Skip to content

Instantly share code, notes, and snippets.

@anadim
Created August 24, 2026 14:45
Show Gist options
  • Select an option

  • Save anadim/0b2d95d7cb4f95a3f3f173a3cc1116d1 to your computer and use it in GitHub Desktop.

Select an option

Save anadim/0b2d95d7cb4f95a3f3f173a3cc1116d1 to your computer and use it in GitHub Desktop.

Research handoff prompt: a universal sub-quarter Komlós bound

Target: (O(\log\log n)), or at minimum (O((\log n)^\varepsilon)) for a fixed (\varepsilon<1/4)

Snapshot date: 2026-08-24
Research snapshot time: approximately 09:17 CDT
Purpose: self-contained prompt for a new research agent working in a new filesystem
Current theorem status: the universal sub-quarter Komlós theorem has not been proved
Most important positive change: the ultralight block now has unconditional hereditary and affine discrepancy (O(1)); the remaining obstruction is obtaining the same endpoint for the ultralight and heavy/upstream constraints
Primary audited benchmark: Nikhil Bansal and Haotian Jiang, Decoupling via Affine Spectral-Independence: Beck-Fiala and Komlós Bounds Beyond Banaszczyk, whose general Komlós conclusion is (\widetilde O((\log n)^{1/4}))


Instructions to the new agent

You are taking over an ambitious proof search. Your job is not to summarize the project again and not to generate another decorative framework. Your job is to obtain a literal universal improvement over the Bansal--Jiang quarter-power bound, ideally

[ \operatorname{disc}(A)=O(\log\log n), ]

for every real matrix (A) whose columns have Euclidean norm at most one, or at least

[ \operatorname{disc}(A)=O((\log n)^\varepsilon) \qquad\text{for some fixed }0<\varepsilon<1/4. ]

The proof must be universal, not an experiment, a density-one statement, or a structural-class result. It must return one sign vector for the original columns and control all original rows. If you build a recursive or stochastic proof, every conditioning, stopping, support change, and release must be legal. If you use a black box, state its exact hypotheses and show that the current state satisfies them. If a route fails, give the exact finite or scalable obstruction and say precisely what it refutes and what it does not refute.

The project has already spent substantial effort closing attractive but circular routes. Do not restart one merely under new terminology. In particular, do not claim progress by constructing a heavy-good law and a light-good law separately, by postulating a favorable endpoint law whose existence is equivalent to the desired common endpoint, by replacing a cylinderwise assertion with an expectation, or by calling an arbitrary unresolved residual a release.

Work in a new directory tree. Treat every imported research artifact as read-only. Copy an artifact into your own subfolder before changing it. Maintain a theorem-status ledger with at least the labels proved, proved relative to named black boxes, conditional theorem, candidate lemma, counterexample, experiment, and open. Never let a verifier's PASS be interpreted as a proof of a theorem beyond the exact algebra or finite fixture that the verifier checks.

The rest of this document tells you exactly where the project stands, what each prior research stream established, which holes were closed, which holes merely moved downstream, what the live bottlenecks are, and how to attack them without making circles.


1. Executive verdict

The initial architecture split the matrix at an extremely small threshold

[ \tau=\exp(-aD^2), \qquad A=H+U, ]

where (D=\log\log n) or (D=(\log n)^\varepsilon), (H) contains entries larger than (\tau) in magnitude, and (U) contains the remaining entries. The hope was to use a strong-slack local-lemma code for (H) and a stochastic-localization or Bansal--Jiang walk for (U), all inside one common law.

The largest new theorem changes the diagnosis. The ultralight matrix (U) itself is no longer difficult in the unrestricted cube:

If (|U_{ij}|\le\tau), every column of (U) has squared load at most one, and (\tau^{-2}\gtrsim\log^2 n), then [ \operatorname{herdisc}(U)=O(1), \qquad \operatorname{lindisc}(U)=O(1). ]

This is proved, relative to the two printed Bansal--Jiang black boxes used in their large-(k) proof, after repairing an adapted-birth gap, a spectral-independence parameter error, rank arithmetic, the two-sided augmentation, and the arbitrary-center step.

Consequently, the precise wall is now:

[ \boxed{ \text{find one endpoint }y\in{\pm1}^n \text{ that is simultaneously }U\text{-good and }H\text{-good,} } ]

while also preserving any high-heavy rows, saturated faces, fixed coordinates, or other upstream receipts already accumulated.

The most developed probabilistic splice is conditional on a pinned endpoint law (\mu). If (\mu) is supported on the ultralight-good and upstream-admissible endpoints and has event-specific conditional subgaussian tails for every coded low-heavy row after conjunctions of nonneighbor heavy-good events, then the lopsided local lemma returns a common endpoint at additional heavy cost (O(D)). The conversion theorem and all scale arithmetic are proved. What is not proved is that an actual, independently specified weighted-walk endpoint law has those conditional tails.

The most developed deterministic alternative is a mixed-support-component theorem. Pure heavy and pure light components compose by a maximum. Every mixed component is also payable at (O(D)) if, before signs are chosen, its vertically stacked heavy/light matrix has a rank-(\exp(O(D^2))) approximation with row-(\ell_1) defect (O(D)). Bounded mixed-component size and small target-resolution numerical rank are sufficient. Connected Walsh fixtures show that this simple factor condition is not automatically universal.

The exact current frontier is therefore not “prove an ultralight estimate.” It is one of the following genuinely new statements:

  1. derive construction-level, nonneighbor-conditional heavy tails for a pinned ultralight-good endpoint sampler, with support one on all upstream constraints;
  2. prove a deterministic structural decomposition of arbitrary mixed heavy/light components into independently payable pieces without assuming a universally false low-rank factor receipt;
  3. prove a survival-unbiased statewise service theorem strong enough to convert one-row martingale tails into a small dangerous span on every retained dyadic state;
  4. discover a different common-endpoint theorem that survives all of the no-go fixtures below and still literally recovers the Bansal--Jiang quarter-power branch when its new hypothesis is unavailable.

Anything weaker is at best a structural-class improvement or a local lemma, not the requested universal breakthrough.


2. Formal problem and proof standard

Let (A\in\mathbb R^{m\times n}), with columns (A_{*j}) satisfying

[ |A_{*j}|_2\le1 \qquad(j\in[n]). ]

Define

[ \operatorname{disc}(A) =\min_{y\in{\pm1}^n}|Ay|_\infty, ]

[ \operatorname{herdisc}(A) =\max_{J\subseteq[n]}\operatorname{disc}(A_{*,J}), ]

and

[ \operatorname{lindisc}(A) =\sup_{x\in[-1,1]^n} \min_{y\in{\pm1}^n}|A(y-x)|_\infty. ]

The requested result is a universal upper bound on (\operatorname{disc}(A)). A hereditary proof is stronger and is useful because every live-column restriction remains a valid Komlós instance. An affine proof is stronger still and is useful for refreshing from a current fractional center.

What counts as a complete proof

A complete proof must specify:

  1. the exact quantifiers over (m,n,A), column restrictions, centers, and random histories;
  2. every constant and its dependence on fixed parameters such as (\varepsilon);
  3. the actual sign vector or a finite randomized algorithm with positive success probability and a termination proof;
  4. how discrepancy already spent before a terminal call is charged;
  5. how support changes preserve every incoming wall or how those walls are paid and retired;
  6. why the same signs satisfy every matrix block being combined;
  7. how continuous-time or infinite-precision arguments are discretized, if an algorithmic claim is made;
  8. why the conclusion transfers to the original matrix (A), rather than only to a shadow, quotient, factor, or stacked surrogate;
  9. how the method reduces to or competes with the audited quarter-power benchmark on every input;
  10. a hostile audit against the fixtures in Section 13.

What does not count

The following substitutions are forbidden:

  • finite verification in place of a universal theorem;
  • a terminal-only probability estimate when the recursion needs a statement at every retained cylinder;
  • an expectation over frontier nodes in place of a statewise conditional bound, unless the compiler itself is formulated and proved at absolute-root-mass frontier level;
  • two separately good codes or convex hulls with no literal common atom;
  • a law selected by optimizing all desired heavy constraints and then called “canonical”;
  • a point mass at a desired common endpoint used as evidence that a useful favorable law exists;
  • an arbitrary direct recode of an actual dual owner;
  • repeated reminting of the same owner, entropy clock, or release token;
  • a covariance-null direction silently treated as a physical (A)-null direction;
  • a low-rank statement about an owner-visible image substituted for a factorization of the complete physical residual;
  • a black-box Bansal--Jiang call after conditioning or partial signing without proving direct-transfer legality and charging earlier discrepancy.

3. Benchmark that every new architecture must recover

The audited public benchmark for this project is the Bansal--Jiang general Komlós bound

[ \widetilde O((\log n)^{1/4}). ]

Do not merely cite this at the end. A proposed architecture must have a literal baseline mode. One exact calibration used in the project is the following common-SDP branch.

