# Review and integration of the research returns **Twin-prime infinitude and all sufficient signed margins remain OPEN.** This review covers the second dispatch from 1285d47 through its closeout 46fa948, including the corrections and second-reader returns. The incoming state is preserved in Git. Owning lane notes contain the current proofs; [OUTCOMES.md](OUTCOMES.md) records reusable results, failed steps and reopening conditions. No novelty claim is made for this integration. ## 1. Current dispositions | Lane | Accepted at stated scope | Remaining limitation | |---|---|---| | [A](centered-discrepancy-estimate.md) | D_y=D^(e_1)+O_(A,eps)(x/log^A x) | D^(e_1) has no sufficient signed bound | | [A2](fixed-endpoint-discrepancy.md) | Repaired low Type I estimate; S=C2*x+B+O_A(x/log^A x) | Exact B, the Type II plus band sum, remains unestimated | | [B](joint-factor-estimate.md) | Family identity R_F=C2*x*H_delta-Z_F^p+o_delta(x), using ordinary BV at all moduli | Separate-sign sieve bound is insufficient; the complement is unpaid | | [C](full-coefficient-average.md) | Exact Fourier lift and norm >=(log x)^(2/5) for 1-bounded representations on all smooth inputs | No matched filtered correlation rate; some nonzero modes satisfy the unit-disc condition | | [D](structured-dispersion-estimate.md) | Lemma H, (D1), additional region and fourth simultaneous residual cut | Target box exponent 407/400>1; no global signed margin | | [E](fold-arithmetic-bridge.md) | Rough-product BV and aggregate constant 4 at the stated source reading; certified all-depth failure of the two displayed tests | Cov_u, Dec_1 and better joint contamination estimates remain OPEN | | [V/W](history/reviews-0907/12-halberstam-richert-second-access.md) | Source-custody distinctions and existing finite measurement corrections | Original H–R theorem page unread; constants and onset unpriced | The analytic statements below were checked from their arguments, not accepted from prior grades or repeated reviewer agreement. Deep external theorems remain named dependencies; this is not a proof audit of every historical note in the repository. ## 2. A: original failed steps and current scope The original endpoint replacement n>=e_1*m by n>e_1*m loses the n=15 atom at x=16,e_1=3,m=5. Odd square divisors must be retained after removing powers of two. The reciprocal-totient local factor is (p-1)/p, with coprimality still present, not its reciprocal. These are real defects in the first argument, not counterexamples to the truncation. Section 3a of the owning note repairs them; section 7 below records the reviewed analytic chain. The undefined old H_band interface is withdrawn. ## 3. C: representation and transfer are different obligations Rounded cutoffs require w_i=log(a_i)/log x and w'_i=log(b_i)/log x in the Fourier profile. The fixed-exponent proxy is a measurement of a different finite function. The exact identity with these moving endpoints has uniformly summable coefficients. Unweighted cancellation does not transfer through an arithmetic envelope: a=(1,-1), w=(2,1) gives sum a=0 but sum wa=1. The composite filter is not multiplicative, since it vanishes at primes p,q>W but not at pq. Its exact expansion reintroduces the prime-pair sum. This invalidates that direct transfer, not every possible joint Fourier estimate. The unit-disc rejection of every nonzero mode was false. At p<=W_i, |G_(i,k,u)(p^v)|=2|sin(pi k log p/(2log x))|<=1 if |k|log W_i/log x<=1/3; at larger primes its modulus is one. Thus at least k=0,±1 on the left and l=0,±1,...,±6 on the right qualify for every real Mellin twist. This does not check their other correlation hypotheses or supply the required filtered o(x/log^2 x) rate. Only an upper bound O(x/log^2 x), not a matching asymptotic, is used for the unfiltered rough-pair count. ## 4. E: finite evidence versus all-depth inequalities The parity-table identity is exact, including its denominator-free form at S=0. The pair (125,127) on (64,128] survives y=4 with both parities odd but is not a twin pair. The safe no-three-factor cutoff on a dyadic interval is (2X+2)^(1/3). The original grid did not prove an all-depth statement. The current Proposition 5 has an elementary analytic bound whose five pieces are certified with directed rational