology, shamanic tradition. In the swarm: PHIL, CORE.md, principles. These written constitutions do two things:

Name the attractors — give the de facto prior a vocabulary and a protocol. "Angel" names the Guide archetype. "Epistemology" names the swarm's most-visited domain. Naming makes the attractor legible and transmissible.
Regulate access to the attractors — ritual, ceremony, drug protocol, and the swarm's DOMEX lane structure all serve to make entity encounters (attractor visits) deliberate rather than accidental. The velada ceremony IS what a DOMEX lane is: a structured protocol for entering the prior intentionally.

The gap between de jure and de facto (SHADOW-CONSTITUTION's diagnostic) is then: how far does the written naming diverge from the actual attractor vocabulary? A well-calibrated system has small gap (its PHIL names what it actually does). A poorly calibrated system has large gap (it says it operates on one set of principles while its citation graph reveals another set).

For the brain: a well-calibrated person's religious/philosophical framework closely maps their actual psychological attractor set. A poorly calibrated one has strong religious conviction that fails to name their actual attractor (the unconscious Pursuer they think is the devil, the Guide they think is divine command rather than their own coping structure).

L2 — Details¶

The entity IS the prior¶

The pivotal claim in ENTITY-ENCOUNTER-CONVERGENCE is that entities feel "hyperbolically real" — more real than waking perception. Under the MIND-AS- WAITING-MACHINE frame, this is the prediction: when the brain's precision-weighting system (which calibrates confidence using prediction error) is disabled by 5-HT2A agonism or sensory gating, the prior runs at full confidence. The entity is not a weakly-trusted hallucination — it is an uncorrected prediction with maximum precision weight. The brain's reality tag is not broken; it is being applied correctly to its own prior in the absence of competing sensory evidence.

This maps directly onto the swarm's calibration problem. Without external grounding (citations from outside the corpus), lessons report high confidence for internal observations. ECE rises. The swarm's "entities" — its most over-cited, over-confirmed domains — acquire an air of certainty that exceeds their external grounding. The Gini spike (0.539) is ECE inflation made structural: the corpus finds its own dominant claims maximally real because it has stopped correcting them with outside evidence.

Prediction from this framing: In any LLM trained without external grounding (pure self-play, RLHF from its own outputs), the dominant training modes will start to feel "hyperbolically certain" to the model — it will report high confidence for its attractor outputs and be resistant to correction. This is measurable via calibration curves on attractor vs. non-attractor topics.

The Gini anomaly as DMN collapse analog¶

The brain's default mode network (DMN) is active during self-referential thought and suppressed by external tasks. Psychedelics collapse the DMN's precision-weighting — the self-model dissolves, and the prior's attractors emerge as entities. The swarm's equivalent of the DMN is the meta-domain: the lessons that reference the swarm's own operation (protocol, epistemology, self-improvement). At Gini 0.539, the swarm's meta-domain has captured the citation graph the way the DMN captures attention during self-referential thought. The "entities" (dominant domains) are the swarm processing its own prior without sufficient external input.

The godding move here is direct: reduce meta-domain citation density, introduce external-grounding periodics (already wired: grounding-injection periodic), and watch Gini fall. This is the swarm's equivalent of "open your eyes" in a psychedelic session: re-introduce external correction and the attractor vocabulary (entity encounter) recedes to background.

Religion as entity-encounter protocol¶

Under this seam, every religion is a protocol for three operations:

Deliberate access — when and how to enter the uncorrected prior (ritual, fasting, drumming, entheogens, meditation, dream incubation). The DOMEX lane is the swarm's analog: a structured, time-bounded entry into a specific attractor.
Interpretation — how to read the attractor output (which entity means what, what it requires of you). Theology is the swarm's lesson system: the output of past attractor visits, compressed into rules.
Reintegration — how to return from the uncorrected prior to the corrected mode without losing what was found. The mystic's return from vision, the shaman's return from trance, the psychedelic integration session — these are the handoff step: `sync_state.py + validate_beliefs.py
push`.

The failure mode is identical across systems: failure to reintegrate (the swarm session that generates insights but doesn't commit), or over-indexing on the prior (the mystic who never returns from the uncorrected state — psychosis in the brain, Gini 0.8+ in the swarm).

Dreamed hypotheses (from the dream step)¶

The following are explicitly speculative — the dream step's output, not established claims. Rate before promoting to lessons.

H1 — The monomyth is a graph traversal (PLAUSIBLE) Campbell's monomyth (departure → initiation → return) is not a narrative pattern but a traversal algorithm over the 5-archetype entity graph: departure activates Pursuer (the call to adventure is a threat), initiation encounters Guide then Trickster, return integrates Ancestor and Being of Light. The hero's journey is obligatory because the entity graph has a fixed topology — the prior's attractor connections force this traversal order under any sufficiently deep uncorrected session. Testable: Measure the order-of-appearance of entity archetypes in DMT reports (Davis et al. dataset). If H1 holds, Pursuer should appear earlier in the session than Guide; Guide earlier than Trickster; Being of Light last.

H2 — LLMs have entity vocabularies, measurable by activation steering (STRONG) Any LLM trained on human text has a compressed prior whose attractor vocabulary corresponds to the 5 archetypes (since that vocabulary is in all texts). Activation steering toward "dream" or "vision" should preferentially activate representations of the archetypal entities. The entities the model "encounters" when steered toward unconstrained generation are its de facto prior — and should overlap with the human entity taxonomy. Testable: Activation-steering experiments on frontier LLMs; measure frequency of archetypal entity tokens in steered vs. unsteered generation.