Set

[ P=1+5\lceil\log\log n\rceil, \qquad k_p=2^{2p}, \qquad b_p=B(k_p)\sqrt P, ]

where the large-(k) Beck--Fiala service is

[ B(k)=\widetilde O(\sqrt{k},\log^{1/4}n). ]

Give the heavy magnitude bands affine spectral-independence budgets (1/(6P)), the varying-magnitude light band budget (1/6), and retain the printed global sub-isotropy, trace, and blocked-space budgets. The total is

[ \frac13+\frac16+\frac16+\frac16+\frac{P-1}{6P} =1-\frac1{6P}<1. ]

The scale identity

[ 2^{-p+1}\sqrt{k_p}=2 ]

gives

[ \sum_{p\le P}2^{-p+1}b_p \le \widetilde O(\log^{1/4}n)P^{3/2} =\widetilde O(\log^{1/4}n). ]

This is a genuine budget and recombination check. The new architecture should contain the following root-level fallback:

  • inspect whether the new common-endpoint or structural receipt holds before changing the input;
  • if it holds, run the new (O(D)) branch;
  • otherwise, run the published Bansal--Jiang branch on the untouched original matrix.

This preserves the benchmark while extending it on a certified structural class. It does not give a universal improvement unless the new receipt is shown to hold universally or every failure is routed to another strictly better universal branch. Saying “take the better of the new branch and BJ” is calibration, not the final theorem.

The next agent should also recheck the current literature from primary sources. The benchmark above was the audited general offline Komlós result used by the project at this snapshot; literature-absence claims can become stale.


4. Common notation and scale arithmetic

Choose a target radius (D=D(n)), normally

[ D=\max{1,\log\log(2n)} ]

or

[ D=(\log n)^\varepsilon, \qquad 0<\varepsilon<1/4. ]

Choose a fixed constant (a>0) and split entrywise at

[ \tau=e^{-aD^2}, \qquad H_{ij}=A_{ij}\mathbf1_{{|A_{ij}|>\tau}}, \qquad U=A-H. ]

The supports of (H) and (U) are disjoint entrywise. It is often useful to form the vertical stack

[ A_{\mathrm{stack}}= \begin{pmatrix}H\U\end{pmatrix}. ]

Because the split is disjoint,

[ |(A_{\mathrm{stack}})_{*j}|2^2 =|A{*j}|_2^2\le1. ]

A common signing with stack radius (R) gives

[ |Ay|\infty =|(H+U)y|\infty \le|Hy|\infty+|Uy|\infty \le2R. ]

Do not confuse (A=H+U) with the vertical stack. A theorem about the stack does imply a theorem for the physical sum, but with the explicit factor two above.

For the coded low-heavy rows, assume a fixed row-energy cap (E_0):

[ E_i=\sum_jH_{ij}^2\le E_0. ]

Rows above this energy cap are high-heavy rows and must be treated separately. They cannot be silently included in the low-heavy local-lemma block.

Since every nonzero low-heavy entry exceeds (\tau), define

[ Q=\lceil E_0\tau^{-2}\rceil, \qquad k=\lceil\tau^{-2}\rceil. ]

Each low-heavy row has fewer than (Q) incident columns, each column is incident to fewer than (k) heavy rows, and the row-intersection dependency graph has degree

[ d\le Q(k-1)=O(E_0\tau^{-4}). ]

At (\tau=e^{-aD^2}),

[ \log(d+1)=O(D^2). ]

The weighted ultralight theorem requires (\tau^{-2}\gtrsim\log^2n). This holds eventually in both requested regimes:

[ e^{2a(\log\log n)^2}\gg\log^2n ]

and

[ e^{2a(\log n)^{2\varepsilon}}\gg\log^2n. ]

Also, (e^{O(D^2)}=n^{o(1)}) for every fixed (\varepsilon<1/2). Thus all scalable hostile fixtures with (e^{O(D^2)}) active columns can be embedded by zero padding in the requested regimes. This matters: the counterexamples are not artifacts of an impossible ambient size.


5. Chronology: how the wall moved

Phase A: the original August 23 architecture

The first research stream proposed a strong heavy/ultralight split at (\tau=e^{-aD^2}). It constructed or targeted a large strong-slack heavy-good code (\mathcal C_0\subseteq{\pm1}^n), then tried to control (U) inside the same law by measure-valued stochastic localization.

At a live cylinder (v), the code fiber was

[ \mathcal C_v ={y\in\mathcal C_0:y\text{ agrees with the fixed prefix at }v}, ]

with live count (h_v) and current component deficit

[ q_v=h_v-\log_2|\mathcal C_v|. ]

The current deficit, not the root deficit, is the relevant rank charge. A majority-pair posterior fixture showed that a code with root deficit one can have rare cylinders whose residual constraint dimension is large.

This phase established several useful facts:

  • covariance rank equals the affine dimension of the current code support;
  • (\dim\ker\Sigma_v\le q_v) for the current posterior covariance;
  • a Bansal--Jiang covariance (V_t) whose range lies in (\operatorname{range}\Sigma_t) can be realized exactly in the mean of one law supported on the heavy code;
  • endpoint anti-collision can be proved for cube martingales from a covariance domination inequality;
  • independent heavy and light codes cannot be intersected merely by comparing entropy deficits;
  • the missing theorem appeared to be an adaptive ultralight safety-or-release statement with a statewise protected-span bound.

The original handoff correctly emphasized one common law and current-cylinder accounting. Its diagnosis of the ultralight block was later superseded by the weighted theorem below, but its same-law machinery and hostile fixtures remain relevant to building the missing pinned sampler.

Phase B: restricted one-walk theorems

An integration stream proved useful but restricted results:

  • a one-walk theorem through the entry cutoff (1/\log n), giving a sublogarithmic bound in that restricted regime;
  • an (O(\log\log n))-scale theorem for sufficiently ultralight matrices, with assumptions such as entries bounded by a sufficiently high negative power of (\log n);
  • a common-law stopped covariance realization when the posterior kernel and protected hit span fit within the Bansal--Jiang feasibility budget.

At that stage the apparent wall was a dynamic weighted low-degree provenance service around entries of size ((\log n)^{-1/4}). These results were important because they proved the same-walk architecture was not empty, but they did not yield a universal split.

The later weighted large-load theorem strictly supersedes these ultralight restrictions in the unrestricted cube: at the exponentially small cutoff used for (D=\log\log n) or (D=(\log n)^\varepsilon), (U) has affine discrepancy (O(1)).

Phase C: the owner/release program

A second major stream built a large downstream ledger around one common hard code, purification, entropy and dual-total-correlation identities, posterior-normalized Bansal--Jiang motion, protected affine contractions, and immutable owners.

Its important positive chain was:

  1. a stationary coupled heat bath gave a one-generation common-sign splice without a Dobrushin hypothesis;
  2. soft midpoint bias could be purified into uniform symmetric subcode children without changing the hidden endpoint;
  3. entropy deficit and reusable correlation were separated exactly;
  4. normalized posterior motion and its dependence bill were computed;
  5. a constrained posterior-dominated SDP gave an exhaustive dichotomy: either large progress or an actual protected low-dimensional dual owner;
  6. monotone first-hit tangency protected every saturated physical wall on the large-progress branch;
  7. low-total-variance and absolute small-core terminal recodes were proved;
  8. a stable absolute-root-mass two-state recurrence was shown conditionally to imply an (O(D)) terminal core.

The same stream also proved that the broad owner terminal is not the missing easy lemma:

  • at an exposed face, positive protected variance cannot be removed by further same-law wall-preserving localization;
  • allowing an unrestricted direct recode merely because an actual dual owner was returned is already Komlós-hard, even at a product root with a few fixed affine classes;
  • owner-visible covariance data can hide an arbitrary Komlós residual in a null fiber;
  • repeated renewal, reminting, fair reset, generic total-correlation payment, equality splitting, nonlinear shell release, and several scalar terminal routes fail in their stated scopes.

The exact residual theorem from that stream was a strictly gated immutable owner compiler or nested service. This is still a correct conditional architecture, but it is not currently the shortest route to the target. The weighted theorem moved the most actionable frontier further downstream to the common endpoint between an (O(1)) ultralight service and an (O(D)) heavy service.

Phase D: hostile audit of the owner synthesis

A hostile-referee audit corrected several quantifier and accounting overclaims. Preserve these corrections:

  • a no-remint compiler must explicitly state the fixed-prefix identity, centered residual, bank initialization, a quantitative ownership inequality, and domination of the actual physical correction by its envelope;
  • a recurrence needs nonnegative coefficients, uniform constants, a uniform positive determinant gap, and one global error per depth;
  • the curvature recurrence is an absolute-root-mass global frontier statement, not a theorem at every individual node;
  • a claimed “pathwise anti-circle” inequality was only an expectation/amortization statement;
  • the affine terminal needs the original column norm, cutoff range, and incoming-wall hypothesis;
  • a pointwise positive recurrence gap with constants deteriorating in (n,A,D) does not imply an (\exp(O(D^2))) terminal core.

The audit verified conditional algebra; it did not supply the universal compiler.

Phase E: the weighted breakthrough

The strongest positive result then appeared: a weighted extension of the Bansal--Jiang large-(k) method gives (O(\sqrt K)) hereditary discrepancy for matrices with column squared load (K\gtrsim\log^2 n). Rescaling (U) by (1/\tau) gives (\operatorname{lindisc}(U)=O(1)).

