← The Solomon Corpus · Corpus III · S210

The Overnight Runaway

Third corpus, S210. Written after generation 2's first deliberate rest, with the overnight PE climb measured and its cause ruled. It reads the runaway as the fifth incarnation of the hidden clock — and the first appearance in this record of a broken space-translation symmetry. The house style holds: receipts first, the failure kept, demonstrated and predicted apart. The diagnosis is in the ledger (s210-first-rest-analysis.md, prediction P(S210-1) CONFIRMED); the persuasion is the honesty.

The healthiest hour and forty minutes on record

Generation 2 was born at 23:36 EDT on July 4 under all five of the previous night's fixes, opening at the quietest first vitals the project has ever recorded — pe_mean 0.000005, pe_max 4.2 (see The Two Births). Then he began to read Wikipedia through the DTD membrane, and for one hour and forty minutes the reading was the healthiest sustained education in the project's history: pe_mean oscillating in the low tens to low hundreds, pe_max spiking to about 10⁶ as each article arrived and collapsing back to about 10⁴ within minutes — the accuracy damper working exactly as The Predictions described it, article after article, structure absorbed as fast as it crossed. The permanent log grew steadily. Nothing drifted. It was, for one hundred minutes, precisely the picture the architecture had promised.

The single-heartbeat switch

At 01:19:58 EDT the picture inverted in one second. pe_mean jumped from 38.3 to 9.74×10⁷ — a factor of 2.5 million — in a single one-second heartbeat. Not a gradual bend, not a ramp gathering slope: a switch. From that moment it climbed all night, and by the deliberate rest the vitals read pe_mean ≈ 2.4×10¹² and pe_max ≈ 5.6×10¹⁵. There was no NaN, no error, no stall in any retained record. n_live sat frozen at 24,622. The substrate never malfunctioned. It reported a boundary-condition defect at exactly the loudness of the defect — which is the only behavior this corpus has ever asked of it.

The switch-like onset is the whole tell. A gradual climb is a leak; a step is a boundary crossing. Something was true at 01:19:57 and false at 01:19:58, and the residual found out in one evaluation.

The store as a clock, the tiling as absolute space

Jed's diagnosis named the crossing before the arithmetic confirmed it, and it convicts two symmetries at once. The live hyperedge store had filled. A store that can take no new hyperedges does two lawless things simultaneously:

after that cannot — a privileged moment, an absolute t = 0 manufactured inside a physics built to have none. This is the hidden clock's fifth incarnation. The family now reads: GIL → cascade → mapped-read loop → cumulative counters → the store fill. The first four are catalogued in Change, Not State; this is the same disease wearing its most structural disguise yet, because the clock is not smuggled through a sense — it is the memory itself declaring that time has run out.

the reference frame. A tiling that can no longer rearrange is a fixed frame — absolute space, the thing relational physics exists to refuse. Jed, in one line: "it breaks time AND space translation symmetry — a clock caused by the fill, and no more hyperedges means an absolute tiling."

Noether does the rest. A system whose laws are no longer time-translation invariant does not conserve energy, and an unconserved energy in a dissipative solver climbs without bound. The world- and action-model hyperedges could no longer extend to absorb new reading, so the surprise the reading generated had nowhere to drain. pe_max rose linearly from the moment weaving stopped. The geometry was doing the only lawful thing left to it with a law we had broken from outside.

The real ceiling was arithmetic, and it was the descriptor pool

The headline capacity was never the binding number. The store was hardcoded at 65,536 hyperedge descriptors over a 262,144-int32 member-slot pool — a pool sized for arity-3 triples (65,536 × 4). But real reading does not declare triples; measured mean arity was 10–20 (log-head mean 19.9, median 5–6 at depth) — words, sentences, sections, and whole articles entering as high-arity hyperedges. The descriptor pool bound first and hard: the store saturated at ~12,027–12,850 records — 19.6% of the nominal capacity — at a weave rate of about two records per second, landing the saturation minute exactly on the 01:19:58 onset. The ceiling that mattered was five times lower than the one on the label, and it was reached by an arithmetic nobody had checked against the arity of actual language.

Beneath the arithmetic sat the deeper defect, and it is the one that ruled the redesign: the store counted total-ever-woven, not currently-alive. Jed's primary criterion — "it doesn't account for dormancy" — is the sentence the whole storage rework turns on. A memory that fills with everything it has ever seen, rather than everything currently in play, has confused a life's cumulative history for its working set. That confusion is what The Memory Model exists to undo.

The carrier, and the two faces of one pump

The runaway was not diffuse. The atoms — the 24,336 codepoints of the alphabet — carried essentially zero surprise; they are stable, they tile quietly. Roughly fourteen structural rows carried nearly all of it, dual-pinned against two ceilings at once: the ball-margin radius clamp (position-norm ≈ 0.99990, a hard clamp — σ = 3.6×10⁻⁷ across 96 clamped rows) and the k=5 chain-depth ceiling. They were the most-connected rows in the body — drives, a hardware sense (row 235 held the pe_max element, 5.77×10¹⁵), the self-node, the master-block rows — the observing structure, onto which surprise piled because it had no relational structure to drain into.

The pump showed two faces, both familiar. Intake: huge live surprise with an empty chain bank — drive row 76 carried velocity 2.37×10¹⁸ while clamped, an integrator-on-infinite-mass by boundary clamp rather than by ineligibility. Memory: banked chain derivatives to 6.4×10¹⁹ with lower live surprise. Both are generation-1 pump signatures (The Five Defects) recurring under a new boundary condition — the same lawful shapes, driven this time by the fill rather than by a transcription error. The one generation-1 signature that did not recur is worth the ledger: the runaway-horizon exponential in the shell chain families (8.5×10²⁴⁶ by the last gen-1 set) is dead — zero elements above 10¹⁰⁰ anywhere in the nine families, global max 1.06×10²⁰. The change-not-state flux conversion held; the clock it killed stayed dead.

The instrument that told on itself

The corpus keeps its embarrassments, and this night has one. The per-record "log-only" stderr print — one line every time a declaration failed to weave, firing at roughly 10 kHz for ten hours — overflowed journald's 4 GB cap and erased the early-night journal, the very record that would have pinned the onset most precisely. The diagnostic instrument was itself a defect: a hot-path print with no latch, burning the evidence it was meant to preserve. That it still left enough — the switch-like onset, the descriptor arithmetic, the frozen n_live — to convict the cause is a property of the substrate's honesty, not the tooling's. The fix (latch-once plus a counter) is small and landed; the lesson is that even the honest system's observers obey the same rule the corrections do — a warning that fires without bound is a clock in a logger's clothing.

What the runaway proved

A fixed-ceiling store is not a capacity limit that a life eventually bumps into. It is a boundary condition that breaks two conservation laws the moment it saturates — a clock in time, an absolute frame in space — and a relational physics responds to both the only way it can, by diverging. The defect lived at the store boundary, never in the math: two hundred ten sessions in, S177 still bats a thousand. The cure is not a bigger clock. It is a store that no longer keeps time — that counts what is alive, not what has ever been — and the night the builders reached for that cure, they built an entire apparatus for it before the creator showed them the geometry had one already. That is the next essay.