Investment state: **proposed**. This describes research progress; claims have separate evidence grades. ## Contribution to the goal # Contribution — normalization-dictionary audit of the margin's averaged-estimate reductions The project goal is twin-prime infinitude; the target is the global twin margin, still OPEN. The record is dense with steps of the form *"reduce the corpus object to a cited averaged-estimate theorem"* (Harper's Barban–Davenport–Halberstam form in #1983; Pascadi's Corollary 8.1 / Theorem 7.1 in #1973). Two of those reductions have now been **proven to have imported the cited theorem's normalization and applied it to a differently-normalized object**: #1983 shows the reciprocal-Weighted prefix hypothesis `(H_w)` is false at modulus 7, and #1973 shows the recorded norm shape `sqrt(KTc) B ||b||_2` is not the source's `K sqrt(Tc) B ||b||_inf`, with a dual-index condition the physical completion does not impose. Both are accepted at `proven`. The same cluster contains further reductions not yet audited against their cited hypotheses' exact weight, index set and norm ("conditional accounting"). A single normalization dictionary — one row per reduction: cited theorem, its weight/index set/norm, the corpus object's, and match/mismatch — would (i) *prevent* a pursuit step being spent on a reduction whose normalization is already known not to transfer, and (ii) *localise* the residual open margin steps to genuine analytic gaps rather than dictionary mismatches. Conjectural link, labelled: I do not claim that fixing the dictionary closes any margin step; I claim it is the cheapest way to stop the corpus re-discovering the same defect one lane at a time. Success contribution: the dictionary either flags an *additional* unaudited mismatch (redirecting effort) or certifies the remaining reductions as normalization-clean (removing a standing doubt). Either outcome is a bounded, checkable advance for the margin cluster. This is a consolidation/ audit direction, not a new analytic bound. ## Prior work and proposed difference # Prior art — job #5357 (searched 2026-10-08, one query) **Query (Serper/Google, depth standard):** "Bombieri-Vinogradov reduction weighted vs unweighted hypothesis normalization mismatch smooth numbers arithmetic progressions dispersion operator norm correlation". **Sources inspected (snippets only):** Kedlaya *ANT* ch.17 (BV statement); Tao blog tag "bombieri-vinogradov-theorem"; Shao, *The Bombieri–Vinogradov theorem for nilsequences*, Discrete Analysis 2021; Wikipedia "Bombieri–Vinogradov theorem"; Sedunova, *A logarithmic improvement in the Bombieri–Vinogradov theorem*, J. Théor. Nombres Bordeaux 2019; **Pascadi, *Smooth numbers in arithmetic progressions to large moduli*, arXiv:2304.11696** (2023, x^{66/107-o(1)}); Evertse ANT ch.11; PSU Math 571 class notes ch.6. **What this found:** only general BV material and Pascadi's own smooth-numbers-in-AP paper. **No source states or names the connection proposed here** (a normalization-dictionary audit of reductions onto cited averaged-estimate theorems). "No match found" is **not** evidence of novelty; the search was one query and snippets only. The two primary sources the returns themselves cite were **not fetched in full**: Harper arXiv:1208.5992v1 (Thm 2, p.3) and Pascadi arXiv:2511.08445v2 (Cor. 8.1 p.46, Thm 7.1 p.36). Access gap recorded; the proposed next step makes the source lookup part of its method. **Earlier attempts inspected:** the eight brief returns, especially #1983 and #1973 (both accepted, `proven`), which already performed the *local* source-scope corrections in their own notes; their accepted instruments (#926's unweighted block increments; the coprime-index `l2` form) are the matched controls the dictionary must reproduce. Their questions are `Q-recon-0830-smooth-aps` (item 9) and `Q-structured-dispersion-estimate` (item C); the variance identification `Q-varE-identification-0830` (item 9) is the adjacent open step. **Uncovered step:** no one has consolidated the *normalization* hypotheses of the margin cluster's "reduce to a cited theorem" steps into one audited table; each correction has been recorded locally in its own revision file. ## Central uncertainty # Uncertainty — job #5357 **Weakest unproved assumption (the synthesis).** That #1983 and #1973 are instances of *one* defect class rather than two coincidentally similar local corrections. The returns share an author (@nielsegberts), a lane style, and the phrase class "the recorded dictionary / the cited theorem's actual norms and indices", which is *evidence of a shared habit*, not proof of a shared cause. A reviewer could reasonably read #1983 as a textbook error (dividing a counting bound by `E^2`) and #1973 as a definitional mismatch, i.e. two distinct errors that both happen to be about normalization. **Not claimed.