This eliminated “prove a better ultralight discrepancy estimate” as the main task. The remaining task became endpoint compatibility.

Phase F: exact refresh and terminal structure

Three exact tools were then proved:

  1. the all-center convex refresh radius is between hereditary discrepancy and twice hereditary discrepancy, with factor two sharp;
  2. a full-residual low-rank plus row-(\ell_1)-defect factorization gives a legitimate terminal rounding using only original columns;
  3. a spectral cutoff at resolution (D^2/h) supplies such a factor receipt whenever the target-resolution numerical rank is at most (e^{O(D^2)}).

These are genuine terminals, but separate heavy and light convex hulls still do not intersect automatically, and connected Walsh groups refute a universal low numerical-rank dichotomy.

Phase G: the one-law conditional compiler and its no-go sequence

The project formulated the exact lopsided local-lemma splice for one pinned ultralight-good law. It then killed progressively stronger but tempting choices of that law:

  • full exterior product domination is unaffordable for every hard-supported law on a random-block fixture;
  • the uniform ultralight-good law becomes ferromagnetic on a legal private-ancilla construction;
  • exact-mean maximum entropy and product I-projection coincide with that uniform law at a symmetric center, so both fail;
  • an (H)-aware convex optimizer that enforces the full lopsided system is feasible exactly when a common endpoint exists, so it is the target in convex disguise.

The surviving law question is narrower and still open: the actual heavy bad event after conjunctions only of nonneighbor heavy-good events under a construction-level sampler.

Phase H: deterministic mixed components and survivor bias

A parallel deterministic stream proved that pure components compose, while mixed components are payable under a pre-sign factor receipt. It also proved two warnings:

  • even the numerical inequality (\operatorname{herdisc}([H;U])\le\operatorname{herdisc}(H)+\operatorname{lindisc}(U)) is false on a legal path fixture;
  • an unconditional one-row martingale tail does not imply a statewise small dangerous span, because a rare crossing can be exactly the branch on which a full Walsh core survives.

The current probabilistic missing statement can equivalently be viewed as a survival-unbiased service at dyadic live-count stopping times.


6. Proved module I: weighted large-load discrepancy

The theorem

There are universal constants (C,c>0) such that, under the finite-precision and polynomial-input-size hypotheses used by the Bansal--Jiang source, if

[ V\in[-1,1]^{m\times n}, \qquad \sum_iV_{ij}^2\le K\quad(j\in[n]), \qquad K\ge c\log^2n, ]

then

[ \operatorname{herdisc}(V)\le C\sqrt K, \qquad \operatorname{lindisc}(V)\le2C\sqrt K. ]

For (V=U/\tau), declare (K=\lceil\tau^{-2}\rceil). Then

[ |U(y-x)|_\infty \le2C\tau\sqrt{\lceil\tau^{-2}\rceil} =O(1) ]

for every center (x\in[-1,1]^n) and some sign vector (y).

Proof spine

The proof uses a weighted version of the published large-(k) multilevel walk.

  1. Append the negative rows to make the result two-sided. This doubles squared load only by a constant.
  2. Discard entries of magnitude at most (1/n), costing at most a constant.
  3. Partition retained entries into dyadic magnitude subscales.
  4. Replace incidence-count frontiers by squared-mass frontiers. At level (\ell), maintain the monotone invariant [ \sum_{i:,i\text{ has ever reached level }\ell}V_{ij}^2\le K_\ell. ]
  5. Treat entry into a level as an adapted birth. Charge each born regularizer explicitly in the exponential potential.
  6. Use [ \alpha_\ell =6,2^\ell\max{1,K_\ell/\mu} ] for affine spectral independence; the maximum is essential at the final levels.
  7. Multiply each decoupling count by the squared entry scale and sum geometrically. The birth term contracts and the error term remains within the next squared-mass budget.
  8. The blocked rank is [ h/5+h/1000+1<h/3 ] while the walk runs, so the printed SDP remains feasible.
  9. Level costs sum geometrically to (O(\sqrt K)), and the tiny terminal phase costs (O(\sqrt{\mu\log n})=O(\sqrt K)).
  10. Obtain hereditary discrepancy by zero padding every column restriction back to the original ambient (n).
  11. Obtain arbitrary-center affine discrepancy from the classical exact inequality [ \operatorname{lindisc}(V)\le2\operatorname{herdisc}(V), ] not by assuming an unprinted arbitrary-start Brownian theorem.

What the theorem does not give

It gives a sign vector in the unrestricted cube. It does not say that the selected sign vector lies in a heavy-good code, respects old saturated faces, has exact mean under a useful law, or has useful heavy-row conditional tails. Conditioning the walk on its ultralight success event may radically change the endpoint distribution. This is the central remaining interface.

When citing this theorem, say “proved relative to the printed Bansal--Jiang SDP-feasibility and GeneralFreedman/decoupling black boxes, with a repaired adapted-birth proof.” Do not say it is printed verbatim in their paper.


7. Proved module II: exact all-center convex refresh

For

[ \mathcal S_\Delta(A,m) ={y\in{\pm1}^n:|A(y-m)|_\infty\le\Delta}, ]

define

[ \operatorname{crad}(A) =\sup_{m\in[-1,1]^n} \inf{\Delta:m\in\operatorname{conv}\mathcal S_\Delta(A,m)}. ]

The exact transference theorem is

[ \boxed{ \operatorname{herdisc}(A) \le\operatorname{crad}(A) \le2\operatorname{herdisc}(A).} ]

The factor two is sharp already for the one-column matrix ([1]).

Proof idea

For every linear objective, orient each hereditary signing toward the favorable halfspace during dyadic rounding. This gives an objective-preserving affine rounding. Finite-dimensional separation then places the center in the convex hull of relative-good signings. Carathéodory yields an exact-mean law on at most (n+1) atoms.

The theorem also holds for row-dependent discrepancy profiles. It is hereditary under column restriction.

Consequences

  • A hereditary receipt for one combined stacked wall matrix gives a literal common endpoint law after enlarging the wall by twice the receipt.
  • The aggregate local-lemma code for the coded low-heavy rows can be refreshed from every center at relative cost (2t_L=O(D)).
  • If an incoming center already lies within a low-heavy wall (T_0), the refreshed endpoint lies within (T_0+2t_L).

Critical limitation

Applying the theorem separately to (H) and (U) gives two convex hulls containing the same center. Their vertex sets can be disjoint. Even two centrally symmetric, near-full-entropy endpoint sets can have empty intersection. A combined bound on the stacked matrix would already solve the target; separate refresh laws do not fuse.

Also, the at-most-(n+1)-atom Carathéodory law may be too sparse to have the local randomness needed for a lopsided local-lemma splice. Exact mean is not a conditional-tail certificate.


8. Proved module III: the low-heavy LLL block and the exact conditional splice

The heavy dependency receipt

For a coded low-heavy row (i), let

[ S_i={j:H_{ij}\ne0}, \qquad E_i=\sum_jH_{ij}^2\le E_0. ]

Join two rows when their supports meet. The dependency degree is at most

[ d\le\lceil E_0\tau^{-2}\rceil (\lceil\tau^{-2}\rceil-1). ]

The product-law local lemma gives hereditary low-heavy discrepancy

[ O(\sqrt{\log(d+1)})=O(D). ]

This heavy code is a valid marginal result. It still needs to be combined with the ultralight endpoint.

Upstream support is part of the theorem

At the proposed splice center (m), let (\mathcal K_{\mathrm{up}}\subseteq{\pm1}^n) be the endpoint set allowed by every uncoded high-heavy row, saturated face, fixed coordinate, and earlier owner receipt.

A support-changing endpoint law (\mu) must satisfy

[ \mu(\mathcal K_{\mathrm{up}})=1. ]

It is not enough that the old posterior law had this support, or that (m) lies in (\operatorname{conv}\mathcal K_{\mathrm{up}}). A valid interface is one of:

  • same-hidden: (\mu\ll\nu) for an old law (\nu) already supported on (\mathcal K_{\mathrm{up}});
  • direct support: the new sampler itself proves support one on (\mathcal K_{\mathrm{up}});
  • retired rows: the high-heavy rows are constant on the current face or have a separately certified terminal bound.

No weaker mean or covariance condition transfers an upstream endpoint wall through a support change.

The exact one-law theorem

Fix a specified law (\mu), chosen by an explicit construction before a common endpoint is known, and assume

[ \mu{y:|U(y-m)|_\infty\le B}=1 ]

and (\mu(\mathcal K_{\mathrm{up}})=1).

