Afterword: FPRD Lab’s contribution

Parameter retuning in the direct reduction for the Euclidean closest vector problem: 1/30-hardness and a 1/28 retained-parameter frontier

Methodological and contribution addendum. The paper’s mathematical scope and review status remain as stated in the paper.

The Lab’s approach

FPRD means Finitely Presented Recognition Dynamics. Its starting point is to distinguish a finite description, the possibly infinite process it generates, and what an observer can learn from that process. Equal answers need not come from equal internal structures. Conversely, an infinite state space may have a short, exact description and useful coordinates; failure to compress it into finitely many states need not end the investigation.

A small example: balanced parentheses

The grammar S → (S)S | ε presents the language of balanced parentheses. To recognize a prefix, keep the number of unmatched opening parentheses, or a failure state if a closing parenthesis arrived too early. An opening parenthesis increments the count. A closing parenthesis decrements a positive count; at zero it enters failure. Failure persists. At the end, accept exactly when the count is zero.

The prefixes ( and ()( leave the same future task: any suffix that completes one completes the other. The prefix (( leaves a different task: ) completes the first two prefixes but not this one. These future tasks are residuals. There are infinitely many counts, yet a finite set of update rules generates their dynamics. This is a classical automata example, not a discovery of FPRD Lab. It illustrates the research question: which coordinates preserve all relevant futures, and what does updating or reading those coordinates cost?

From viewpoint to proof

In practice, the Lab fixes the observable and the allowed operations, compares presentations of the same problem, and seeks an exact translation or invariant. A failed shortcut is useful when a counterexample identifies the missing coordinate or hypothesis. Small computations test particular uncertainties; a proof must explain why the result holds beyond those tests. Costs, relation sides, boundary conditions, and the direction of each translation remain explicit.

AI-assisted exploration, source comparison, executable checks, and adversarial reconstruction support this work. Their agreement is not automatically independent evidence: shared assumptions must be exposed. A formal development certifies its stated formal theorem, not every interpretation attached to it. Each paper retains its own evidence and review status.

The approach is most useful when a poor representation hides a reusable construction, a compatibility condition, or a sharp obstruction. Outside recognition theory, it can guide the search for a structural reformulation without producing a literal residual dynamical system. The account below distinguishes that methodological influence from the ingredients of the proof.

FPRD’s contribution to this paper

Here the useful change of presentation was numerical. The source reduction fixed several large parameters. The Lab treated them instead as coupled resource coordinates: field size, support cutoff, and moment horizon. It then asked which inequalities the retained proof actually needed, rather than changing a displayed exponent in isolation.

Eliminating those internal coordinates exposed a frontier at 1/28 for the retained strict inequalities. The paper gives a concrete 1/30 specialization and checks the finite constants as well as asymptotic exponents. These numbers describe different claims: feasibility below the retained frontier, an explicit specialization, and a barrier for this particular system of obligations. The frontier is not an impossibility theorem for other reductions.

The representation audit mattered at several interfaces. Binary Hamming weight becomes squared Euclidean distance, so the approximation exponents differ by a factor of two. Support must be discarded separately for each table type, rather than by taking an unjustified common intersection. The coefficient needed in binary soundness must survive into the Euclidean wrapper. A discrepancy between one imported formal wrapper and the paper’s coefficient requirement was addressed explicitly instead of assuming a literal substitution preserved the proof.

Inherited ingredients and the boundary of the contribution

OpenAI’s direct reduction supplies the one-hot encoding, reconstruction from polynomial moments, valuation and clause-matching arguments, and parity lift. The Lab did not replace those constructions. Its contribution is the resource parameterization, elimination, finite-constant bookkeeping, and corresponding retuning of the verification interfaces.

This is an application of the Lab’s habit of separating representation costs and proof obligations. It is not a construction of a new language-residual system or an independent derivation of every imported lemma. The general written retuning and the archived executable Lean specialization also have different scopes: the latter establishes the concrete 1/30 reduction, while a separate formal result treats feasibility of the parameter inequalities. The full reduction is not uniformly formalized for every exponent below 1/28.

Where to inspect the contribution: the paper, Sections 1.2–1.3, 2, 5.5, and 7. These distinguish retained source mathematics, the repaired interfaces, and the exact formalization boundary.