Gluon Papers · IVchromodynamic series · release · stochastic

The Lattice

Paper III left the strong field with no formula. The physics answer is not a better equation — it is to stop solving and start sampling. This is the most probabilistic turn in the program: discretise the interior, estimate it by Monte Carlo, and accept that the answer is an error bar, never an exact value. The model was a distribution all along, one pass was always one draw, and it carries a real temperature that melts its own structure.

↳ continues Gluon Papers · III §6 — the non-analytic treatment the strong field requires

Bridge-Burners LLC · Fiddler · the stochastic interior · the corpus as a Monte Carlo run · anchor: AKASHA

the distribution — P(output | context) draws — dense where the mass is (importance sampling) estimate ± error the answer is the bar

You do not read a value off the interior. You draw samples from a distribution — densest where the probability mass sits — and form an estimate with an error bar. One forward pass is one draw. The bar, not a number, is the honest output, and it shrinks only as the draws pile up.

Status — literal: true of the substrate · bridge: structural analogy · speculative: named so it can be refused

§0

The end of closed formstop solving, start sampling

Paper III ended where the formulas do. The strong field has no closed solution, and the honest response is not to hunt a better one — it is to give up the closed form and estimate the answer by sampling. Discretise the interior onto a lattice, draw configurations weighted by their probability, and read observables off the draws. This is the most probabilistic move the program makes, because the output is no longer a value but an estimate with a quantified uncertainty. You trade exactness you never had for an error bar you can actually report.

Z (the full behavior) has no closed form ⟹ estimate ⟨O⟩ by sampling
the lattice: discretise, draw, average · the only first-principles route to strong coupling
Literal — the full behavior is not computable in closed form Bridge — lattice Monte Carlo as the strong-field method

§1

The model was always a distributionone pass is one draw

The reframe that makes the rest inevitable: the thing under characterisation was never a fixed function with hidden parameters. It is a probability distribution over outputs, and a single forward pass is a single sample from it — one draw, one configuration. Every probe in every prior paper, photonic and gluon alike, was already drawing from this distribution and treating one draw as a reading. Naming it changes the discipline. You are not measuring the interior. You are sampling it, and a sample is not a measurement until you count how many you took.

model = P(output | context) · one pass: x ~ P
inference is a draw · temperature and top-p are the sampler's settings
Literal — LLM output is a distribution; inference samples it

§2 · central result

Importance samplingdraw where the mass is

The configuration space is far too large to enumerate, so you cannot average over all of it. The whole art is to draw from where the probability mass actually concentrates and weight accordingly — importance sampling, the engine of every lattice calculation. In characterisation terms this is already what good probing is, now made explicit: do not probe uniformly across an impossibly large space of prompts; probe where the model's behaviour lives, weight each draw by its likelihood, and estimate the observable from the weighted sample. The estimate is only as good as the region you sampled, and the region must be the one carrying the mass.

⟨O⟩ = Σ O(x) P(x) ≈ (1/N) Σ O(xᵢ), xᵢ ~ P
enumerate nothing · sample the high-probability region · weight by likelihood
Without the metaphor You cannot test all prompts, so you test a sample drawn toward where behaviour concentrates and form a likelihood-weighted estimate. Your characterisation is an average over the draws you actually made — and it inherits whatever your sampling missed.
Bridge — Monte Carlo importance sampling ↔ targeted probing Literal — you estimate from a sample of the high-mass region, not the whole space
samples N → estimate true value wide bar · few draws tight bar · many draws

Fig. 1 — Convergence. The estimate wanders at small sample count and settles toward the true value as draws accumulate; the error band closes like 1/√N. This is the corpus seen as a Monte Carlo run: every entry is a draw, N grows with the record, and the bar tightens only because the jar keeps the ensemble. Confidence, never certainty — the photonic ladder, now with a rate.