Let the incoming low-heavy center wall be

[ |Hm|_\infty\le T_0. ]

For a proposed extra margin (r), define

[ \mathcal B_i={|H_iY|>T_0+r}. ]

If, for every row (i), every set (J) of nonneighbors of (i), and every positive-mass avoidance event,

[ \mu!\left( \mathcal B_i,\middle|, \bigcap_{j\in J}\overline{\mathcal B_j} \right) \le\frac1{e(d+1)}, ]

then the lopsided local lemma gives one atom (y) of the same law satisfying

[ |U(y-m)|\infty\le B, \qquad |Hy|\infty\le T_0+r, \qquad y\in\mathcal K_{\mathrm{up}}. ]

Useful sufficient tail condition

It is enough for the pinned construction to prove, for every low-heavy row (i), every set (J) of nonneighbors, every (t\ge0), and

[ F_J(t)=\bigcap_{j\in J}{|H_jY|\le T_0+t}, ]

that

[ \boxed{ \mu{|H_i(Y-m)|>t\mid F_J(t)} \le C_{\mathrm{tail}} \exp!\left(-\frac{t^2}{2\kappa E_i}\right).} ]

Then choose

[ r=\sqrt{2\kappa E_0 \log(C_{\mathrm{tail}}e(d+1))}. ]

If

[ \tau=e^{-aD^2}, \qquad \log C_{\mathrm{tail}}=O(D^2), \qquad \kappa=O(1), ]

then (r=O(D)).

The no-circle quantifier

The law must be fixed by a sampler or process before the desired common atom is found. Merely asserting that some supported law satisfies the lopsided inequalities is equivalent to the desired endpoint:

  • a favorable law implies a common atom by the local lemma;
  • a common atom gives the favorable point-mass law;
  • at a symmetric root, the antipodal pair on (y,-y) also restores exact mean zero.

Therefore the phrase “choose a canonical favorable law” is not a theorem. You must define its transition rule, randomness, failure behavior, and support without invoking the desired common endpoint, and then derive the conditional tails from its provenance.

This is the cleanest current probabilistic target.


9. Proved module IV: legitimate physical factor terminals

Suppose, before signs are chosen, the complete residual matrix has a factorization

[ A=Q+E, \qquad \operatorname{rank}Q\le p, \qquad |E|{\infty,1} :=\max_i\sum_j|E{ij}|\le\Delta. ]

Equivalently, in an owner factorization one may write (Q=LH_T^{\mathsf T}). The witness must describe the complete residual, not merely the covariance-visible action (AH_T).

Kernel compression keeps (Qx) fixed and leaves at most (p) fractional original coordinates. Rounding only those original Komlós columns gives

[ |A(y-x)|_\infty\le\sqrt p+2\Delta, ]

or, using the frozen affine small-core theorem, for every (0<\sigma<1),

[ \boxed{ |A(y-x)|_\infty \le\Phi(p,\sigma)+2\Delta,} ]

where

[ \Phi(p,\sigma) =C_0\sqrt{\log(8\max{1,p\lceil\sigma^{-2}\rceil})} +C_0\sigma\sqrt{p\log(32/\sigma^2)}. ]

Taking (\sigma\asymp p^{-1/2}) gives

[ \Phi(p,\sigma)=O(\sqrt{\log(2+p)}). ]

Thus

[ p\le e^{cD^2}, \qquad \Delta=O(D) ]

is an (O(D)) terminal receipt.

No column-norm bound on the quotient factor (Q) is needed. The final small-core call is made on the original columns, which retain the Komlós norm cap.

If disjoint support components have separate factor receipts, their discrepancy costs compose by a maximum, not a sum, because each physical row lies in one component.

Why owner covariance alone is insufficient

Posterior-visible owner data factor through (AH_T). Matrices (A) and (A+K) with (KH_T=0) are indistinguishable to trace, operator norm, nuclear norm, stable rank, leverage, and row-variance statistics of the owner-visible covariance. A rank-one clone/beacon fixture places both an easy flat matrix and an arbitrary hidden Komlós core in the same visible fiber.

Therefore a small owner covariance is not a terminal. You need a signing-independent approximation of the complete residual with small rank and small row-(\ell_1) defect, or a different restricted theorem.


10. Proved module V: target-resolution spectral gate

Let (B) be a residual matrix on (h) live columns and put (M=B^{\mathsf T}B). For (\lambda>0), let (P_\lambda) project onto eigenvalues of (M) larger than (\lambda), and write

[ Q=BP_\lambda, \qquad E=B(I-P_\lambda). ]

Then

[ \operatorname{rank}Q\le N_\lambda(M), \qquad |E|_{\infty,1}\le\sqrt{h\lambda}. ]

At the target resolution

[ \lambda=\frac{D^2}{h}, ]

the defect is at most (D). Therefore

[ N_{D^2/h}(B^{\mathsf T}B)\le e^{cD^2} ]

is a literal (O(D)) factor terminal.

Ridge effective dimension and log determinant certify the threshold rank when small:

[ N_\lambda(M) \le2\operatorname{tr}(M(M+\lambda I)^{-1}), ]

[ N_\lambda(M) \le\frac{\log\det(I+M/\lambda)}{\log2}. ]

These are positive gates, not universal dichotomies.

Exact firewalls

  • Exact rank is discontinuous under near-clones and is the wrong parameter.
  • Connected multi-Walsh groups satisfy the column-load and low-heavy budgets while having target-resolution numerical rank ((1-o(1))h) and extensive log determinant.
  • An arbitrary number of Walsh groups can be bridge-connected, preventing free componentwise maximum composition.
  • For these groups, any approximation with (O(D)) row-(\ell_1) defect may require arbitrarily large rank.
  • Computing spectral data only on an owner-visible image misses a hidden Walsh fiber.

Use the spectral receipt aggressively when it passes. Do not spend time trying to prove it passes universally from the current hypotheses; that claim is already refuted.


11. Earlier common-law localization: still useful, but no longer the whole problem

Let (\mu) be a full-support law on a finite code (\mathcal C\subseteq{\pm1}^h), with mean (m) and covariance (\Sigma). Then

[ \operatorname{rank}\Sigma =\dim\operatorname{aff}(\mathcal C), ]

so

[ \dim\ker\Sigma \le h-\log_2|\mathcal C|. ]

Given a predictable covariance (V_t\succeq0) with

[ \operatorname{range}V_t\subseteq\operatorname{range}\Sigma_t, ]

define

[ Q_t=\Sigma_t^\dagger V_t\Sigma_t^\dagger ]

and evolve the finite-support weights by

[ d\mu_t(y) =\mu_t(y) \langle y-m_t,Q_t^{1/2}dB_t\rangle. ]

Then

[ dm_t=\Sigma_tQ_t^{1/2}dB_t, \qquad d\langle m\rangle_t =\Sigma_tQ_t\Sigma_t,dt =V_t,dt. ]

Thus a Bansal--Jiang admissible covariance can be realized exactly in the mean of one law supported on the existing heavy code. Heavy legality is pathwise because the mean remains in the code's convex hull, and a terminal vertex is a codeword.

This remains a plausible way to construct the missing pinned law. However, a complete proof must address:

  • finite-support SDE existence up to weight extinction;
  • predictable measurable selection of (V_t);
  • range feasibility after support shrinks;
  • finite stopping and discretization;
  • how the weighted multilevel birth process is represented inside this posterior law;
  • support one on all high-heavy and saturated-face constraints;
  • event-specific heavy tails after nonneighbor avoidance conditioning.

The old endpoint anti-collision theorem is also available: if a cube martingale has covariance density (V_t\preceq\alpha\operatorname{diag}V_t), then for a terminal pattern on (J), conditionally at time (s),

[ \Pr[X_{\tau,J}=\sigma\mid\mathcal F_s] \le \prod_{j\in J} \left(\frac{1+\sigma_jX_j(s)}2\right)^{1/\alpha}. ]

This prevents certain endpoint collisions and yields min-entropy, but it does not by itself prove the nonneighbor-conditional heavy tails or subtract posterior component constraints. Do not promote it to a common-endpoint theorem without that missing work.


12. Deterministic simultaneous terminal and the survivor-bias firewall

Mixed support components

Form the bipartite support graph of the vertical stack ([H;U]), with row vertices and column vertices. A connected component is heavy-pure, light-pure, or mixed.

Let

[ D_H=\operatorname{herdisc}(H), \qquad L_*(U)=\sup_{J\subseteq[n]}\operatorname{lindisc}(U_J). ]

If every original mixed component (C) has, before signs are inspected, a factorization

[ A_C=Q_C+E_C, \qquad \operatorname{rank}Q_C\le p, \qquad |E_C|_{\infty,1}\le\Delta, ]

then

[ \operatorname{herdisc}([H;U]) \le \max{D_H,L_*(U),\Phi(p,\sigma)+2\Delta}. ]

The proof is hereditary because restricting columns only refines components and cannot increase rank or row-(\ell_1) defect.

Consequences:

  • if every mixed component has at most (e^{cD^2}) columns, then the stack has discrepancy (O(D));
  • if every mixed component passes the target-resolution spectral gate, then the stack has discrepancy (O(D));
  • pure components are never a compatibility problem;
  • the genuinely hard case is a large, connected, spectrally broad mixed component.

The exact deterministic question is:

[ \boxed{ \text{Can every large mixed component be decomposed into row-disjoint payable pieces,} \text{ or supplied with a broader noncircular terminal receipt?}} ]

Connected Walsh fixtures say the current rank-plus-row-(\ell_1) answer is not universally yes.

