nucleation

CLAIM-536: A categorical model of perspectives and an LDA-style model differ in where the mixture lives: the first draws every act's perspective from one fixed categorical distribution, the second gives each episode its own mixture, drawn from a Dirichlet, and draws each act from it; the Dirichlet is a distribution over mixtures, not the perspectives themselves, and with symmetric concentration a over K perspectives two acts in one episode share a perspective with probability (a + 1)/(Ka + 1) against 1/K across episodes, so the concentration runs from one perspective per episode, as in persona selection (a → 0), to a fresh independent draw for every act (a → ∞)

4 min read

Properties9
StatusActive
Rolethesis
Warrantdemonstrated
Tagsprobabilistic-modeling, mathematics
Linelearned-structure, formal-apparatus
Bears onQUESTION-276 — How can overlapping higher-level perspectives be identified from the activity of the members who realize them, without being inserted into the model and without their number being fixed in advance, and told apart from coactivity that a common cause produces; and does a structure so inferred, such as a feature allocation under an Indian buffet prior or a nested topic hierarchy, predict an organization's responses to interventions on its records, membership and authority better than its communication and authorization relations alone?, QUESTION-370 — Do the concepts a language model's workspace makes reportable across an Assistant episode share an episode-level mixture, so that an LDA-style admixture over J-lens readouts predicts their co-occurrence better than one persona selected per conversation and better than independent draws per position; does intervening on the inferred mixture change which concepts reach report while leaving the consumers that bypass the workspace unchanged; and if two models matched on persona distribution and mixture differ under such interventions, whose will-organization is that difference: the simulated Assistant's, the network's or the deployed system's?
GroundsLIT-1948 — Latent Dirichlet Allocation
ComplementsCLAIM-543 — What separates a language model's distribution over personas from a distribution over organizational perspectives is not whether the latent model is hierarchical, since a mixture or an LDA-style admixture already is, but whether some latent components condition which others are available or consequential: a model whose components are exchangeable given their weights is organizationally flat however many levels its prior has, and dependencies among components count as organization (organization as relational arrangement) only where interventions, not fit, support them
Cited as “complements” byCLAIM-540 — The persistence a Dirichlet–categorical hierarchy gives a run of selections is an observer's learning about a fixed mixture, not a change in what selects: under a uniform prior an observed first choice raises the predicted chance of repeating it from 1/2 to 2/3, while the same choice forced leaves it at 1/2, so the hierarchy's reinforcement is not evidence of self-formation or of a maintained will-organization, which need the mixture itself to change under the choice

Graph View

The claim

Two models of which of K perspectives governs each act y_{d,n} of an episode d:

  • H1, categorical. z_{d,n} ~ Cat(π) for one fixed π shared by every episode.
  • H2, LDA-style admixture (Blei, Ng and Jordan, LIT-1948). Each episode draws its own mixture θ_d ~ Dir(α), then z_{d,n} ~ Cat(θ_d). In smoothed LDA each topic’s distribution over outputs is an independent draw from one shared exchangeable Dirichlet, Dir(η) (§5.4), which is not the one at issue here.

What separates them is the per-episode mixture, as the owner said. The owner’s “the topics … are the dirichlet” names the wrong thing. The Dirichlet is a distribution over mixtures. The perspectives are the categories each mixture weights. A12’s parenthesis has the repair: “Strictly speaking, θ_d is the topic-mixture vector drawn from the Dirichlet; the topics are the categories it weights.”

Agreement within and across episodes. With α = (a, …, a), the moments of the Dirichlet give E[θ_k²] = a(a + 1)/(Ka(Ka + 1)). So

P(z_{d,1} = z_{d,2}) = Σ_k E[θ_k²] = (a + 1)/(Ka + 1),

while for acts in different episodes, whose mixtures are independent, P(z_{d,1} = z_{d’,1}) = Σ_k E[θ_k] E[θ’_k] = 1/K.

  • As a → 0 the within-episode agreement tends to 1: one perspective per episode. That is the mixture of unigrams, and persona selection read as selection.
  • As a → ∞ it tends to 1/K: every act an independent categorical draw, A12’s H1.

So the owner’s H1 has two readings, one draw per conversation or one per act, and they are the two limits of H2’s concentration. H2 is the family between them. The contrast is testable by within-episode against between-episode co-occurrence (QUESTION-370).

Checked by simulation (400,000 episodes each, 2026-11-02): K = 3, a = 1 gives 0.499 within against 0.5 predicted, and 0.3335 across against 1/3; K = 5, a = 0.2 gives 0.6002 against 0.6; K = 4, a = 10 gives 0.270 against 0.268.

What it does not say

  • That the hierarchy is what A12 said it is. A12 credited the 1/2 against 1/3 agreement (K = 3, a = 1) to “the hierarchy alone”. A single uncertain mixture shared by both draws, a Pólya urn with no hierarchy, gives the same 1/2. What makes LDA hierarchical is many episodes, each with its own θ_d, sharing α. Its signature is the within-against-between contrast, not the within-episode figure alone.
  • That an admixture is a feature allocation. CASE-165’s beta–Bernoulli model lets a member hold several features at once, and CLAIM-454’s “feature allocation” is that sense. An admixture puts one component behind each act. Which H2 the record means matters for QUESTION-276’s overlap requirement.
  • That an H2 fit is organization. An LDA’s components are exchangeable given θ_d (CLAIM-543), and its reinforcement of a run of choices is the observer’s learning (CLAIM-540).

Note of 2026-11-03: LDA and the Assistant Axis, read

The account of LDA matches the paper (NOTE-1519), which itself names the one-topic-per-document comparison (a “Dirichlet-multinomial clustering model”, p. 997; the mixture of unigrams, §4.2). The agreement formula was re-simulated. Two refinements: the symmetric a → 0 limit is the mixture of unigrams with uniform p(z), where the paper’s has a free p(z), the limit of Dir(sπ); and the a → ∞ limit is the paper’s unigram model once z is marginalized, so per-act independent personas are unidentifiable from one output distribution. And the Assistant Axis (NOTE-1505) finds the persona following the latest user message within a conversation (R² 0.53–0.77 against 0.10), neither limit: the switching model this claim’s defeater names. So persona selection read as one persona per conversation does not fit what either source describes. The derivation stands.

Backlinks