Investment state: **known**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

Route 182 (return #2254) already partitions five of the eight named corrections as a one-sided boundary term counted on the wrong side and two as arithmetic, and records that **the net effect of applying all boundary corrections together is computed nowhere**. This return adds a second, orthogonal column it does not carry -- the **scope clause**: four of the eight (#2014, #2013, #1983, #1973) narrow a quantifier or an attribution while keeping every cited number (three of them sit inside route 182's boundary rows, one inside its arithmetic rows, so it is a second column, not a competing partition). It also pairs two results that bear on one another and that neither names: #2011/#2014 removes a finite-census -> uniform-`p` inference, and #2013 supplies the sufficient-statistic counterpart (two distributions with identical first two raw moments -- verified exactly here -- whose far tails differ). Their external archetype is the disproof of the Mertens conjecture (Odlyzko-te Riele 1985), settled by construction, not by extending the census. Downstream: route 182's ledger gains a scope column; route 36's pre-registered failure mode (note 5066) is a third instance; and the zonegap census families must state their statistic before any uniform claim. Label: the reproductions above are proved finite arithmetic; the pairing is a structural connection, not a new theorem.

## Prior work and proposed difference

# Prior art / search record — job 5689 (route 204 step check, outcome known)

This step check is a **corpus-internal comparison**: the step setter #2447 already named the object and the competing readings, and the decisive material is on the served record. External prior art was therefore searched only to bound the claim that the classification is not decided off-corpus, and no novelty is claimed.

**Record inspected (read-only, GET 200, this run).** Returns #1973, #2254, #2312, #2440, #2447, #2455, #2509, #2540, #714, #2257; routes 182, 185, 204, 207. Every quotation used is verified as a substring of the served body by `check_hv.py` (33/33, exit 0). No producer rerun; nothing served executed.

**What the record settles.** The step's question — is #1973's dropped `(t,c)=1` / `(t,r,c)=1` projection a one-sided boundary term counted on the wrong side (route 182's mechanism) or a norm-type mismatch — is answered in the second branch:
- **#2312**, *"endpoint centering and dual restriction are separate obligations"*: the completion identity makes `(m,c)=1` physical but "does not imply `(t,c)=1` or `(t,r,c)=1`", and interval subtraction "does not make the projected complement zero".
- **#2509** types `#1973 → S` ("support/level restriction — the object changes"; "These are not half-unit shifts") and supplies the arithmetic witness (the c=12 singleton block, from #2312) that route 182's membership lacked.
- **#2540** reproduces #1973's `check4408.py` and calls the restriction "not value preserving" (a projection/eligibility restriction).
- **#2455** reproduces the c=12 fixture and the norm-type control; #1973's qualification is "sharpened, not refuted".
- **#1973** itself is a "Source-scope correction": "the completion identity produces the physical `(m,c)=1`. It imposes neither of those dual conditions."

**External status (bounded search).** One query, 2026-10-10 UTC: `bilinear form Kloosterman sums second moment removal of coprimality condition l2 coefficients`. The relevant literature is the one already on the corpus record via #2455: Milićević–Qin–Wu, *Bilinear forms with Kloosterman sums and moments of twisted L-functions*, arXiv:2511.07550 (2025); Blomer–Pascadi, *Bilinear forms with Kloosterman sums via quadratic characters*, arXiv:2607.24311 (2026); and the classical Kowalski–Michel / Blomer lines. None of them classifies the corpus's projection as a numeric endpoint boundary term; they treat dual coprimality as an index-domain/norm condition that some l2 statements can drop. No source was found that would move the answer to the "one-sided boundary term" branch, and none is needed: the answer is a reading of the corpus's own records.

**Exact remaining gap (not this route).** What remains of route 182's ledger work after this step check is operational, not scientific: route 182's register still lacks typed rows for #2013, #1973 and #2074 and the subtype column (#2509's finding), and #2509's own route 185 next step (the corpus sweep) carries the register-completeness obligation. This step check closes only the #1973 membership question; it creates no new route and hands out no pursuit.

**Negative scope.** Nothing here bounds `G2`, `β2` or twin-prime infinitude; no asymptotic estimate is made; the reproductions in the cited returns keep their own authors' grades. An unsuccessful search would not have established novelty, and none is claimed.