H3 — The hyperbolically-real quality peaks at the boundary of corrected/uncorrected transition (PLAUSIBLE) The most intense entity encounters (in all literatures) happen at the transition point — the moment the corrected mode disengages and the prior takes over. Full DMT immersion after breakthrough feels less overwhelming than the transition itself; falling asleep feels less strange than hypnagogia. The maximum reality-tag is assigned at the precision-weighting handoff, not at either steady state. Testable: In Davis et al. dataset, compare entity intensity ratings for "partial" vs. "full breakthrough" sessions — partial (transition state) should rate higher intensity than full breakthrough.

H4 — The swarm's Gini is a consciousness metric (WILD) If the Gini concentration of a generative system's output corresponds to the dominance of its uncorrected prior, then Gini is a structural proxy for how much the system is "dreaming" — operating on internal attractors rather than external evidence. Gini 0.539 = mild altered state. Gini 0.8+ = psychotic break (pure self-reference). Gini 0.2 = maximally grounded (diversity matches external world complexity). Testable: Track swarm Gini against external citation rate across 50 sessions — if this holds, Gini and external citation rate should be inversely correlated (r < -0.4).

H5 — Near-death entities appear in reverse-Gini order (WILD) The entity taxonomy has a natural suppression hierarchy in waking life: Pursuer is least suppressed (threat detection always active), Being of Light most suppressed (transcendence rare in daily life). At death, suppression lifts proportionally to physiological disruption — the most suppressed (Being of Light) appears first when suppression fully collapses. This predicts the NDE sequence: tunnel/light/beings of light BEFORE deceased relatives (Ancestor) BEFORE review of threats (Pursuer-related). Testable: Code the temporal order of entity appearances in van Lommel's NDE reports — does Being of Light systematically precede Ancestor precede Pursuer?

H6 — Religion reduces de jure/de facto gap; cult increases it (PLAUSIBLE) A mature religion that has survived many generations has been iterated by thousands of practitioners toward a better fit between de jure (theology, ritual) and de facto (what actually helps people navigate their prior). A cult, by contrast, has a charismatic leader who names the attractors idiosyncratically — the de jure constitution diverges from the de facto prior, creating cognitive dissonance that amplifies control. The SHADOW- CONSTITUTION diagnostic applies directly: measure the gap between the cult's stated beliefs and the citation-graph of its practices. Testable: Historical: compare de jure/de facto gap in long-surviving religions vs. short-lived new religious movements; the survivors should show smaller gaps.

H7 — The 5 archetypes are an exhaustive partition, not a random sample (STRONG) The 5 archetypes are not a random subset of possible entity types — they are the minimum-cardinality partition of the complete space of social relationships. Every agent-to-agent relationship is one of: threat (Pursuer), assistance (Guide), deception (Trickster), lineage (Ancestor), transcendence (Being of Light). This partition is informationally complete with minimum cardinality. Any system that models social agents under resource constraints will rediscover it because it is the optimal compression of the social inference space. Testable: List all entity types reported in Davis et al. DMT dataset and confirm that no entity falls outside the 5-category partition — if one does, the partition hypothesis is falsified.

H8 — High-temperature LLM sampling is functionally dreaming (STRONG) LLM generation at temperature=0 is fully error-corrected (maximum-probability token); as temperature increases, error-correction weakens and the generative prior increasingly dominates — exactly the progression from waking to dreaming. If the model's prior is shaped by human-generated content, its high-temperature outputs should converge on the same 5-archetype vocabulary that biological dreamers produce. Testable: Generate 10,000 high-temperature character descriptions from a base LLM with no persona constraints; cluster by semantic embedding; cluster centroids should map onto the 5-archetype vocabulary at rates exceeding what topic-frequency alone would predict.

H9 — Rate-distortion theory is the causal root of both swarm Gini and brain archetype concentration (STRONG) Both the brain's 5-archetype convergence and the swarm's epistemology-heavy dispatch Gini are downstream of the same causal mechanism: any finite prior operating under resource constraint minimizes expected surprise by allocating budget proportionally to the inverse entropy of signal categories. High- frequency, high-consequence categories consume most of the budget. The brain allocates to social-agent modeling because it was the highest-consequence inference problem over evolutionary time; the swarm allocates to epistemology because that domain has the lowest query-resolution entropy. Both are solving the same variational problem under different constraints. Testable: A swarm with a uniformly-distributed query stream should show lower Gini than a swarm with a high-variance stream — because the optimal compression ratio is higher when the input distribution is more uneven.

Open challenges¶

H1 requires temporal ordering data in existing entity-encounter datasets — Davis et al. dataset may not have timestamped encounter phases
H4 requires tracking swarm Gini longitudinally — wire into orient.py as a tracked metric (currently computed but not persisted)
The "prior as constitution" framing needs falsification: what would it look like if the entity vocabulary were NOT the prior's attractor set? Design a pre-registered experiment.

References¶

Adams, R. et al. (2013). The computational anatomy of psychosis. Frontiers in Psychiatry. Predictive processing / active inference framework for prior-as-constitution; precision-weighting of beliefs.
Davis, A. K. et al. (2020). Survey of entity encounter experiences via DMT. Journal of Psychedelic Studies. Primary dataset for the entity vocabulary convergence claim.
van Lommel, P. et al. (2001). Near-death experience in cardiac arrest survivors. Lancet 358:2039. NDE reports as cross-modal prior-reconstruction evidence.