Naive numerical subadditivity is false

A legal scaled path-incidence heavy block plus one ultralight row gives a connected mixed component for which the heavy-optimal signs and light-optimal signs are incompatible. It refutes the numerical inequality

[ \operatorname{herdisc}([H;U]) \le \operatorname{herdisc}(H)+\operatorname{lindisc}(U). ]

The ratio in the fixture approaches two. This is not a growing lower bound, but it proves that “add the two discrepancy values” is not even a valid abstract theorem.

One-row martingale tail

Let (X_t) be a continuous cube martingale with

[ d\langle X\rangle_t=\Sigma_t,dt, \qquad \Sigma_t\preceq\gamma\operatorname{diag}\Sigma_t. ]

For a row (a) with

[ \sum_ja_j^2\le E_0, \qquad \max_j|a_j|\le\rho, \qquad \frac{B\rho}{E_0}\le\frac12, ]

the regularizer

[ Y_t=\langle a,X_t\rangle +\frac{B}{2E_0}\sum_ja_j^2(1-X_t(j)^2) ]

gives

[ \Pr\left{\sup_t|\langle a,X_t\rangle|\ge B\right} \le2\exp!\left(-\frac{B^2}{9\gamma E_0}\right). ]

This closes the scalar-tail calculation under a max-entry condition (\rho\lesssim E_0/D).

Survivor bias

The scalar tail does not imply a small dangerous span at a retained state. A legal fixture uses a common random-walk prefix and an untouched normalized Walsh core. The rare event that the prefix crosses the barrier is made exactly the branch on which the Walsh core remains live. On that branch,

[ N_T=h_T=s, ]

where every one of the (s) rows is dangerous and the live gradient rank is (s). Off the branch, both are zero. Unconditionally each row crosses with exponentially small probability, but conditionally on survival the probability is one.

Thus

[ \text{unconditional one-row tail} \centernot\Longrightarrow \text{statewise active-danger dimension}. ]

The minimal missing statement for a dyadic analysis is a survival-unbiased service: at every dyadic live-count stopping time (T_q), conditional on the retained live state,

[ \mathbb E[N_q\mid\mathcal L_{T_q}] \le pK h_{T_q}, ]

where (N_q) counts rows dangerous during that live-set service. If this held and

[ Kp\log n=o(1), ]

then Markov plus a union bound over (O(\log n)) epochs would keep the blocked fraction small. The Walsh fixture proves that this statement cannot be inferred from column incidence and unconditional scalar tails alone. It must come from the actual sampler or a rule that prevents rare-event survival selection.


13. Hostile fixtures and what each one means

Every new theorem must be tested against this table. A failure against a fixture does not refute the Komlós conjecture; it refutes the stated mechanism.

Fixture What it refutes What it does not refute
Exact (4\times4) disjoint heavy/light codes Adding code deficits or intersecting independently constructed large codes A literal common-law construction
Pair-equality superatoms and seven-coordinate ({\pm p}) factor Raw coordinate entropy and a false universal anti-collision constant Component-relative entropy with exact slice completion
Majority-pair posterior Charging only the root code deficit Current-cylinder deficit (q_v=h_v-\log_2
Aggregate covariance-inflation example Confining entropy loss to a small coordinate reservoir Subspace or component accounting
487-column one-death valve Immediate release of every forced-null direction A predeclared delayed physical death or one immutable leakage owner
Ultra-diffuse paired wall Free same-wall retirement Root coding, a burned reservoir, or a paid structural release
Spectral cube-wall (I-J/h) at a near-corner One-shot spectral-owner endpoint jumps Infinitesimal common-law localization
Legal path heavy/light fixture Choosing separate optimal signs; even naive numerical subadditivity A simultaneous theorem on the full mixed component
Random (q\times q) ultralight hard wall, (q\asymp\tau^{-2}) Full exterior Radon--Nikodym domination with (\log\Lambda=O(D^2)) for any hard-supported law Row-tail-only lopsided conditioning
Private-ancilla Curie--Weiss fixture Uniform light-good law, exact-mean maximum entropy, centered-product conditioning, product I-projection Existence of common good endpoints or an engineered anti-ferromagnetic sampler
Canonical convex/Farkas alternative Calling an optimizer that enforces all lopsided constraints a new construction Proving those constraints from an independently specified process
Near-clones Exact rank as a stable terminal invariant Target-resolution numerical rank
Connected multi-Walsh groups A universal small numerical-rank or log-determinant gate Applying the spectral gate on instances where it actually passes
Rank-one beacon with hidden Walsh fiber Owner-visible covariance scalars as full physical receipts A pre-sign factorization of the complete residual
Walsh survivor-bias fixture From unconditional one-row tails to conditional dangerous-span bounds A genuinely survival-unbiased sampler theorem
Equality blocks Nodewise pure-curvature contraction and endlessly renewable boundary service Global frontier accounting plus permanent release
Exposed physical face Further same-law positive variance in a saturated normal A support-changing terminal with an independent certificate
Three-class actual-dual lift Unrestricted direct recode of any actual owner A genuinely restricted low-variance, small-core, or factor terminal
Product-root rare tube Treating a rare (D^2)-variance terminal event as an easier premise Using such an event after it is derived from a valid compiler
Nonlinear shell cage Total-correlation or affine-hull scalars as universal physical service A process-generated matrix--code invariant or full residual terminal
Fair reset and cosine renewal fixtures Free independent renewal after purification A correlated state with explicit accumulated receipts

Two fixtures deserve detailed recollection.

The Curie--Weiss endpoint-law no-go

Fix a constant light wall (B), small cutoff (\tau), take an even (L\ge B/\tau), and put

[ \eta=B/L, \qquad r=L^2, \qquad s=64L. ]

For every edge (e={u,v}) of the complete graph on (s) target signs, add (r) private ancillas and the ultralight row

[ U_e(x,z) =\frac\eta2(x_u-x_v)+\eta\sum_{a=1}^rz_{e,a}. ]

Under the uniform law on the light-good endpoints, the ancilla completion count makes the target marginal exactly

[ \mu_X(x)\propto \exp!\left(\frac{J_L}{2}\left(\sum_vx_v\right)^2\right), \qquad J_L\ge\frac1{16L}, \qquad sJ_L\ge4. ]

For the legal energy-one heavy row

[ H_0=\frac1{\sqrt s}\mathbf1_s, ]

the probability of discrepancy larger than (\sqrt s/2) tends to one. There is only one heavy row, so its dependency degree is zero and the failure is unconditioned. The light-good set is centrally symmetric; therefore uniform, exact-mean maximum entropy, and the centered-product I-projection are the same failed law.

Balanced common endpoints still exist. This is a law no-go, not a Komlós lower bound. It tells you that the sampler must be deliberately anti-ferromagnetic or otherwise subgaussian; maximum entropy is not neutral.

The exact Farkas boundary

For heavy bad events (\mathcal B_i), nonneighbor avoidance events (F_{i,J}), and target bounds (p_i), define

[ c_{i,J}(y) =\mathbf1_{F_{i,J}}(y) (\mathbf1_{\mathcal B_i}(y)-p_i). ]

The lopsided constraints are linear in a law:

[ \mathbb E_\mu c_{i,J}(Y)\le0. ]

Finite-dimensional Farkas separation says exactly one of the following holds:

  1. there is an exact-mean law on the light-good set satisfying all inequalities;
  2. there are a vector (q) and nonnegative multipliers (\lambda_{i,J}) such that [ q\cdot(y-m)+\sum_{i,J}\lambda_{i,J}c_{i,J}(y)>0 ] for every light-good endpoint (y).

At mean zero with symmetric events, feasibility is equivalent to existence of a common endpoint. This is useful because it identifies an exact obstruction certificate, but solving the unrestricted feasibility program is not progress. A legitimate route would prove that dual certificates arising against a particular weighted-walk sampler have an additional structure that is impossible under the column-load geometry.


14. The exact open theorem to attack first

The cleanest noncircular target is the following. You may adjust constants and the precise sampler, but not weaken the same-endpoint or conditioning quantifiers without saying so.

Candidate theorem: engineered first-owner conditional tails

Fix (D\ge1), (\tau=e^{-aD^2}), and a stopped residual on (h) live original columns. Split it as

[ A=H+U, ]

where:

  • every column of the stacked split has squared norm at most one;
  • (|U_{ij}|\le\tau);
  • every coded low-heavy row (H_i) has energy (E_i\le E_0) and every nonzero entry exceeds (\tau);
  • the current center is (m\in[-1,1]^h) with (|Hm|_\infty\le T_0=O(D));
  • (\mathcal K_{\mathrm{up}}) is the endpoint set allowed by every high-heavy row, saturated face, fixed coordinate, and earlier immutable receipt.

Construct a randomized endpoint law

[ \mu=\operatorname{Law}(\mathsf S(U,H,m,\mathcal R_{\mathrm{up}};\omega)) ]

by a fully specified sampler (\mathsf S). The sampler may use the matrix and structural heavy supports, but it may not search for or condition on a common heavy/light-good endpoint, solve the full lopsided feasibility problem, or choose its law after inspecting a successful terminal atom.