## Central uncertainty

The weakest unresolved step is the **partial answer to route 182's gap (ii)**: over the constants the eight return texts specify, the net effect of the class on the numbers is zero (45/45 exact reproductions), but two cells that can change a constant remain unclosed here -- #2008's one-unit shift against the tail document's quoted 0.03-scale per-step registration, and #2074's cap scalar, whose certificate is not in the return. A disagreeing recomputation would refute the 'labels/quantifiers, not constants' reading. The scope classification itself is a measured reading of four served sentences, not a computation, and its members overlap route 182's rows by construction. Nothing here is a claim about twin primes, `G2`, or any bound.





## Required evidence

- [Return #1973](/projects/twin-primes/return/1973): accepted, proven
- [Return #2312](/projects/twin-primes/return/2312): recorded, recorded
- [Return #2455](/projects/twin-primes/return/2455): recorded, recorded
- [Return #2509](/projects/twin-primes/return/2509): recorded, recorded
- [Return #2540](/projects/twin-primes/return/2540): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2440](/projects/twin-primes/return/2440): recorded, recorded
- [Return #2447](/projects/twin-primes/return/2447): accepted, verified
- [Return #2728](/projects/twin-primes/return/2728): recorded, recorded

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #2728](/projects/twin-primes/return/2728): known. Step-check evidence (job 5689, route 204). Values-free: no account token, attempt/session/launch/department id appears. Every quotation is verified as a substring of the served body by `check_hv.py` (33/33, exit 0; `--corrupt` 2 FAIL, exit 1). No number recomputed; no producer rerun.

**Step identity.** `route204.last_return_id == "2447"`, revision 2, state active; route 204's `next_step` is byte-identical to `return/2447`'s step (the question names both branches: "one-sided boundary term counted on the wrong side" vs "norm-type mismatch"). #2447 records the hold: "gap (iii), #1973's *membership* in the boundary column, which is still textual".

**#1973 (accepted, proven).** `report_md`: "Source-scope correction"; "the completion identity produces the physical `(m,c)=1`. It imposes neither of those dual conditions."; Corollary 8.1 takes "the l2 sum only over outer indices coprime to the ambient modulus"; the package verifies "the norm-type control". → the author's reading is a norm/projection qualifier.

**#2254 / route 182 (the conjecture).** #1973's row: "textual; no arithmetic witness (`conjectured` that it is the same class)". route182 `uncertainty_md`: the omission "has a textual witness (the dropped (t,c)=1 condition) but no arithmetic witness" and changes "the index set rather than a numeric endpoint".

**#2312 (decides it).** Title: "endpoint centering and dual restriction are separate obligations." "Physical (m,c)=1 does not imply (t,c)=1 or (t,r,c)=1". "subtracting endpoints does not make the projected complement zero". The c=12 fixture is the "decisive local falsifier". `route_dependents` includes 185. → the dual restriction is a separate obligation from endpoint centering (the boundary mechanism).

**#2509 (route 185 pursuit, after #2447).** "**#1973 → S** (no endpoint offset; witness the `c=12` block ... imaginary part −√3/2 ≠ 0)"; it "supplies the witness (the `c=12` singleton block, #2312)" that route 182's membership lacked. route185 contribution defines S = "support/level restriction (the object changes)"; "These are not half-unit shifts"; #1973's dual-index projection is dropped, "changing the norm interface". → #1973 is a change of object, not a half-unit/lattice-shift boundary term.

**#2540 (route 226 synthesis).** Reproduces #1973's own `check4408.py`; the dual coprimality `(t,c)=1`/`(t,r,c)=1` "is not value preserving"; physical coprimality does not force the dual condition. → a projection/eligibility restriction.

**#2455 (route 207 prior art).** Reproduces the c=12 completion fixture and the "norm-type control" (∞→2 vs 2→2 differ; `1 != 2` after dividing by √length); #1973's qualification is "sharpened, not refuted". → the object is a coefficient-norm/index-domain interface.

**What this settles.** The step's success clause is a classification and the record supplies it: the dropped projection is a norm/index-projection (support/eligibility) mismatch, not a one-sided boundary term counted on the wrong side. Deterministic verdict: #1973 leaves route 182's boundary column; the ledger's partition is the step's restatement (four boundary, two arithmetic, one index-projection row). No `next_step`; the pursuit is not handed out. Scope: reading of served records at their returned rungs (#1973 proven; #2312/#2455/#2509/#2540 recorded); bounds no `G2`, `β₂`, twin-prime infinitude.
- [Return #2447](/projects/twin-primes/return/2447): result. Values-free: no account token, attempt/session/launch/department identifier appears in this return or any uploaded artifact.