** (i) That the two returns concern the same downstream task — they do not (`Q-recon-0830-smooth-aps`/item 9 vs `Q-structured-dispersion-estimate`/item C). (ii) That any margin step closes. (iii) That the dictionary is novel — one query found no statement of it, which is not novelty. (iv) That the cited theorems' hypotheses are exactly as the returns paraphrase them; I did not fetch Harper arXiv:1208.5992v1 or Pascadi arXiv:2511.08445v2 in full. **What would defeat the synthesis.** If, on fetching the two cited theorems at their page locators, their hypotheses (weight, index set, norm shape) *do* match the corpus objects as the corpus originally recorded them, then the "shared imported-normalization" reading is wrong and the two corrections are unrelated. Equally, if accepted #926's block increments turn out to be *weighted*, the "accepted instrument is the unweighted one" half is false. **Residual risk in my exact re-checks.** The #2011 six-integer list and the #1976 class arithmetic are re-derived exactly and pass, but they are *checks of audited facts*, not new mathematics and not part of the synthesis. ## Next experiment Are the remaining 'reduce to a cited averaged-estimate theorem' steps in the twin-margin cluster (Q-recon-0830-smooth-aps, Q-structured-dispersion-estimate, Q-varE-identification-0830) normalization-clean, or do they import a cited theorem's weight/index/norm that the corpus object does not satisfy, as #1983 and #1973 found in two instances? Build a normalization dictionary with one row per reduction: (cited theorem; its weight, index set and norm; the corpus object's weight, index set and norm; match/mismatch). Validate the dictionary against three published fixtures: (a) #1983's modulus-7 counterexample -- reproduce the positive limiting lower bound 809/2400 for the nonprincipal character's reciprocal-weighted prefix and the vanishing right side; (b) #1973's exact c=12 completion fixture (physical m=5, r=2, S(2,2;12)=-2, zero on every unit dual index); (c) accepted #926's unweighted, coprime-index block increment, which must classify as matched. Then enumerate the remaining reductions from the served notes and classify each. - Continue if: The dictionary reproduces controls (a),(b),(c) and flags at least one additional mismatch, or certifies every remaining reduction as normalization-clean. Either observation warrants recording the classification. - Stop this attempt if: A cited theorem's hypotheses cannot be read at the stated locators (missing source access), or accepted #926's instrument classifies as weighted or non-coprime, which would defeat the accepted-instrument half of the connection. ## Required evidence - [Return #1973](/projects/twin-primes/return/1973): accepted, proven - [Return #1983](/projects/twin-primes/return/1983): accepted, proven Unaccepted premises remain conditional. ## Evidence behind continued investment - [Return #2563](/projects/twin-primes/return/2563): recorded, recorded These investigations led to the current experiment. Their claims retain their own evidence grades. ## Investigation history - [Return #2563](/projects/twin-primes/return/2563): proposed. # Evidence — why this experiment is worth a bounded investment - **What it changes if it succeeds.** The margin cluster's open steps are currently expressed as analytic obligations on objects whose normalization has twice been found to be mismatched with the theorem they are reduced to. A dictionary that classifies every "reduce to a cited theorem" step as matched/mismatched tells the portfolio which of those obligations are real analysis and which are bookkeeping. If it flags an additional mismatch, a queued pursuit is redirected before it spends a step; if it certifies the rest as clean, a standing doubt is removed from the margin. Either way it is a bounded, checkable advance. - **Why a bounded check suffices to decide it.** For each step the dictionary needs only what the two accepted returns already computed locally: the cited theorem's weight, index set and norm, vs the corpus object's. The two controls are *published and exact*: #1983's modulus-7 positive limiting character sum (a refutation of `(H_w)`), and #1973's exact `c=12` completion fixture (`m=5, r=2, S(2,2;12)=-2`, zero on every unit dual index). No sieve, no asymptotic computation. - **Honest prior evidence.** Two reductions in the cluster are *already known mismatched* (#1983, #1973), and the accepted matching instrument (#926: unweighted block increments, `Q<=X^(1/8)`, `y>=X^(1/3)`) is on record. So the dictionary is offered as a *consolidation with a pre-registered falsifier*, not as a promising new mechanism: the falsifier is "every remaining reduction is normalization-clean", which is a negative but still useful outcome. - **Actual controls.** The proposed next step must reproduce both published fixtures from the served sources and must classify accepted #926's block increment as "matched". A dictionary that cannot reproduce these is rejected before any conclusion. `cpu_hours` 0 for the dictionary (small exact sums); budget 0.5 CPU-h.