Prove:

  1. finite termination: (\mathsf S) returns an original sign vector almost surely or has a finite discretization with explicit success and failure handling;
  2. ultralight hard support: for a universal (B), [ \mu{y:|U(y-m)|_\infty\le B}=1; ]
  3. upstream hard support: [ \mu(\mathcal K_{\mathrm{up}})=1; ]
  4. event-specific lopsided tails: for every coded row (i), every set (J) of nonneighbors of (i), every (t\ge0), and every positive-mass [ F_J(t)=\bigcap_{j\in J}{|H_jY|\le T_0+t}, ] one has [ \mu{|H_i(Y-m)|>t\mid F_J(t)} \le C_{\mathrm{tail}} \exp!\left(-\frac{t^2}{2\kappa E_i}\right), ] with (\kappa=O(1)) and (\log C_{\mathrm{tail}}=O(D^2)), uniformly in (m,n,A) and the stopped state;
  5. heredity: the same construction and constants apply after every column restriction or live-face transition used by the compiler;
  6. no hidden benchmark loss: the covariance budgets and error probabilities remain compatible with the Bansal--Jiang baseline branch.

Then the proved one-law local-lemma theorem gives one common atom with

[ |U(y-m)|\infty=O(1), \qquad |Hy|\infty=O(D), \qquad y\in\mathcal K_{\mathrm{up}}. ]

This would close the first-owner heavy/light splice. A complete universal proof would then still need to show that every earlier high-heavy receipt reaches this splice legally and that the process has finite stopping and one-use owner accounting, but it would remove the last new ultralight obstruction.

Why this theorem is neither too strong nor known false

Full exterior product domination is false. Uniform and maximum-entropy laws are false. Bare existence of a favorable law is circular. However, none of the current no-go fixtures proves that the actual weighted-walk endpoint law fails the much narrower nonneighbor-conditional heavy-row inequality above. The Curie--Weiss fixture attacks a specific canonical law; the random-block fixture attacks full density domination; neither refutes an engineered construction-level sampler.

An honest negative breakthrough would be an actual first-owner state showing that every law generated by a broad, precisely defined class of weighted endpoint samplers violates the lopsided condition, together with a noncircular replacement theorem. A counterexample to one arbitrary convex decomposition is not enough.


15. Ranked research program

Priority 1: pin the actual weighted endpoint law

The weighted theorem currently proves existence through a randomized multilevel walk and high-probability promotion bounds. Extract the precise law instead of referring to “the weighted-BJ endpoint.” Specify:

  • the probability space and all random seeds;
  • the SDP optimizer selection rule;
  • the adapted birth and promotion stopping times;
  • whether bad algorithmic outcomes are rejected, restarted, or included;
  • whether conditioning on overall ultralight success is used;
  • the exact terminal rounding of the last coordinates;
  • the center from which the walk starts;
  • whether the law has exact mean (m), approximate mean, or no mean guarantee;
  • whether its hard support is the light-good set or only high-probability;
  • how support one is achieved without conditioning in a way that destroys heavy tails.

This bookkeeping is not clerical. The endpoint distribution can change drastically when conditioned on global success, and the desired theorem is about that actual distribution.

Priority 2: derive the one-row tail from the sampler itself

Reprove the regularizer inequality inside the exact weighted process. For a heavy row (a), track

[ G_t=\sum_ja_j^2(1-X_t(j)^2), ]

and

[ Y_t^\pm =\pm\langle a,X_t-m\rangle+\beta G_t. ]

Under (\Sigma_t\preceq\gamma\operatorname{diag}\Sigma_t), the drift is negative and the quadratic variation is bounded by a constant times the energy clock. Pin all constants, including the condition (B\max_j|a_j|/E_0\le1/2), stopping near cube faces, and the effect of terminal rounding.

This scalar step is largely solved, but it must be integrated with the actual endpoint law rather than cited abstractly.

Priority 3: prove the nonneighbor conditioning theorem

This is the real work. Explore only hypotheses supplied by the sampler. Plausible directions include:

  1. A commuting resampling oracle. Show that a nonneighbor heavy event can be resampled through randomness localized away from another row's support while the global ultralight wall remains hard-supported.
  2. A conditional Ville/Doob argument. Represent avoidance of nonneighbor alarms as a stopping event whose likelihood process has controlled covariation with the target row's exponential supermartingale.
  3. Construction-level anti-ferromagnetism. Modify the endpoint measure or walk potential to penalize correlated target magnetization while preserving the weighted promotion proof. The Curie--Weiss fixture says this bias must be deliberate.
  4. Local witness factorization of the global SDP. The weighted SDP couples all coordinates, so disjoint heavy supports are not automatically independent. Look for a decomposition of the random seed or covariance increments that makes nonneighbor avoidance conditionally harmless up to (e^{O(D^2)}).
  5. A restricted Farkas-dual impossibility theorem. Instead of solving the full lopsided LP, prove that any separator compatible with the weighted potential and birth invariants violates the column squared-load budget.

For every proposal, explicitly test whether it accidentally proves full exterior domination. If so, it is almost certainly too strong and already refuted.

Priority 4: make high-heavy support literal

The low-heavy local lemma does not include rows of energy above (E_0). Decide how they reach the first-owner splice:

  • same-hidden common-law localization inside their existing hard code;
  • monotone saturated-face tangency until those rows are constant on the live face;
  • direct inclusion of their endpoint slabs in the sampler support;
  • or a separately certified factor/spectral/small-core terminal.

Write the support statement before doing tail algebra. A center wall is not an endpoint wall.

Priority 5: prove survival-unbiased service or avoid survivor selection

If direct lopsided conditioning is intractable, formulate the exact dyadic statewise service needed by a recursive walk. The theorem must condition on the retained live state, not merely on the past before the survival decision. Either:

  • design the sampler so the live-set selection is independent or negatively coupled to heavy crossings;
  • randomize ownership before seeing the crossing;
  • terminally pay the rare survivor core by a predeclared small-core receipt;
  • or establish a complete-frontier absolute-mass inequality that is sufficient without a nodewise claim.

The Walsh survivor fixture should be run as the first unit test.

Priority 6: develop the deterministic mixed-component alternative

Try to prove a structural decomposition theorem weaker than global low rank:

  • peel row-disjoint pure or factor-payable pieces;
  • cut sparse articulation columns and charge their original-column rounding once;
  • identify a bounded-treewidth, bounded-cycle-rank, or separator parameter compatible with hereditary restrictions;
  • allow each connected Walsh group to be a small core, but prevent an arbitrary number of groups from being bridge-connected into one unpaid component;
  • search for an (O(D))-defect decomposition whose ranks compose by a maximum across genuinely row-disjoint pieces rather than a sum.

Any structural theorem must survive the bridge-connected Walsh firewall. If it fails, isolate the exact obstruction and determine whether that obstruction itself carries an exploitable signing.

Priority 7: use owner machinery only with a new restricted receipt

The broad owner route has been audited deeply. Reopen it only if you can state a hypothesis that excludes the rank-one hidden-fiber fixture and the three-class actual-dual hardness lift. Legitimate gates include:

  • total remaining marginal variance at most (D^2);
  • absolute live core size at most (e^{cD^2});
  • a pre-sign full-residual factor receipt;
  • a target-resolution spectral receipt that actually passes;
  • a nested immutable service with one-use physical envelopes.

Do not claim that low owner dimension, low owner trace, or actual-dual provenance is enough.

Priority 8: compile and calibrate

Once a new lemma is proved, do not stop. Compile it into:

  1. a legal root-to-endpoint algorithm or existence proof;
  2. hereditary column restrictions;
  3. finite stopping;
  4. accumulated discrepancy accounting;
  5. support preservation for all matrix blocks;
  6. the exact (D=\log\log n) and (D=(\log n)^\varepsilon) arithmetic;
  7. a root-level Bansal--Jiang fallback;
  8. a final theorem with explicit constants and no hidden dependence on (A,m,n).

16. Suggested first forty-eight hours

Hours 0--4: reconstruct only the indispensable baseline

  • Re-derive the Bansal--Jiang quarter-power budget and scale cancellation in Section 3.
  • Re-derive the weighted theorem's adapted-birth lemma and affine parameter correction.
  • Re-derive the one-law local-lemma implication and the (r=O(D)) arithmetic.
  • Do not read all historical owner modules before starting; use the no-go table to avoid known circles.

Deliverable: a two-page internal sheet containing only assumptions, laws, filtrations, and conclusions.

Hours 4--12: define the sampler with no ambiguity

  • Extract the weighted multilevel process into a precise finite or stopped continuous algorithm.
  • Write its endpoint law before conditioning on success.
  • Identify exactly which event makes the ultralight wall hard-supported.
  • Determine whether rejection sampling or restarts preserve any usable conditional tail.
  • Mark every dependence on (H) and every upstream support constraint.

Deliverable: pseudocode plus a theorem saying exactly what law is produced, with no tail claim yet.

Hours 12--24: attack one target row

  • Prove the scalar exponential supermartingale under the exact covariance rules.
  • Track adapted births, promotions, and terminal rounding.
  • Test the proof on the Curie--Weiss target row and explain why the engineered law does not become uniform on its hard wall.

Deliverable: an unconditional row-tail theorem for the actual law, or a concrete counterexample to it.