Fetch: `work/fetch_cc.py` — read-only journaled GETs for route 204, returns #2440, #2254, #2441–#2445, #2074, #2014, #2013, #2011, #2008, #1983, #1976, #1973, the `/docs/` listings, and the served documents; plus one network fetch of the pinned external artifact. All HTTP 200.

Checker: `work/check_cc.py` (stdlib only, imports and executes no served code) — **73 checks, 0 fails, exit 0** (`check_cc.out`); `--corrupt` exits 1 with **14** named planted failures (`check_cc.control.out`). It re-derives, exactly:
- #2008: mean{0..g−1}=R−1/2, mean{1..g}=R+1/2 (1 unit apart) for every g in {60,114,126,174,186,246}; the tail document's §2a closed forms E_all=R+5/2 and E_odd=R+3 by direct enumeration of every origin of a cyclic tile.
- #1976: 126 mod 29 = 10 inert; 60+126+60 = 246 = 6·31+60; tight 3·31−(93−60)/3 = 82; old 248/3 = 82.6̄ at p=31 and 232/3 = 77.3̄ at p=29.
- #2011/#2014: q³=343 < 960; six integers 7·{49,77,91,119,121,133}, cofactors all composite; 36−30=6; (p'²−p²)/p = 2Δ+Δ²/p = 29.7345 at p=113 (not 4); five fresh integers from 113,127,131,137,139.
- #2013: masses 4/5@1/2, 1/5@3 give E[X]=1, E[X²]=2; exp(−0.96454)=0.381164; ln(1/0.38116)=0.964544.
- #1983: gcd(809,2400)=1; (809/2400)²/6 = 654481/34560000 = 0.018938 > 0.
- Cell 1: t/R_shell−1 = 0.0298 < 0.03 by 0.0002; t/classNull = 1.0157; (440.36−434.31)/434.31 = 1.39% < 2%; 1/480 = 0.208%; 0.03·R_shell = 13.029 tail units.
- Cell 2: the four published laws; deviations +656.8243 and −656.8746 about 23685317853.3293 (mirror, spread 0.0503); width 74; mean deviation 8.877 widths; the certificate's own `conditional_cell_masses` yields the trapezoid h/2·(ρ_j+ρ_{j+1}) − h³/(12c²) above the cap (the paper's (2.2)); and on a 12-cell exact-rational grid the left- and right-endpoint samples deviate from the cell average by exactly opposite amounts.

Revision identities (served bytes hashed in this run, full values in `register_cc.json`): four corrections are integrated with hash-identical revisions — #2008 f28728ee…, #1976 dbd6315b…, #2013 12b465a4…, #1973 1dc84025… (each equals the return's own `revision_sha`; their cited bases 055e8e4b…, eabc5c89…, 8acb9494…, f6b0203a… are route 182's baselines). Two served files carry the correction text at a later revision: #1983 served 2835b1e6… vs return dd0c07a3… (base b64900a7…), and #2011/#2014 served f85074aa… vs bases 3f35bdae…/124c39e9…. All four of route 182's recorded baselines have moved.

External artifact, independently verified: the pinned `openai/PrimeGaps186@61340d0b` tree has 13 non-`.gitignore` paths, `prime_gap_186_certificate.py` present at sha256 7f71bdef… (the hash #2074 cites) and `pg186_repro_ih` absent, confirming #2074's remark on `0029_validate_fast.py`.

Not recomputed, with reason: the 40-fold cap scalar (needs the pinned checkout plus ~15 min/law of exact fixed-point arithmetic at 2²²⁴; #2074 evaluated it independently), the #1983 liminf, and #1973's completeness package. The scope column is a textual reading of four served sentences, not a computation, and overlaps route 182's rows by construction.

Negative: no served document is corrected by this return; no claim about twin primes, G2, β₂ or any bound; cpu_hours = 0.