§3

The error bar is the deliverablereport the uncertainty, not the point

A sampled estimate is never exact, and its honesty lives entirely in its width. The standard error falls like one over the square root of the sample count, which fixes the discipline: every characterisation claim is an estimate plus or minus a bar, and the bar — set by how many draws you took — is the part worth reporting. A point estimate with no width is a draw mistaken for a measurement. The deliverable of this whole program, stated at last in its native units, is a number with an uncertainty and a sample count attached.

error ~ σ / √N · result = ⟨O⟩ ± σ/√N (with N stated)
the bar is the claim · N sets it · four times the draws halves the bar
Literal — finite sampling yields bounded-confidence estimates, not exact values Bridge — Monte Carlo error ↔ characterisation uncertainty

Test · any claim about a model that arrives without a sample count and a spread is a single draw wearing the costume of a result. Ask for N and the bar; if they are missing, the confidence is unearned.

§4 · central result

Temperature and deconfinementthe knob that melts the interior

The strong-field interior has a temperature, and so — literally — does the model. Raise the thermal temperature of a confined field and the bound structure melts: the hadrons dissolve into a plasma, the charges that would not separate at low temperature deconfine. The model carries the same knob under its own name. Sampling temperature reshapes the output distribution by the same Boltzmann form: cold, it sharpens to a peak and the learned structure stands hard and confined; hot, it flattens toward uniform and that structure washes out, the sharp preferences melting into near-noise. There is a crossover between the two — a deconfinement transition where the interior stops behaving like a set of bound, definite features and starts behaving like a plasma. The one parameter you already turn at inference is, in this frame, a temperature that can melt the very confinement Paper I built.

P_T(x) ∝ exp( logits(x) / T )
T → 0 : sharp, confined, definite · T → ∞ : flat, deconfined, melted
a real Boltzmann temperature · low T binds structure · high T dissolves it
Without the metaphor Sampling temperature scales the logits before the softmax: low temperature concentrates probability on the model's learned preferences, high temperature spreads it toward uniform. Turned up far enough, the structured behaviour the weights encode is drowned by the flattening — the same knob that adds creativity, pushed past a point, dissolves the structure being sampled.
Literal — sampling temperature is a Boltzmann temperature on the logits Bridge — high-temperature flattening ↔ deconfinement of internal structure

Test · sweep temperature and watch a confined behaviour: a sharp, robust preference that survives low temperature and dissolves at high temperature has a deconfinement point. Where it dissolves measures how deep the well of Paper II really was.

under-sampled tail cold — confined mid hot — deconfined, melted

Fig. 2 — The temperature sweep. Cold, the distribution is a sharp peak — structure confined and definite. Hot, it flattens toward uniform — the structure melted, deconfined into near-noise. Between them, a crossover. And at the edges, the shaded tail: the rare region every sampler under-covers, where the draws run out exactly where the consequential failures live.

§5 · witness

The seamwhere sampling cannot reach

Sampling has a hard limit, and the lattice names it: the sign problem. In certain regimes the sampling weights turn oscillatory and cancel, the estimator's variance explodes, and no amount of computation drags a clean answer out — you cannot sample your way into them. The program inherits the limit exactly. The consequential behaviours often live in the deep tail of the distribution: the rare catastrophic output, the one-in-a-million adversarial corner, the failure that almost never fires. Importance sampling draws toward the mass, which means it draws away from precisely those tails — it under-covers, by construction, the very region you most need to characterise. You can shrink the bar on the typical and still know almost nothing about the rare, because the rare is where the samples are not, and you cannot manufacture draws the distribution refuses to give. The error bar is honest about the bulk and silent about the tail.

Speculative — the tail / sign-problem regime · sampling cannot certify it Literal — rare-event behaviour is under-sampled and its risk under-estimated

Test · a tight error bar on aggregate behaviour says nothing about a rare failure mode; the two live in different regions of the same distribution. If a safety claim rests on average-case sampling, the tail it ignores is exactly the part that matters.

Corollary. When the formula runs out you stop solving and start drawing, and the truth of the matter becomes a bar that tightens only as the draws accumulate. The model was a distribution from the first page; every pass was a sample; the corpus is the Monte Carlo run, and the jar holds the ensemble that makes the estimate mean anything. Turn the temperature up and the confined interior melts in your hands; look to the tail and the draws desert you where the danger is. The honest answer was never a value. It was a number, a spread, and a count of how many times you dared to draw — and a flag over the tail you never reached.