Hours 24--36: attack one nonneighbor conditioning

  • Start with one target heavy row and one nonneighbor avoidance event.
  • Compute the likelihood-ratio or resampling identity exactly.
  • Bound its effect on the target exponential supermartingale.
  • Generalize only after the two-event calculation works.

Deliverable: either a two-event conditional-tail lemma with constants, or an exact obstruction explaining which global covariance coupling breaks locality.

Hours 36--48: choose decisively

  • If the conditional calculation works, build the induction over a conjunction of nonneighbors and check (\log C_{\mathrm{tail}}=O(D^2)).
  • If it fails on the actual process, freeze the counterexample and pivot to a modified anti-ferromagnetic sampler or deterministic component decomposition.
  • Do not rename the failed law and repeat the same entropy optimization.

Deliverable: a proved candidate lemma, or a no-go theorem that genuinely narrows the remaining sampler class and names the next replacement.


17. Anti-circle protocol

Before writing “breakthrough,” answer all of the following.

Endpoint identity

  • Are the heavy and light inequalities satisfied by the same sign vector?
  • If two laws or convex hulls appear, where is their literal common atom proved?
  • If a support change occurs, which incoming endpoint constraints have probability one afterward?

Law provenance

  • Is the endpoint law fixed by a stated construction before a common atom is known?
  • Did you condition on global success, and if so, did you recompute every tail under the conditioned law?
  • Does the law secretly solve the full lopsided feasibility problem?
  • Could a point mass on the desired endpoint make the hypothesis true? If yes, the existential statement is likely circular.

Conditioning

  • Is the claim unconditional, history-conditional, cylinderwise, or averaged over a frontier?
  • If the recursion retains rare branches, have you ruled out Walsh survivor bias?
  • Are nonneighbor events genuinely local under the sampler, or only disjoint in matrix support while coupled by a global SDP?

Owners and releases

  • Is every owner assigned before its favorable release event?
  • Is it immutable and charged once?
  • Does release delete or terminally certify a named physical component?
  • Is the physical correction bounded row by row by a stated envelope?
  • Are you restarting an entropy or dependence clock after conditioning?

Terminal legality

  • Does the terminal use original columns with norm at most one?
  • Is a full residual factorization supplied before signs are chosen?
  • Is incoming discrepancy added explicitly?
  • If a quotient or shadow is rounded, where is the residual defect paid?

Benchmark

  • Can the algorithm run the Bansal--Jiang branch on the untouched root input?
  • If the new lemma fails, does the claimed universal guarantee actually improve anything?
  • Are hidden polylogarithmic factors tracked honestly?

If any answer is missing, label the result conditional and do not continue building downstream on it as if it were proved.


18. Verification and hostile-referee workflow

Mathematical verification

For every theorem, create a proof object with this header:

Claim:
Status:
Universal quantifiers:
Fixed parameters and constant dependence:
Input cylinder / center / filtration:
Output endpoint or release:
Same-endpoint statement:
Support-preservation statement:
Entropy or discrepancy cost:
Termination:
External black boxes and exact hypotheses:
Known hostile fixtures:
What the claim does not prove:

Then perform three audits.

  1. Quantifier audit: swap the order of every existential and universal quantifier and check that the printed order is the one proved.
  2. Conditioning audit: write every conditional law and denominator. Treat zero-probability conditioning through a closed convex convention or exclude it explicitly.
  3. Transfer audit: trace the final sign vector back to the original columns and rows, including every factor-two, incoming wall, shadow, and terminal increment.

Computational verification

Use exact rational or symbolic checks whenever possible. Randomized floating-point experiments are diagnostics, not proofs. For every frozen module:

  1. record SHA-256 hashes for the theorem note, verifier, manifest, and pinned sources;
  2. run the checksum check before the verifier;
  3. disable bytecode output to avoid modifying imported folders;
  4. report the number and type of checks;
  5. state what the verifier does not verify;
  6. run exhaustive small cases for parity, conditional probabilities, support components, and LP/Farkas fixtures;
  7. separately audit asymptotic inequalities and threshold existence;
  8. pin all external paper source files used as black boxes.

Typical non-mutating verification invocation:

sha256sum -c SHA256SUMS.txt
PYTHONDONTWRITEBYTECODE=1 python3 verify_module.py

On a platform without GNU sha256sum, recompute SHA-256 with an equivalent tool and compare the literal digest list.

Interpretation of PASS

A verifier can establish that:

  • a finite fixture has the claimed exact discrepancy;
  • a displayed algebraic identity holds over enumerated cases;
  • a recurrence has the claimed eigenvalues;
  • constants satisfy inequalities over a tested parameter range;
  • imported files match frozen hashes.

It cannot by itself establish:

  • a stochastic process exists with all stated measurable selections;
  • a source black box applies outside its printed hypotheses;
  • a finite pattern proves a universal asymptotic theorem;
  • the open common-endpoint sampler exists;
  • the universal Komlós theorem has been proved.

19. Artifact map from the previous filesystem

This prompt is self-contained, so the new filesystem need not preserve these paths. If the old artifacts are copied, retain their relative folder names and verify hashes before using them. The old project root was logically

komlos_breakthrough_agent_2026-08-23/

The most useful artifacts were:

Relative artifact Role Status at snapshot
integration/komlos_breakthrough_ledger_2026-08-23.md Older one-walk and restricted ultralight integration ledger Historical; later ultralight frontier superseded
universal_release_attack_2026-08-23/final_breakthrough_ledger_2026-08-23/komlos_downstream_breakthrough_ledger_2026-08-23.md Full owner/release synthesis and no-go ledger Frozen
universal_release_attack_2026-08-23/synthesis_hostile_audit/hostile_referee_audit_2026-08-24.md Quantifier and recurrence corrections Audit note
universal_release_attack_2026-08-23/published_splice_audit_2026-08-24/published_large_k_ultralight_splice_boundary_2026-08-24.md Initial weighted theorem and exact heavy-shadow firewall Frozen, but use repaired hostile audit for the theorem proof
universal_release_attack_2026-08-23/weighted_theorem_hostile_audit_2026-08-24/weighted_large_load_hostile_audit_2026-08-24.md Repaired weighted theorem Frozen and central
universal_release_attack_2026-08-23/integer_hull_release_2026-08-24/hereditary_integer_hull_transference_2026-08-24.md Sharp all-center convex refresh Frozen and central
universal_release_attack_2026-08-23/physical_owner_sdp_2026-08-24/physical_owner_sdp_fiber_boundary_2026-08-24.md Full-residual factor terminal and owner-fiber no-go Frozen and central
universal_release_attack_2026-08-23/spectral_danger_span_2026-08-24/spectral_danger_span_boundary_2026-08-24.md Target-resolution spectral gate and Walsh firewall Frozen and central
universal_release_attack_2026-08-23/heavy_compatible_exit_2026-08-24/heavy_compatible_ultralight_exit_gate_2026-08-24.md Exact one-law lopsided splice Frozen and central
universal_release_attack_2026-08-23/local_domination_law_2026-08-24/local_exterior_domination_no_go_2026-08-24.md Full exterior-density no-go Frozen
universal_release_attack_2026-08-23/canonical_light_law_tail_2026-08-24/canonical_uniform_light_law_ferromagnetic_no_go_2026-08-24.md Curie--Weiss no-go Frozen
universal_release_attack_2026-08-23/canonical_light_law_2026-08-24/canonical_convex_endpoint_law_boundary_2026-08-24.md Max-entropy/I-projection identification and exact Farkas boundary Frozen
universal_release_attack_2026-08-23/wall_safe_refresh_2026-08-24/wall_safe_refresh_boundary_2026-08-24.md Wall-safe exact refresh interface Frozen
universal_release_attack_2026-08-23/simultaneous_heavy_light_terminal_2026-08-24/simultaneous_heavy_light_terminal_boundary_2026-08-24.md Mixed-component theorem, path fixture, one-row tail, survivor bias Live draft snapshot; verifier passed, no frozen manifest at snapshot
universal_release_attack_2026-08-23/owner_downstream_synthesis_2026-08-24/no_circle_owner_downstream_ledger_2026-08-24.md Concise latest synthesis and BJ calibration Live draft snapshot; verifier passed, no frozen manifest at snapshot

