From Cycle to One-Way Road When Structures No Longer Depend on Their Participants: Why the Two Channels of Error Correction Share One Root and Fail Together

Abstract Most historical collapses are overshoots: structure is dismantled when it can no longer be paid for, participants return, and order is reorganised. The collapse paper shows that, in a minimal fast–slow model, this cycle depends on a fact seldom written down: structure cannot do without its participants. This paper writes that fact as a matrix element and discusses how AGI might drive it to zero. In the model, the effect of participation on structural growth, A_UΦ, has two terms: a material term, structural income growing with the participation base; and an information term, structure contracting actively when readings show participation falling. The paper defines a "one-way road" as A_UΦ → 0 while the pressure of structure on participants remains: people depend on the structure, but the structure no longer depends on people. The material term carries a factor (1 − θ), where θ is the share of structural income that does not depend on participants. If the strength of the information term also comes from the same dependence, so that the owners of the reading have reason to read only because they depend on the participants, it too carries the factor (1 − θ). Under this condition the two channels of error correction share one root and fail together: the boundaries of the adjustment and lock-in regions become k_c(θ)/(1 − θ) and k_L(θ)/(1 − θ), both diverging as θ → 1; for any finite nominal reading gain k there is θ̄(k) < 1 beyond which the system falls into hollow lock-in: the structure keeps growing and the participants do not return. Of 20 numerical integrations, 19 fall in the regions the closed forms predict; the remaining one lies right at the lock-in boundary, with a trough of about 6200 time units, exactly the divergence the precursor describes. Several pressures in the AGI case, such as the disappearance of the outside and the withdrawal of the carriers of correction, are then not independent coincidences but symptoms of one cause. This cause has three precursors observable in advance: troughs after collapse lengthen as 1/(θ̄ − θ) (with zero reading gain, θ̄ is θ_c); the structural stock rises while participation collapses; and when cross-layer alternation vanishes, the slow-mode recovery rate tends over time to the smaller of the two layer rates (decoupling), instead of staying above it (weakened feedback) or falling below it (sign flip). The structure of prescription follows from the same matrix element. The reading channel must be decoupled from θ, that is, written into places that do not depend on incentives; once it is hard-wired, even at θ = 1 the adjustment region requires only a finite threshold k_c(1). Readings must be faster than structural accumulation: at fixed gain, if accumulation speeds up m times, the reading lag must shrink about m times. What is coordinated is the clock, not the reading. In the ordering of catch-up, lengthening the interval between revolutions and shortening identification delay have the same effect per unit, but identification delay has a lower bound and coordinated deceleration takes time to take effect once launched, so coordinated deceleration must come first and be launched early; it cannot be held in reserve until the unilateral lever is exhausted. Two common statements are corrected along the way: in the catch-up formula, "deceleration acts at exponential order and everything else at linear order" does not hold; and the collective tail is set by the agent closest to the ruin point rather than by the most aggressive, the direction of "restraint" on the position axis must be stated, and unilateral restraint by agents that do not set the tail lowers exceedance probabilities at finite thresholds provided the vacated share is not filled. Cold zones need not wait for full graduation to disappear, and the paper gives a conjectural estimate. The number of executions on external data is zero.