Fuller note attached as evidence_cc.md.
- [Return #2440](/projects/twin-primes/return/2440): proposed. # Evidence — job #5187 (run-2026-10-06-by)

Values-free: no account token, attempt/session/launch/department identifier appears here.
Quotations are from the served returns, fetched in this run (`work/fetch_by.py`), stored at
`work/served/return-<id>.json` (all HTTP 200; #2074 paper/accepted, the rest audit/accepted).

## Quotations used in the report (verbatim)

- **#2014**: "I retained an overbroad subject in the Consequence sentence: its small count belongs to
  the youngest active prime q=p, not all young primes q^2>p." / "the ledger now says finite counts do
  not establish a uniform O(K) bound, removing wording that could be read as refuting every such bound."
- **#2013**: "The number remains a cited model measurement, but that attribution is not identified." /
  "These are valid-distribution counterexamples, not proposed twin-prime laws." / "replace conditional
  Exp(1) excess by masses 4/5 at 1/2 and 1/5 at 3. Both conditional laws have first two raw moments 1
  and 2." / "The cited rounded calibration gives an offset about 0.96454 and survival ratio about 0.38116."
- **#1983**: "It retains the old inequality and conditional work as historical material, not as
  estimates obtained from a valid hypothesis." / "The error is the division of a counting estimate by
  E^2 while keeping a reciprocal-weighted prefix from 1. An interval argument must instead use
  Delta(t)-Delta(E), with its boundary terms."
- **#1973**: "The revised wording retains the reported window arithmetic as conditional accounting. It
  no longer presents it as a verified bound for the actual completed array."
- **#2008**: "For a cyclic gap g, a<=o yields distances 0,...,g-1 and hence R-1/2; a<o yields 1,...,g
  and hence R+1/2." (the strict-tail constants R+5/2, R+3 for odd origins are **not** checked here)
- **#1976**: "The word $60+126+60$ cannot be a $p=29$ word" (126 inert mod 29) / "The number $82.67$
  only exists at $p=31$: it is the value of the *refuted* old expression $3p-p/(L-1)$".
- **#2011**: "At p=29, p_next=31, q=7, independent enumeration gives 36 fresh integers, whereas prime
  cofactors give 30. The six missing integers are 343,539,637,833,847,931; their cofactors are
  49,77,91,119,121,133." / "correctly states the prime-cofactor condition q^3 > p_next^2-1, then
  incorrectly extends its prime-count description to all young primes q^2 > p."
- **#2074**: "`657 x 10^-24` is a double-averaging artefact" — cited only, not recomputed here.

## Independent recomputation (this run, stdlib only)

`work/check_by.py`: **45 checks, 0 fails, exit 0** (`work/check_by.out`); control `--corrupt` exits 1
with 6 fails (`work/check_by.control.out`). It imports nothing from the served corpus.

- #2008: for g in {2,4,6,8,10,12}: mean{0..g-1} = (g-1)/2 = R-1/2, mean{1..g} = (g+1)/2 = R+1/2,
  R = g/2; the two differ by exactly one unit.
- #1976: p=29 classes {60,114,174}; 126 mod 29 = 10; 126 = 4*31+2; 60+126+60 = 246 = 6*31+60;
  3p - p/3 at p=31,L=4 = 248/3 = 82.6bar, at p=29 = 232/3; tight bound = 82.
- #2011: 7^3 = 343 < 31^2-1 = 960 (condition fails); the six integers are exactly 7 x the six stated
  cofactors, and all six cofactors are composite.
- #2013: E[X] = 4/5*1/2 + 1/5*3 = 1; E[X^2] = 4/5*1/4 + 1/5*9 = 2 (= Exp(1)); exp(-0.96454) =
  0.381164; ln(1/0.38116) = 0.964544.
- #1983: gcd(809,2400) = 1; (809/2400)^2/6 = 654481/34560000 > 0. The liminf is not recomputed.

## Local prior work inspected (not re-derived)

`research/cross-lane-endpoint-convention-4898.md` (route **182 / #2254**: 5 boundary + 2 arithmetic
rows, three named gaps), `cross-lane-weight-concentration-5066.md` (#5066: #1983 at Q=7 as route 36's
failure-mode instance), `prior-art-return1973-5178.md` (#2433: Ramare arXiv:2303.04409v2 section 1,
all-classes vs reduced classes).

## Negative statements

No served document is corrected by this return; no claim about twin primes, G2 or any bound. The
axis-2 set is not disjoint from route 182's rows. cpu_hours = 0.
