# Primeoire review and proposed next campaign, 2026-09-05 Internal review at commit `6ab754e`. This is a proposal for Chris, not a change to the adopted research programme. The existing manuscripts and research summaries have been left unchanged. No external publication or contact was made. The project has a useful exact representation of the problem and substantial computational work. It has no demonstrated route to infinitely many twin primes. However, the latest assertion that the tile itself cannot support such a route is stronger than the evidence warrants. The correct distinction is between what the structure encodes and what the estimates currently extract. ## 1. Review scope and what survives Read the entry documents, target, lens, attack and question registries, search conventions, the upper-bound derivation, the relevant variance and anchored formulas, and the records underlying the parity and bilinear closures. The lower-bound manuscript was inspected for its statement and construction setup; its full Kalmynin–Konyagin adaptation has **not** been independently audited in this review. Neither the entire historical corpus nor every external citation has been referee-checked. **VERIFIED computationally.** `node research/qc.js --full` exited successfully: 14 corpus checks clean, the checker self-tests passed, and 251/251 numerical checks passed. This included recomputing the maximum twin-slot gaps through 37#, and checking the recorded gap witnesses at 41# and 43#; the latter two were not fresh exhaustive maximality searches. A successful gate verifies its listed assertions, not every mathematical sentence in the repository. The added [validation script](../../review-0905-validation.js), whose output is written by the existing embed tool, independently compares the browser fold kernel with direct divisibility marking through T19. It also exhausts the CRT covering identity at T7 and checks the four-sign identity in section 4 below. These are finite checks of definitions and algebra, not evidence for a new asymptotic law. **DERIVATION CHECKED, using the cited sieve theorem.** The application in [beta2-note](../../../paper/beta2-note.md) has the right structure. With `H = z^(beta+epsilon)` and sieve level `D = z^(beta+epsilon/2)`, the weighted remainder is `O(D log^7 D)` while the positive main term has order `H/log^2 z`. Their ratio tends to zero. This gives the stated uniform upper bound with `beta = 4.266450284…`. I found no defect in this application; this review did not re-read all the archived book pages or establish its novelty. **ELEMENTARY AND CHECKED.** The fold census, crystallization with the exclusion of 1, and the CRT product for pair correlations are sound. Almost-all phase estimates do not select the integer origin. A subquadratic uniform G2 bound would suffice for twin primes. None of these statements is a new proof of infinitude. ## 2. Corrections needed before another campaign ### A. The parity obstruction is overextended in the latest summary [README Status](../../../README.md) says every statement about the tile is a residue statement and concludes that every argument using it is parity-blind. That conclusion does not follow. Exact divisibility information, including its arrangement on a finite interval, determines primality there. Full inclusion–exclusion is an exact identity. The difficulty is controlling its signed remainder uniformly as the interval grows. The classical obstruction concerns specified classes of sieve bounds and their available errors. A parity-twisted comparison sequence can preserve their usable asymptotic statistics; it need not preserve every exact residue count or every possible fact about the original integers. The stronger general obstruction in [Tao's 2014 discussion](https://terrytao.wordpress.com/2014/11/21/a-general-parity-problem-obstruction/) also explicitly relies on a pseudorandomness hypothesis. It cannot serve as an unconditional impossibility theorem for all arguments on the tile. This distinction was already made in [the obstruction audit](attack-obstruction-audit.md) and [the exact-data adversary experiment](attack-parity-adversary.md). The latter even records finite instances where exact congruence information defeats its adversary. Those computations are not an asymptotic escape, but they directly contradict the unrestricted information-loss claim. **Repair:** every exclusion must name the method class, retained statistics, error tolerance, and relevant quantifiers. Keep the parity checklist, but do not classify a proposed proof by the presence of residue notation alone. This does not reopen any specific failed inequality. ### B. The number 4 is not a proved universal floor The README states that everything in `(2,4]` needs input outside the standard sieve axioms. Its own [refuted registry](../../REFUTED.md), under the floor-at-4 entry, withdraws precisely that claim. The DHR value is an achieved sifting threshold, not the optimal threshold proved for every possible sieve. [Ford's notes, section 3.1](https://ford126.web.illinois.edu/sieve2023.pdf) table known **upper bounds** on the optimal sifting limit and state that its exact value at dimension 2 is unknown. **Repair:** say that the available DHR lower function is zero at the target parameters, and that an improvement of its threshold is open. The failure of current methods is real; a universal barrier at 4 has not been established by the cited material. This correction supplies no improved sieve. ### C. The Jacobsthal implication loses a quantifier [The new consolidated paper, section 5](../../../paper/two-class-jacobsthal.md) and the README infer `g(x#)=o(x^2)` from `G2(x#)0` would imply the little-o conclusion. An infinitely-often upper bound is weaker again. These targets must not be interchanged. **Repair:** preserve the distinction between a constant at exponent 2, a strict power saving, and an infinitely-often result. This is a correction to the cited deduction, not a proof that no other implication can exist. ### D. CRT does not prevent adverse alignments from combining [THE-LENS section 3](../../THE-LENS.md) says different primes' worst alignments cannot compound at one location. For any specified residues `a_p`, CRT supplies an `s` with `-s = a_p (mod p)` simultaneously for every prime in the tile. In a window translated by `s`, its forbidden pairs are exactly `{a_p,a_p-2}`. Thus the full tile realizes every such joint choice. The correct statement is already proved in [two-class-lower-bounds section 1](../../two-class-lower-bounds.md): G2 minus one is the maximum covered interval under freely translated pairs with separation 2. The validation script exhausts all phase choices at T7. The larger paired Jacobsthal function also varies the common even offset. Its larger values therefore do not establish protection from adverse translations in G2. Any protection near the actual origin must be proved using the origin and the interval length. ### E. The headline numerical band drops an outlier The README's full-ladder range `0.45 to 0.53` omits the value at x=37. [The bound producer](../../a144311-full-ladder.js) and the paper's own table give approximately 0.594 there. For the reported normalization over x>=11, the full range is approximately 0.446 to 0.594. Excluding a point must be explicit; a local power fit is not an asymptotic exponent. ### F. One shifted Liouville average is not a complete proof specification The README names `sum_(p<=X) lambda(p-2)=o(pi(X))` as the clean missing input. Such cancellation would concern the parity of the number of factors of p-2. Negative Liouville value alone does not imply primality: for example, `107-2=105=3*5*7`. A proof must also supply appropriate roughness or factor-count restrictions and show that the estimate survives those restrictions. **Repair:** call this an example of difficult parity-sensitive information, not a sufficient substitute for the conjecture without a written implication. Likewise, Type II is a concrete class of useful inputs, not a proved necessary form for every conceivable future proof. ## 3. What the finish line actually requires Use the ordinary count `T(X) = #{n <= X : n and n+2 are prime}`. The target is exactly `T(X) -> infinity`. Equivalently, the dyadic intervals `(2^j,2^(j+1)]` contain a twin opener for infinitely many j. This asks for neither every zone nor every translation of an enormous primorial period, nor a Hardy–Littlewood asymptotic. Sparse successful scales are sufficient. The existing weak Zone Postulate already records this quantifier distinction. Changing the target alone proves nothing, but it avoids importing an unnecessary uniform maximum-gap problem. A sufficient lemma being as hard as the final problem is not itself a refutation: the useful question is whether it comes with independently justified input or merely renames the desired conclusion. ## 4. A useful correction to the Liouville experiment's interpretation [The earlier lambda ledger](attack-lambda-ledger.md) correctly proves that signs become constant on survivors after full crystallization. Its off-depth test, however, looked for unusually large pair covariance and a signal stronger at shift 2 than at shifts 4, 6, and 8. Failure to find that anomaly does not exclude the usefulness of proving decorrelation. All those fixed admissible shifts have related prime-pair conjectures; a useful estimate need not be unique to shift 2. Here is an exact way to see which sign matters. Fix X sufficiently large, `y = ceil((2X+2)^(1/3))`, and the openers `X0, put `a=A/S`, `b=B/S`, and `cov=C/S-ab`. The prime-pair proportion is `[(1-a)(1-b)+cov]/4`. Consequently, sufficiently small negative covariance, together with controlled marginals below 1, would help prove a positive count. A large exceptional covariance is not required. The new script checks the identity on finite intervals only; this is elementary bookkeeping, with no novelty claim. **The unresolved obligations are all retained:** at this depth a useful lower bound on S, bounds on a and b conditioned on *both* numbers surviving, and the necessary covariance bound have not been established here. One-variable rough-number marginals cannot simply replace those conditional marginals. This identity is a diagnostic for evaluating a proposed estimate, not an announced route through the wall. Do not repeat the old correlation sweep under a new name. ## 5. Proposed campaign, in order **First: repair the specification.** Apply the corrections above to the live summaries and their downstream decisions, with the normal history record. Replace blanket closures with their actual theorem hypotheses or measured failure criteria. Preserve the many valid finite refutations. This is the immediate prerequisite for sensible selection of the next mathematical task. **Second: make the framework reproduce Chen on long intervals.** The benchmark is a proved lower bound for primes p for which p+2 has at most two prime factors, on the usual long counting range. Chen's theorem is stated and developed in [Tao's lecture](https://terrytao.wordpress.com/2015/01/29/254a-supplement-5-the-linear-sieve-and-chens-theorem-optional/). The repository already proposes this in [bv-import-survey section 4](../../bv-import-survey.md); it is not a new idea or a new theorem. What is missing is a completed derivation in the framework. Retain the fold engine, but track the partner's prime-factor states and import the actual lower-sieve and prime-distribution estimates. Use a whole long interval, rather than demand that the theorem work separately between consecutive prime squares. The latter was the extra obstruction in [the bilinear transplant note](attack-bilinear-transplant.md), whose local failure remains valid within its stated scope. **Acceptance criterion:** a complete classical proof with every imported hypothesis matched to the chosen sequence, and a small independent numerical check of the identities. Relabelling the theorem as an anchored bias is not completion. If the framework cannot carry this benchmark without loss, resolve that before trying to surpass it. **Third: specify one arithmetic estimate beyond that benchmark.** Choose the actual non-negative sequence, comparison sequence, factor ranges, and coefficients before selecting a theorem. Extract the signed sums over products of two varying factors, with the fixed shift 2 still present. Write a conditional implication from a stated error bound to an unbounded twin count. Record the required error relative to the positive main term and whether the estimate is needed at every scale or only infinitely often. [Ford–Maynard's prime-producing sieve framework](https://arxiv.org/html/2407.14368v1) is a suitable reference for this specification: it relates specified Type I and Type II information to prime lower bounds and gives counterexamples in parts of its parameter space. Its introduction was checked in this review; the full paper was not audited. It is already cited in the transplant note. It supplies a framework for pricing an estimate, not the missing estimate for shifted primes. **Acceptance criterion:** one precise unproved inequality, a checked theorem that would consume it, and a proposed mechanism for proving it. If the input is only a divisor-count bound already covered by a closed route, stop that attempt. If the proposed estimate is just the twin count in disguise, label the equivalence and do not count the rewriting as progress. **Fourth: compute only to test that mechanism.** Use scales and controls fixed before the run, include a parity-twisted comparison where applicable, and preserve factor multiplicities. A success must support the error estimate the proof actually needs. More G2 terms, prettier folds, a smaller fitted slope, or a high-significance anomaly without a proof implication would not meet this criterion. The variance-limit project can continue separately as a worthwhile theorem project, with no claim that an ensemble limit forces origin occupancy. The recommendation is to keep Primeoire as the exact language and laboratory, while changing the next deliverable from another gap-growth conjecture to a classical benchmark followed by one precisely priced arithmetic estimate. There is no established estimate in this review that finishes the conjecture. ## 6. Reproduction and limits Run from the repository root: ```sh node research/qc.js --full node research/qc/embed.js --check research/review-0905-validation.js ``` The full gate was run before these review artifacts were added. Afterwards, the strict fast gate passed with zero findings, and the validator's code, embedded output, and kernel-input hashes all passed `embed.js --check`. The new validator was executed through the embed tool, so its output and kernel input are hash-bound. Its independent mathematical content is the direct marking comparison, the exhaustive CRT check, and the sign identity; none is a test of the proposed future analytical estimates. No existing manuscript has been silently promoted, withdrawn, or corrected by this review.