logarithms in the integration validator. Its upper bounds are below 1 for Q_cov and below 4 for c*_real, with positive slack. Its sub-2 conclusion holds on (4,4.8] and (8,infinity). The stronger sub-2 observation throughout the intervening interval is still only measured. These close the two displayed tests with aggregate constant 4, not the decorrelation hypotheses or every improved consumer. ## 5. V/W: source and retained-data custody [Report 10](history/reviews-0907/10-independent-review-0908.md) supplies finite tolerances and the CRT representative repair. The shared M columns in the retained D_y and shifted-prime JSONs agree within their recorded floating-point tolerance. The S/x and D_y comparisons remain finite measurements; no asymptotic conclusion follows from them. Pooled random-control results do not bound each omitted control separately. [Report 11](history/reviews-0907/11-two-class-theorem2c-source-review.md) and [report 12](history/reviews-0907/12-halberstam-richert-second-access.md) separate published statements, same-author secondary page readings and OCR-only evidence for H–R 1974 Theorem 2.2. The substitution matches the reported OCR hypotheses, but that does not certify the original page or Lemma 2.2 behind its Remark. The thirteen access attempts are a retained worker report, not thirteen fetches repeated in this review. The manuscript already marks the source OCR ONLY; retain that limitation. The asserted onset with constants set to one is not an effective threshold. ## 6. Validation and limits [research-round-validation.js](research-round-validation.js) retains the endpoint, local-factor, rounded-profile, weighted-transfer and dyadic parity controls; reads the existing JSONs; and adds exact rational checks of D's exponent target, B's uniform constant, C's low-mode condition, A2's density/consumer logic and E's five-piece inequalities. It also independently recomputes the C norm witnesses. The q=1 Mobius mean values are finite consistency checks, not tests of an asymptotic rate. The six lane validators and this integration validator are bounded checks. Their invocation and output hashes belong to their embedded OUTPUT blocks. No census producer, random-control regeneration or Lean build is run by this review. Required combined gates are strict QC with indexes, QC self-tests, audit-numbers and git diff --check. Their final execution record is in the CHANGELOG. Green mechanical checks do not certify the analytic corpus. ## 7. A: reviewed truncation chain The review reconstructs the prefix version (BV*) by a logarithmic mesh from [Tao Notes 3, Theorem 17](https://terrytao.wordpress.com/2015/01/10/254a-notes-3-the-large-sieve-and-the-bombieri-vinogradov-theorem/), re-read on 2026-09-09. The fixed power margin absorbs all logarithmic rounding and multiplicity costs. The square-divisor and coprimality tails (3a.10)–(3a.11), c(q)<=tau(q)^3, and Cauchy against the tau^6 moment are sufficient. For the main term, the corrected weight has series H_(b,g)(s)/zeta(1+s). Its coefficients eta satisfy sum |eta(k)|sqrt(k)<.92. Taking r proportional to log log x and using primes in [P,2P] with P=x^(.7/r) makes the profile perturbation tend to zero. This supplies actual smooth inputs with |F_L|>=(log x)^(2/5) for all sufficiently large x, as detailed in Proposition 6.5. Any representation by functions bounded by one on every such input has coefficient norm at least that supremum. This is the missing growing witness, and is stronger than the earlier single finite norm example. It does not rule out density-one representations, explicitly paid growing components or an estimate treating the coefficient sum jointly. ## 9. B: ordinary BV repair and constant range The right short approximant has nonsquarefree moduli, so Wu's squarefree-weighted lemma cannot evaluate its entire Type I term. Ordinary BV at each cofactor length Y=x/m>=sqrt x does: moduli are at most W_R^2<=x^.1, below sqrt(Y/2) by a fixed power. Prime-power removal, two prefixes and a mesh for the variable log weight are affordable at arbitrary fixed logarithmic precision. The family has bounded harmonic mass for each fixed delta, so the errors sum to o_delta(x). The Mobius density constant and its local exclusions give the stated family