QUESTION-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?

Why it is a question

Four conditions pull against each other, and each turn of the exchange added one.

1. Not inserted. “Organizational perspectives are patterns we attempt to identify in that machinery—not additional agents we insert above it” (T19), so “there is no PerspectiveState inside CorporateState”. This is CLAIM-418’s first part as a constraint on method.
2. Overlapping. Participation is not exclusive (“there is no requirement that ∑_p w_ep,t = 1”, T19). T22: “perspectives need not be mutually exclusive, and we shouldn’t have to decide in advance how many there are.” A partition into clusters cannot express this, and T21 noted that classical hLDA gives each document one path through its tree, which the nested HDP relaxes (LIT-002, NOTE-023). The reading confirms it: in hLDA two documents share a topic only by sharing the whole path above it (THEORY-969).
3. Not common cause. T18 §5: “a dynamical mode is not automatically a perspective. A high-variance activity pattern might merely reflect a shared reporting schedule.” T22: “correlated activity may be caused by common external events. Employees responding to the same incident don’t necessarily constitute one emergent perspective.” That is CASE-101’s first population, whose members “respond alike to a common input”, and QUESTION-063’s shared drive against coupling. A check filed with CASE-166 finds three cross-departmental modes among employees who never interact.
4. More than the graph. T22 §6: “Can inferred participation predict the effect of changing an authorization pathway, beyond what the authorization graph already tells us? … If an IBP detects an organizational mode but that mode adds nothing once communication and authority are modeled, we should be reluctant to identify it as a causally important perspective.” And: “We’d also need to allow the inference procedure to find no persuasive organizational perspectives at all.”

The record holds the parts. TERM-072 says its organizational sense “would be shown wanting … if sense 1 could not be told from coordinated feedback in general”. CLAIM-327 gives a candidate criterion: a collective variable is in the dynamics when members read it, condition on it and regenerate it. CLAIM-135 says recovery is reconstruction, identification and constitution, and identifiability reaches the second at most. CLAIM-154 and CLAIM-188 price naming a recovered unit, which an inferred number of features is. CLAIM-427 says that recovering modes in the framework’s own terms counts for nothing. What no entry asks is the combined question for a population of members whose higher-level organization is to be inferred.

The candidate methods

• hLDA and the nested HDP (T21): a topic hierarchy over the content of members’ messages (LIT-1542, NOTE-1176). T21 said not to “identify its topics directly with perspectives”, and to “allow the hierarchy to lose if it doesn’t provide useful explanatory compression”. The record already holds a nonparametric topic model fitted jointly with an author network (LIT-013’s Twitter-Network Topic Model, NOTE-008), which is the topics-and-relations combination T21 proposed. T21 also named the author–topic model (LIT-1546) and the mixed membership stochastic blockmodel (LIT-1553).
• A feature allocation under an Indian buffet prior (T22, developed at T23): which members hold which latent features, with their number left open (the review T22 cites, LIT-1556, NOTE-1172). T23’s dynamic beta–Bernoulli form is CASE-165, with its checks. T23 recommends fitting it to a mechanistic simulator that does not itself assume the prior (“Option A”), so that “feature discovery can be tested against a world that did not itself assume IBP-distributed employees”.

Either way, CLAIM-454 says what a recovered feature could amount to: at most a constituent of a perspective, whose organization lies in relations the allocation leaves open.

What the readings of the candidate methods show

All seven works T21–T23 cite were read on 2026-10-26.

• Only the feature allocation overlaps as condition 2 asks. hLDA’s overlap is nested: a member shares a theme with another only by sharing everything above it (THEORY-969). In the author–topic model and the blockmodel a member’s participation is a vector on the simplex, so more of one theme is less of another (NOTE-1180, NOTE-1179). The beta-process family is the one whose participation has no sum-to-one constraint.
• Inferred memberships are not identified. The author–topic model and the blockmodel each leave a continuum of changes of basis free, beyond the permutations their papers name (THEORY-972, THEORY-993; CLAIM-148’s note of 2026-10-26). So a comparison under condition 4 has to be made on predicted edges and intervention outcomes, never on the memberships themselves.
• Condition 4 needs extending the relational baseline. The blockmodel is the natural model “given the communication and authorization relations”, but as published it models only one static graph’s edges, with no dynamics or interventions; it would have to be extended to the intervention outcome. The topic models cannot pass T21’s own falsification test (identical messages, different authorization pathways), as T21 expected of “a content-only hLDA model”: their likelihoods see only words and authorship.
• Time and groups have established forms, with gaps. The dependent IBP lets features appear and fade over a covariate such as time without fixing their number, with each feature’s prevalence held fixed (THEORY-1006). The hierarchical beta process shares a feature library across groups with group-specific prevalences, but each member belongs to one fixed group (THEORY-1015). A beta-process hierarchy over overlapping and changing groups is not in the literature read, and is an open part of this question.
• An intermediate layer is still co-occurrence. Correlated latent-feature models put a layer of categories over the features (THEORY-1012), a candidate for something between a feature and a perspective. A category is a pattern of co-occurrence, so by CLAIM-454’s standard it is not a perspective either; and in those models a category can switch features on, never off.

What would count as an answer

A procedure that (i) recovers overlapping modes with their number inferred, (ii) separates them from modes a shared exogenous driver produces in a model with no interaction, and (iii) predicts the effect of intervening on the coupling (team records, delegation, membership, authorization pathways) better than an equally flexible model given the communication and authorization relations and no modes, on held-out interventions. A showing that no such procedure adds anything beyond the relational graph would answer it too, in the negative.

What would not answer it

• A factorization that fits trajectories, however flexible.
• Recovering the prior’s structure from data generated under that prior, which tests the inference, not the structure (CLAIM-203’s note of 2026-10-26). hLDA’s paper has the published example: eight of ten trees recovered from corpora simulated under its own prior (THEORY-1010).
• A mode named after the departments or roles that make it up.