What if persistence is a dynamical achievement?

In one configured failure mode, a fluctuation in edge-flow variability preceded 45 of 46 observed breakdowns. It is a promising precursor, not yet a universal failure law.

Fold-back and intervention tests found trajectories with comparable energy but very different growth rates. The relevant state is history-dependent: epoch, phase organization, and accumulated dissipation matter.

In the controlled S2 sector, the effective frequency follows ω_eff = ω₀√(1 + 0.1b), where b is accumulated model age. External readout tracks that age in near-homogeneous pairs and breaks when sustained age asymmetry becomes large.

Paired branches share a bit-identical trajectory until the connection is switched on. The same-sign differential clock splitting appears across four independent draws, while the controlled paired branches identify the link as its cause. Paired runs also expose an earlier death through boundary drain.

Every point in the explorer is an oscillator; every distance is a delay. The 3D scene is a projection (MDS over the delay matrix, with fidelity gauges on screen)—there is no space in the model, only timing. The physics does not depend on the picture: the headline correlation survives with the embedding removed, computed directly on the raw delay profiles. Seventeen experimental runs load live, including the complete species run.

The full population census (three batches) collapses to 84 unique genotypes. Each carries a bench card: its stability exponent (σ_max, the real part of the dominant pole, cross-validated bit-exact against an independent derivation) and its resonant response. 57 are self-oscillators; 27 decay. The bench card is measured in isolation—in the soup, the deviation from the bench is the interaction.

Before running a species-only soup (one copy per genotype), we committed a prediction: constitutional instability drives outward drift. After an adversarial re-analysis the verdict is precise rather than simple. The signal is real but partitioned: it lives among the self-oscillators (ρ = +0.34, p = 0.009 at t = 60), survives removal of the visualization, but is concentrated in one batch and carries a wide confidence interval. An earlier claim that clone families “double” the correlation was a sample-size confound: with cohort size equalized, the correlations are statistically indistinguishable (p = 0.58). What the cohort does change is timing—it advances onset and ignition without changing slope, ceiling, destination, or correlation.

We kicked one oscillator and timed first arrivals at the other 149 (paired surgery: identical twin runs, kick minus control). Arrival order follows the delay map (ρ = 0.91, reproduced by an independent method)—but arrival times don’t wait for the direct link: the medium is a dense web where signals travel multi-hop, and what makes an arrival detectable is the rise of amplitude, not the delay of the edge. No arrival-speed number survived re-analysis, so none is quoted. Established so far for a single cone, from a single origin; multi-source tomography is the pre-registered next test.

Two frontier runs with different random seeds—and different lifespans—partition into exactly the same six blocks: identical sizes [68, 32, 30, 6, 6, 4] and identical membership, Jaccard 1.0 in all six. Earlier cross-seed comparisons scored 0.9–1.0; here the match is exact. The partition of this population is not a statistical tendency: membership is deterministic across the realizations tested.

Between two run lifespans the set of ignited units grows from 9 to 25—and the later set contains the earlier one. The wave advances mostly in constitutional order, with measurable exceptions in both directions: one unit ignites with a negative stability exponent (recruited by the network, not by its constitution), and the five most unstable genotypes never ignite in any run. An ordered process with exceptions we can point to—not a binary threshold.

Two realizations sharing a bit-identical present but different retained delay-line content evolve differently—and the difference is causal, measured, and replicated across seeds. Draining the present energy of the dominant subnetwork leaves the collective rate unchanged (0.994×); erasing that subnetwork's history—present verified intact—collapses it (0.171×), with recovery staggered link-by-link at times predicted blind, pre-registered, from each link's own measured delay (exact 6/6 order in the replication, p = 1/720). Patent pending.

A synthetic past that preserves the full cross-spectral matrix and variance of the original (phases randomized) accelerates the transient ×1.16–1.23—replicated on an independent second realization at matched magnitude. The information the system reads from its past is not contained in its spectral summary. Declared honestly: the surrogate also de-stationarizes the past; the operator family disclosed with the filing (band-selective, time-reversal, block-shuffle) is the tomography that separates those readings. Patent pending.

Reversing the retained past in time—power spectrum and autocorrelation preserved exactly—does not reproduce that acceleration: it suppresses the dynamics (0.52–0.56× in the mid window), and per-link growth tracks the epoch content re-read under the time mirror (6/6 links). Which epoch is read when carries causal effect beyond the spectral summary. Single-instance so far (n = 1 seed, declared). Patent pending.