The claim
A9 §3, after Flack’s macaque work (LIT-1122, unread): “The critical step isn’t simply that an observer can compute a long-term average. The animals themselves appear to use information encoded in these more stable relationships to guide subsequent interactions.” A9 separated three stages:
- Statistical regularity. An observer finds that many fast interactions yield a stable aggregate.
- Member-accessible structure. Members recognize or estimate it and use it when choosing what to do.
- Self-reinforcing organization. Behaviour conditioned on it preserves, transforms or regenerates it.
“A slow variable is not just something that changes slowly. It is something whose history has become sufficiently stable and accessible that components can organize their future behavior around it.” A fanbase’s reputation (CASE-103) and a program’s standing (CASE-102) are A9’s examples: recruits, donors, visiting fans and stewards condition on the accumulated past, and their decisions make the future that sustains or changes it. “The time-averaged history has become a contemporary causal influence.”
The record holds a special case: CLAIM-277, where durable trust is the temporal regularity of endorsement and then directs endorsement. This claim says what that case has in general, and gives QUESTION-065 a candidate for a target the system supplies itself: what its own members condition on.
A9 §4 wrote the tests: timescale separation (τ_Z ≫ τ_X, not sufficient), predictive usefulness over a matched baseline (I(Z_t; X_{t+h} | B_t) > 0, which depends on the baseline), endogenous maintenance, and causal uptake under intervention (“giving new students different narratives about a university’s traditional rival”).
What it does not say
- Not that the variable individuates a collective. A9 §7: five slow variables sharing a referent “could be produced by largely independent mechanisms that happen to reference the same sports team.” Whether they form one collective needs a further test of integration, such as interventions on one changing the others through mechanisms internal to the collective.
- Not that the information is new. A9: if Z_t is computed from a history already in the baseline, “there is no additional Shannon information to gain”; its value may be economy or accessibility to bounded members.
- Not agency. A collective organized around a reputation need not be able to act.
Note of 2026-10-18: endogeneity without slowness
The exchange of 2026-10-17 re-derived this claim’s separation of slowness from endogeneity without citing it (A1 §1, A3, A4), and built it as constructions (CASE-116). Arm A, a slow register no member reads, is this claim’s “timescale separation (τ_Z ≫ τ_X, not sufficient)” made a construction. Arm B is the converse, which the claim lacked. Members read a register and rewrite it (Z_{t+1} = Z_t ⊕ N₁, with N₁ ~ Bernoulli(0.4)). Its second eigenvalue is 0.2, so it forgets on the members’ own timescale, and do(Z_t) changes the next value (0.4 against 0.6). So timescale separation is not necessary for the third stage either, for weak maintenance. A4’s limit: “This demonstrates only weak endogenous maintenance. If maintenance is defined to require robust long-term persistence, this counterexample no longer establishes the separation.” The claim is about slowly changing variables, so arm B does not cut against it. It says the test list should not read slowness as a requirement of endogeneity. CLAIM-331’s “slowness is a mark of a level and not its definition” says the same for levels.
Note of 2026-10-27: Flack 2017 read, and the third stage is not hers
Flack’s general statement (LIT-1630, read in NOTE-1216; THEORY-1061) changes how this claim’s attribution should be read. Stage two, members recognizing and using the variable, is what she calls apparent downward causation, with two refinements the claim lacks. Members condition on their own estimate, which may be wrong and need not be shared, and one member tuning is enough (§§1, 3). Stage one has no separate place: the variable is computed by the members from the start, “endogenous coarse-graining” (§2). Stage three, use regenerating the variable, is in neither of her lists of criteria (§1, §7). What reinforces itself in Flack is agreement among members’ estimates: “positive feedback induced by the downward causation itself” (§3). And where the claim says only the third stage makes an endogenous level, Flack says use alone is weak downward causation, and a new level needs her strong criteria: predictive, robust, used nearly universally, agreed on, with rising micro–macro mutual information (§1). So the three stages are the exchange’s synthesis of 2026-10-16, with regeneration where Flack has convergence. The claim’s content may still be right; it is not Flack’s. Flack 2012 (LIT-1122), which this paper defers to, stays unread.
Note of 2026-10-29: Flack 2012 and computational mechanics, read
Flack 2012, LIT-1122, is now read (NOTE-1351), so the summary’s “unread” and the last sentence of the note of 2026-10-27 no longer hold. Its words match A9’s “appear”: that the animals read and condition on the slow variable is what “the data suggest”, argued from earlier studies, and for relative power not yet worked out. Against the claim’s three stages:
- Statistical regularity is shown, for the signal network.
- Member accessibility is asserted, not shown.
- Regeneration is the paper’s §4 argument that fighting lowers after signals and so consolidates the relation. No measurement in the paper supports it.
So the paper this claim quotes is evidence for the first stage only.
Shalizi and Crutchfield, LIT-1126 (NOTE-1372), bear on A10’s offer of causal states as “the minimal history a collective must keep to predict its future”. The reading adds two limits to the slow state that offer names. The theorems need the observed process fixed in advance and stationary, and the authors call relaxing stationarity “unclear”, which a collective under turnover may need. And the causal states are minimal only for the whole future: for a finite horizon a coarser state can predict as well, so “the” slow state depends on the horizon of prediction.
The claim stays Proposed.
Note of 2026-10-29: the two Flack 2006 papers, not reached
Two of the earlier studies Flack 2012 argues the second stage from are now registered, and both are Unavailable: no lawful copy was reached.
- Flack and Krakauer 2006, “Encoding power in communication networks”, LIT-1840, which builds the power measure from the signalling network and relates it to how others solicit and accept intervention. Whether it shows members conditioning on the consensus rather than on the signals themselves turns on its methods.
- Flack, Girvan, de Waal and Krakauer 2006, “Policing stabilizes construction of social niches in primates”, LIT-1828, the knockout of policing that Flack 2012 rests part of its argument on, and the one intervention in the literature this claim draws on.
So the second stage still has no evidence the record has read. One thing can be said from the descriptions alone. The knockout removed the policers, an input to the power structure, not members’ access to a slow variable. Even a clean result would bear on the third stage, regeneration, more directly than on the second, and it is not the intervention the defeated_if names.
The claim stays Proposed.
Note of 2026-10-30: what holds it back
No case in the record shows the second or third stage for a real collective: Flack 2012 (NOTE-1351) shows the first stage only, the two Flack 2006 papers that argue the second are Unavailable, and the intervention on members’ access that the defeated_if names has not been run or read. Flack 2017 (NOTE-1216) also puts weak downward causation at use alone, against the summary’s “only the third makes the variable an endogenous level”, and the claim has not replied.