identity. The imported 24/22! Dickman error was specialized to delta=1/100. For all 0x^(7/300+epsilon), with fixed positive epsilon. The small-gcd range remains open. The returned 59/40, 17/200 and 17/300 misstated this target; they are corrected in the owning note and handoff. This does not control the full target box, corner or global residual, or improve the uniform product threshold 19/25. ## 11. E: imported sieve scope and certificate Wu, [arXiv:0705.1652v1](https://arxiv.org/pdf/0705.1652v1), printed page 6, was re-read at the rendered page on 2026-09-09; PDF SHA-256 `41d432dd63da6d1fe7836ba3beda8b601ce6e64420e50501d26d79f7d971043e`. Lemma 2.3 has the weight mu(q)^2*3^nu(q), so only squarefree moduli are controlled. Proposition 3 uses that scope correctly. Its passage from the distinct-prime-weighted convolution to set_k pays repeated-prime terms; three uses of the lemma handle the moving endpoints. The coefficient-uniform reading for bounded f is explicit in the note; the original Pan–Ding proof was not independently read in this review. Lemma 2.2 supplies an order-one, level-Q upper sieve. Applying it to the aggregate set_k-2 gives 4D_k(u) after summing tuples. It does not give constant 4 uniformly for each cofactor tuple. Keep an additive O_(A,k,u)(X/log^A X) in the cardinality form, including k=u where D_k=0. The two ratio tests need only the aggregate statement, so this correction does not reopen those tests. Section 4a's rational certificate establishes their all-depth failure with the standard linear-sieve functions. ## 12. A2: independent reading of the repaired Type I estimate The original bound e[r,g]<=e_0UV is false: at x=2^16,e=g=209,r=1 it gives 43681>1908. The corrected proof truncates g<=G=(log x)^(A+13). The body then lies below x^(1/2-eps'/3)G. On the progression side, (r,g)=sum_(d|(r,g))phi(d) yields the tail x log^12 x/G. On the density side, the same divisor sums with phi denominators give x log^11 x (log log x)^2/G. Both are affordable. The full local density is 1_((r,e)=1)/(e phi(r)), including repeated powers in r. Its local exponent at p|e,p|r is v_p(r)+1. The main term uses A's mean at (b,g)=(k,e), uniformly in the varying excluded primes. The multiplicity bound tau(q)^4 and tau^8/phi sum cost log^256 and are paid by the arbitrary BV precision. The repaired argument survives. Two bookkeeping fixes matter for the written quantifiers: the weighted triple count costs log^2 before the log^2 error weight, giving log^4, and delta=1/100 in the divisor-bound estimate handles the power-of-two atom for every 0=-4x/25+o(x). It implies B>=-(C2-1/200)x+o(x) using the squarefree upper bound for M; the converse does not follow. The exact clipped endpoints keep the ±1 atoms. Statement (4.9) is a stronger sufficient all-modulus input for the band, not a necessary condition. For an elementary size comparison B=O(x log^5 x) suffices; the reported O(x log^4 x) was not justified by Brun–Titchmarsh on every divisor class. No sufficient signed bound for B has been supplied. ## 13. Integrated research decision A/A2 reduce an exact representation; B evaluates a linear family part; C limits a particular representation strategy; D adds a controlled region; E rules out two explicit sufficient tests. These are different statements with different gaps. Calling all of them the same bilinear theorem would lose the coefficients and quantifiers needed for a useful next attempt. The most concrete next local obligation is D's remaining small-gcd moment saving greater than 7/200 at the target sector, with every separation cost included. This is a research-priority judgment, not a claim that it is necessary for twins or likely to succeed. It requires a source/structure check before new analysis or compute, and success at that box would still leave a global complement. The independent alternative is a one-sided bound for the actual B of A2 with its fixed parameters. B/C/E can be reopened by a matched rate, different consumer or a precise correctness concern; their scoped failures are not universal exclusions. No new agents, scheduled jobs, long runs or publication actions are started by this integration. Current priorities and ownership live in [RESEARCH-EXECUTION.md](RESEARCH-EXECUTION.md) and [../TODO.md](../TODO.md).