Contribution
The paper gives a bottom-up mechanism by which a group of minimal deciders becomes a unit that decides in the same minimal sense. Before, each module of strongly coupled units settles on a convention and cannot change it. Afterwards it switches as a block when the other modules pull it. The mechanism changes only when units decide, never what they decide, and each change is paid for by an immediate individual gain. The authors offer it as a solution to what v2 calls “the hard problem of transitions in agency”: a transition that is both “systematic” (driven by individual incentives) and “meaningful” (it changes outcomes).
Key insight
A decider that alternates between being sensitive but inactive and being active but insensitive is held in place by its strongest partners only while they are acting during its sensitive phase. If the strongly coupled partners synchronise, they all become sensitive at the same moment, the internal pull vanishes for that moment, and the only influence left is the weak pull from outside. “Synchronisation changes the ‘attention’ from themselves to their environment - from the ingroup to the outgroup.” The same strong couplings both reward synchrony now and trap the group later, which is why individually motivated synchrony finds exactly the groups whose joint switching pays.
Assumptions
- The game. A symmetric modular coordination game, Eq. 1: c_ij = 1 within a module (⌊(i−1)/k⌋ = ⌊(j−1)/k⌋), 0.01 across; N = 300, k = 10. Energy E(t) = −Σ_ij c_ij s_i(t) s_j(t) (Eq. 2). There are 2^30 local optima and two global optima. Constraints are consistent, so all can be satisfied at once, and modules do not overlap (footnote 2).
- The unit (v2, Eq. 3). s_i(t+1) = γ_i(t)(I(∂U_i/∂s_i) + ε_i(t)), with ∂U_i/∂s_i = Σ_j c_ij s_j(t) (Eq. 4), I(x) = max(min(100x, 1), −1), ε uniform in [−0.01, 0.01], and the decision cycle γ_i(t) = (sin(ωt + θ_i) + 1)/2, ω = 0.01. (v1’s Eq. 3 had a decay term s_i(t)(1 − τ) with τ = 0.1 and γ_max = 0.1.)
- The timing rule (Eq. 5). θ_i(t+1) = θ_i(t) + ν Σ_j (π/4) c_ij sgn(s_i s_j) sin(θ_j − θ_i), ν = 3·10⁻⁵. Phases attract when the pair’s states agree with the sign of the coupling and repel otherwise. It is derived in Appendix A from a cycle-averaged pair utility U_ij ≈ c_ij sgn(s_i s_j)(π/2 + (π/4) cos(θ_i − θ_j)), on the approximation that s_i is nearly sinusoidal with constant sign over a cycle. The rule is “loosely analogous” to Kuramoto-type entrainment; it is a gradient on payoff, not a physical coupling through the units’ outputs.
- The evolutionary version. Adaptive dynamics under strong selection, weak mutation: each unit is a homogeneous population with a heritable θ, mutations drawn N(θ_i, σ) with σ = 0.5, fixed if the focal population’s average fitness over a cycle rises.
Key results
- Without entrainment (Fig. 4). Each module reaches internal agreement fast; modules agree with each other with probability about 0.5. “In 50 independent simulations, the global optimum is not found.” Units “decide many times” but “always re-make the same decision”.
- With entrainment (Fig. 5). Phases synchronise within modules; the “waking-up” of modules is “quite sudden (and less than t=20,000 for most modules)”, at different times for different modules. With the learned phases, 50 runs from random states “all find one of the two globally optimal solutions”. Blocks of ten switch as one, “acting like a single macro-scale unit”.
- The phases move before the benefit (Fig. 5B). “The theta values of a module change systematically in the direction of increased synchronisation even before a collective exhibits collective action.”
- Competency rises with entrainment (Fig. 6). Phases sampled every 1,300 steps and tested in side runs: without entrainment the system mostly lands on the worst local optima (half the countries each way); after about 20,000 steps every side run reaches a global optimum.
- Specific, not global, synchrony (Fig. 7A). With synchrony imposed across all units, the system “does not exhibit transitions or the increased energy minimisation behaviour”. “Specifically-organised synchronisation is required to enable one country to respond to another.”
- Transfer (Fig. 7B). After entrainment the modules solve a non-stationary sequence of games that reward a varying middle block of countries for disagreeing: “it has learned how to move in a way that finds solutions”, for “the class of problems with this module structure”.
- By selection (Fig. 8). 7 modules of 7 populations reach the same within-module synchrony by mutation and selection alone, about 10⁶ timesteps against 10⁴ for Eq. 5.
- Hierarchy (Fig. 9). N = 343, 7 × 7 × 7, Eq. 6 couplings 1, 0.01, 0.0001. Modules wake, then meta-modules after about 40,000 steps with ν ten times larger. The transition is not scale-invariant: “higher-level modules do not have the ability to respond to selection as a single unit”; a further transition “would require higher-order traits (e.g. meta-phase differences)” (v2).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Payoff-driven phase adjustment synchronises units within strongly coupled modules, using only local information | strong (in the model) | Fig. 5B, Eq. 5; one parameter setting, one game |
| C2 | Within-module synchrony makes each module switch as one unit and the system reach a global optimum | strong (in the model) | Figs. 5C–D, 6; 50 of 50 runs against 0 of 50 |
| C3 | Global synchrony does not produce this; the synchrony must match the module structure | moderate | Fig. 7A, one condition, caption-level report |
| C4 | The learned organisation transfers to other games over the same modules | moderate | Fig. 7B, one non-stationary sequence |
| C5 | The same result arises by individual-level selection on timing | moderate | Fig. 8, one 7 × 7 run reported |
| C6 | The mechanism iterates to a second level, but not scale-invariantly | weak | Fig. 9, one run |
| C7 | Cell-cycle and reproductive synchrony in development and major transitions may work this way | weak | analogy in Discussion, no model of either |
Method
Simulate N units by Eqs. 3–4 from random states and phases, with or without the phase rule (Eq. 5); sample the learned phases and test them in fresh runs; impose global synchrony as a control; switch the game during a run; replace Eq. 5 by mutation and selection on θ; repeat on a two-level modular game. Energy at the end of each run is the score.
Concepts
- Decision cycle — the periodic alternation of a unit between an undecided, sensitive phase near a saddle point and a decided, acting phase, set by γ_i(t) and the unit’s phase θ_i.
- Entrainment (as used here) — the change of θ_i by Eq. 5 or by selection, toward partners whose states currently agree with the coupling.
- Inner alignment — synchrony of decision cycles inside a group.
- Transition in agency — a group gaining the same minimal competency its members have: choosing between two (collective) states so as to reduce (collective) stress.
- Hard problem of transitions (v2) — that a structure created by individual incentives either aligns with what individuals already do (and is redundant) or opposes it (and would not be created).
Connections
It is the model behind Watson’s agency essay (LIT-990, read in NOTE-747, which cites it as “under submission”), and continues the Watson, Buckley and Mills work on self-modelling Hopfield networks and associative learning in adaptive networks. The units are Waddington-style bistable deciders; the energy function follows Hopfield; the timing rule is offered as an analogue of Kuramoto entrainment (Izhikevich and Kuramoto). It addresses Okasha’s fitness-1/fitness-2 problem in the evolutionary transitions in individuality.
Bearing on the record
- LIT-989’s seed claim is half right. “Aligning the timing of components makes a higher-level unit that acts as one” is what the paper shows. “A case for the mutual-entrainment test of CLAIM-172” needs two qualifications. The paper’s entrainment is a payoff-gradient rule on phase, not timing pulled by the partner’s output as in HKB (THEORY-069). And the result turns on which units are synchronised: synchrony imposed on all units makes no unit (Fig. 7A).
- CLAIM-172 (mutual entrainment as a sufficient mark). Support, with a
sharpening. Within-module entrainment is mutual and does make a unit
relative to the target of resolving between-module constraints. But
shared timing alone is not what made it: the control with global
synchrony has common timing and no unit. Relative to this target, the
mark that works is synchrony matched to the strong couplings. Whether
that control counts as mutual entrainment is doubtful, since it is
imposed, so it is not a counterexample to CLAIM-172 as worded. CLAIM-172’s
defeated_if(a model of each side taking the other as input predicts as well as the pair model) is met trivially here: the micro equations predict everything, and the authors say “the higher-level agent is not a variable in any of our equations of motion”. Read literally, the defeat condition would be met by any fully specified simulation. It needs to be read as descriptive economy, or reworded. - CLAIM-157 (the block criterion). The game’s interaction matrix is nearly decomposable throughout: within 1, across 0.01. The block is there before entrainment and after it. What changes at the transition is not the coupling but the timing, and with it whether the block acts as one relative to the other modules. So a block that would clear CLAIM-157’s threshold need not behave as a unit. That is consistent with CLAIM-157, which disclaims agency, and it shows the two come apart in a concrete case: one system by coupling strength, and a responding unit only after synchrony.
- QUESTION-035. Here the boundary is set by the modules, and the timing process finds it: “the structure of this synchronisation (i.e. who is synchronised with whom) implicitly defines the membership of the collective”. The boundary is given in advance by c_ij (footnote 2), so the paper does not show a boundary drawn where none was built in. On duration, the learned phases persist once formed and carry over to new games; dissolution is not studied.
- THEORY-232. It names this paper as “Tissot et al., under submission, not held”. It is now held as LIT-989. THEORY-232 says the learned relationships “pick out groups that ‘work together’”. The learned relationship is phase, not coupling weight, and the groups match modules the couplings already define. THEORY-232’s comparative measure is computed here in effect: total energy at equilibrium with the old timing against the new (Fig. 6).
- New THEORY. THEORY-265 holds the model result.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- One game with clean, non-overlapping, consistent modules. The authors say natural interactions will be “of various different strengths, exhibiting complex (sometimes unresolvable) frustrations”.
- The phase rule is postulated (or derived from an approximation that holds sign constant over a cycle). Appendix A’s final equation writes ν Σ_j c_ij dU_ij/d(θ_i − θ_j), and dU_ij already carries c_ij, so as written it gives c_ij² where Eq. 5 has c_ij. The hierarchical section calls the phase rate σ where elsewhere it is ν. Both versions carry these slips.
- Single runs or 50-run batches at one parameter setting; no sweep over the coupling ratio, ν, ω or noise, and no statistics beyond counts.
- The new level has “no internal computation or information integration”, no boundary in space and no metabolism; the authors say so.
- The developmental and evolutionary interpretations are analogy.
Open questions
- Does payoff-driven timing find groups when modules overlap or constraints are frustrated?
- Does entrainment through the units’ outputs (Kuramoto or HKB coupling), rather than a payoff gradient on phase, give the same result?
- What dissolves a woken module: a change in the game, noise in θ, or drift? The paper shows formation and transfer, not ending.