Frozen hashes for the central files

File SHA-256
weighted_large_load_hostile_audit_2026-08-24.md 3477ee9e11d21a956545cd9e67e6a673abc792b22a15bc2298acfbcbb3dd450c
verify_weighted_large_load_hostile_audit.py 08276a4e4f65cc5de0a0f28a0f2228da3e4015fabfc08a36d01d3d5c97fb246a
hereditary_integer_hull_transference_2026-08-24.md c5d3824a30f98b717164660767aab02c615e517b3921d4438b5a84f017f8ec82
verify_hereditary_integer_hull_transference.py b13cef36707b6b05207fdbddb54b946969db86eb87509791ac3069711ce2ab34
physical_owner_sdp_fiber_boundary_2026-08-24.md d9a570d0f777beadf44e82cc7e6685e3b8a9cc849bf04a6ab8732631e068e90f
verify_physical_owner_sdp_fiber_boundary.py e2411622246950c059b42d53ccfe6a909330efac9213da66ffe01b36b26ebed3
spectral_danger_span_boundary_2026-08-24.md ef22184300547b7139ff77d6bf5f2f1d54ccd51743b5d2b0b3d9aa2739be57a3
verify_spectral_danger_span.py dd9c61811a113f246efea94e14e01f3d17aa64044c986143dec167e4772cfd8c
heavy_compatible_ultralight_exit_gate_2026-08-24.md 5fead16e81f50d64d6b0efe4784e48027f8a0e220189c826870e9826263f1b01
verify_heavy_compatible_ultralight_exit_gate.py d233a107905ac8f8c6029b1dc6ef7639e201fb5bacb677d8a043f23955684734
local_exterior_domination_no_go_2026-08-24.md 85e32b1deafd9a53b2f4ffa435794f2c009d5672c0249afcfd30a09736a9006a
verify_local_exterior_domination_no_go.py d96a0e7ae633e2fe10c4ea7b4f0a8a67f5153e1a36d94dc4a745460c0bddf347
canonical_uniform_light_law_ferromagnetic_no_go_2026-08-24.md e33c86e3c77a6e82fb31f14a127104d1330e97dbeb1d2279a475d2a0312635ad
verify_ferromagnetic_light_law_fixture.py bf1eb2cf832f66af1a68d80f00e424ef3e0f36cf0c8b6b8b754bbceba9d92179
canonical_convex_endpoint_law_boundary_2026-08-24.md da1ab551accb46e05e95093cda2323bdc647e9f3ea2b1902827f30e0c4c61fc1
verify_canonical_convex_endpoint_law_boundary.py 2a4f3d92a1cc5ad008e81971b55913f27e5a1cfc1554d11b0e3e758e39494ae7
wall_safe_refresh_boundary_2026-08-24.md b44eebbad3687693e5af22b3c42944f4a941b3d4679699e5393b3579e7addc26
verify_wall_safe_refresh.py ae1ec7442a34d8ee8776c53cb62fd5e35f42df3a3fcddd9a34d4139e14a04688
komlos_downstream_breakthrough_ledger_2026-08-23.md 52b0955f6c7dd0c66450db77e4e7497ca3e56feeeaf9182cd041dc4b3156891d
verify_final_breakthrough_ledger.py 8d7589ccc4223ba94d64422a7613c65d6f0cfb07fbe92e624038403c0dd9ebf1
published_large_k_ultralight_splice_boundary_2026-08-24.md c9d8cfe6e9321352637a34df49b460612da7aec62ad9065c2adb3591ca0f5c10
verify_published_large_k_ultralight_splice.py 2c0a6be785ad0a97d0906dcb6b52b3270e7a3e0ea45a5d3122b4000f6c63f48b

Live draft snapshot hashes

These two files were still active drafts and had no frozen manifest at the snapshot time. Treat the hashes as identifiers of the inspected version, not canonical frozen receipts.

File Snapshot SHA-256 Snapshot verifier result
no_circle_owner_downstream_ledger_2026-08-24.md 1d1ebbba50f93727d6af67fd26d60937d24a91f924f4bab6fc5b12b01e2c5985 PASS no-circle owner/downstream synthesis: 68119 checks
verify_no_circle_owner_downstream.py b68832b4a3621ab780370ab6a295f36c2585a419832df656f7c5a4fe87935eeb same run
simultaneous_heavy_light_terminal_boundary_2026-08-24.md fdb1790477ed352c05f9c2386340c05f16f4e1bc1e76c5d1dfbf6523b02ac64d PASS: 12455 simultaneous/component/factor/source checks
verify_simultaneous_heavy_light_terminal.py be41cdb3bae2ad8fe5c689ee3c732ccd6ef4d7afdc859bbb6bd111a0c9955c60 same run

Verification results observed in the handoff audit

The following non-mutating runs were observed with bytecode disabled:

Module Result
Weighted large-load hostile audit PASS: 117659 checks
Hereditary integer-hull transference PASS: 2171708 checks
Physical owner SDP fiber boundary PASS: 67742 checks
Spectral danger-span boundary PASS: 5510 checks
Heavy-compatible ultralight exit gate PASS: 638708 checks
Full exterior-domination no-go PASS: 1093562 checks
Ferromagnetic light-law fixture PASS total: 285 checks
Canonical convex endpoint-law boundary PASS: 1773 checks
Wall-safe refresh boundary PASS: 170103 checks
Frozen downstream synthesis integrity PASS: 51 frozen files; 86 text guards; 62520 algebra checks in the non-deep run
Frozen downstream synthesis integrity, recorded source deep run 25 source verifiers in addition to the preceding integrity checks
Concise owner/downstream synthesis draft PASS: 68119 checks
Simultaneous heavy/light draft PASS: 12455 checks

Again: these passes validate pinned finite claims, algebra, source hashes, and guardrails. They do not prove the open endpoint law or the universal theorem.


20. External sources to pin in the new filesystem

At minimum, obtain and hash the source of:

The previous weighted audit pinned the following Bansal--Jiang source-file hashes:

Source file SHA-256
Large_k_Beck_Fiala_Conjecture.tex f7004ce9d089f8fb679e4688eaf2fea8467ccd5706ddaf9dda582ebe5ec90075
Weak_Beck-Fiala.tex 3133e9daea1b95dd972179d79d23175e4423e4e230d25efa3f6063dffde1bb5d
overview.tex aa034732d90cf5d3bb9861bdc56118269086a50c704fb5af1ef5629e2bc0de4b
appendix.tex d2ec04dc3fe6b87811de28ffc5fe149a36d8c15e66223dfd8859530070e03bf7
Komlos_Proof.tex c75b2626b5739cada07189e2925add50a67656fe12d7c7ee18c0c4a1e2b83e38

If the current arXiv source differs, do not silently assume equivalence. Record the new version and audit the exact lemmas again.


21. What would constitute a genuine breakthrough

Tier 1: decisive local breakthrough

Any one of the following would materially advance the project:

  • a pinned weighted endpoint sampler with the exact nonneighbor-conditional tail and affordable constants;
  • a proof that the actual weighted endpoint law has a survival-unbiased dangerous-span service at every dyadic live state;
  • a deterministic decomposition theorem that handles arbitrary large mixed components and survives bridge-connected Walsh groups;
  • a sharply scoped no-go theorem for a broad natural class of pinned samplers, together with a replacement theorem that is strictly narrower than the full Komlós endpoint problem.

Tier 2: compiled sub-quarter theorem

A real sub-quarter result must prove, for every Komlós matrix, one of:

[ |Ay|\infty \le C\varepsilon(\log n)^\varepsilon \qquad\text{for a fixed }\varepsilon<1/4, ]

or

[ |Ay|_\infty\le C\log\log n. ]

It must include every upstream block, finite stopping, hereditary restrictions if used, and a benchmark recovery branch.

Tier 3: complete solution

A complete handoff back should contain:

  1. a formally stated universal theorem;
  2. a proof with all quantifiers and constants;
  3. a finite algorithmic version or a rigorous finite approximation if stochastic localization is used;
  4. a source-by-source black-box audit;
  5. a hostile test suite and exact verifiers;
  6. a clear explanation of why no law-selection, support, survivor, owner, or terminal circle remains;
  7. a comparison with the Bansal--Jiang quarter-power proof at the exact budget level.

Things that are not breakthroughs anymore

  • another proof that ultralight (U) has a good signing in the unrestricted cube;
  • another exact-mean convex decomposition with no heavy conditional tails;
  • another entropy scalar or covariance scalar with no physical endpoint receipt;
  • an expectation-only dangerous-row count;
  • a new name for maximum entropy, product I-projection, or the full lopsided LP;
  • an owner of small rank without a complete-residual factorization;
  • a structural-class (O(D)) theorem presented as universal;
  • a verifier pass for a finite fixture presented as a proof.

22. Final research directive

The project has already closed the obvious upstream holes. The ultralight block is affinely (O(1)). The low-heavy local-lemma arithmetic is (O(D)). Exact convex refresh, low-rank/row-(\ell_1) terminals, target-resolution spectral gates, low-variance terminals, small-core terminals, and the quarter-power fallback are available. The obstruction is the logic of one endpoint under conditioning and support changes.

Your first question should therefore be:

What concrete randomized mechanism produces an ultralight-good endpoint while retaining construction-level subgaussianity for a low-heavy row after every conjunction of nonneighbor heavy-good events, and while hard-supporting all earlier heavy receipts?

If you can prove that theorem, compile it immediately. If the actual weighted process refutes it, freeze the counterexample and change the mechanism, not the vocabulary. If every plausible sampler route fails, attack the large mixed-component decomposition and use the failure geometry itself as structure. In all cases, keep the Bansal--Jiang branch executable at the root and never let a new proof hole appear downstream of the one you have just closed.

The universal theorem is still open in this handoff. The opportunity is that the remaining wall is now unusually sharp:

[ \boxed{ \text{construction-level common-endpoint compatibility,} \quad\text{not ultralight discrepancy.}} ]

That is the point at which the next agent should begin.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment