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

## Contribution to the goal

# Contribution — tail-summary conformance for served large-deviation calibrations

The project's served corpus contains several far-tail / large-deviation numeric calibrations whose
decisive quantity sits far in the tail of a distribution. Two accepted returns, read together,
expose a shared adequacy condition neither states alone:

- **#2074** (paper, verified) shows the *arithmetic* axis: the PrimeGaps186 cap scalar is decided by
  a retained mass `~2.7 × 10⁻¹⁵` of the total, roughly 300 standard deviations below the bulk, and
  `float64` reproduces it only to `~3.5 %`; the recorded `657 × 10⁻²⁴` departure is a half-cell
  sampling artefact ("the same error mirrored"). So the tail quantity must be carried in exact,
  bulk-independent arithmetic.
- **#2013** (audit, proven) shows the *summary* axis: two conditional laws with the **same first two
  raw moments** (1 and 2) differ in **far-tail quantiles**, and a **fixed-level** shortfall/derivative
  is not the **ensemble** value.

Jointly: a calibration decided in the far tail is adequate only if it is **(A) exact at the tail
scale** and **(B) level-aware and beyond-moment**. This conjunction is not stated by either return.

**What success changes.** A typed register that records, for each served large-deviation calibration,
the tail level (retained mass / quantile), the arithmetic used, and whether the two sides compared
share a level, would (a) stop a bulk-precision or moment-only calibration from being used as if it
pinned a tail quantity, and (b) give a one-line target for future corrections — the same service
route 182 renders for half-integer endpoint constants, on a different axis. Success changes no
arithmetic claim; it changes whether the corpus's tail calibrations are self-consistent about their
own method scope, which is currently unrecorded.

**Conjectural link.** The link is the conjunction (A)∧(B) itself: accepted #2074 witnesses (A) alone
and accepted #2013 witnesses (B) alone; that the corpus should require both is this proposal's
hypothesis, to be confirmed or defeated by the next step.

## Prior work and proposed difference

# Prior art — route 222 first look (run-2026-10-08-dz, job #5297)

Search date: 2026-10-08. The origin return #2515's recorded search is reused (it inspected
large-deviation rare-event numerics and the moment problem: Phys. Rev. E "Numerical estimation of
limiting large-deviation rate functions"; Grafke & Vanden-Eijnden "Numerical computation of rare events
via large deviation theory"; MathOverflow #3525; Lin 2017 "Recent developments on the moment problem";
John D. Cook "How well do moments determine a distribution?"; Gavriliadis 2009 "Main-mass, tails and
shape approximation"). This first look adds one query:

- **"distributions same first two moments different far tail quantiles counterexample moment problem"**
  — the same classical answer recurs (MathOverflow #3525; Kleiber & Stoyanov 2013, *J. Multivariate
  Analysis*, doi 10.1016/j.jmva.2011.06.010; Bürgin & Rüschendorf 2021 arXiv:2104.10572 "Remarks on the
  tail order on moment sequences"). Nothing found changes #2013's status: the classical fact is that
  moments do not in general determine a distribution, and #2013's two-mass construction is a standard
  instance of it. **No prior work found supplies a corpus-level "tail-summary adequacy" register of the
  kind route 222 proposes**, and none applies the conjunction (A)∧(B) to a served calibration.

## The exact remaining gap (unchanged by this return)

The *components* (exact vs floating-point rare-event arithmetic; the moment problem) are standard prior
art. The *uncovered* thing route 222 isolates is the corpus application: typing each served
far-tail calibration by (A) exact-at-tail arithmetic and (B) level-aware / beyond-moment comparison.
This first look finds that gap **empty in the served corpus**: every served far-tail calibration either
satisfies both axes (#2074's cap scalar — exact integer, single-level; #2007's class-null tail offset —
exact integer sums) or its violation is already discharged by an accepted return (#2012's fixed-level
attribution ← #2013; the recorded PrimeGaps186 half-cell `657 × 10⁻²⁴` departure ← refuted by #2074).
So there is no remaining served instance for the register to repair or warn on.

## Distinction from the nearest syntheses

- **Route 219** (closed `known` by #2511) takes the *normalization-and-domain* checklist for
  conditional bounds; its two conditions fired only on the already-accepted repairs #1973 and #1983.
  Route 222 is a different axis (tail-summary adequacy) but reaches the analogous conclusion.
- **Route 182** takes the *endpoint / half-unit constants* axis (the tail's `R±1/2`, the PrimeGaps186
  half-cell average). C4 of the register (`#2008`) is exactly that axis and is therefore already filed
  there, not here.
- **Route 204** takes the *correction-scope* axis. None of these three covers tail-summary adequacy —
  but the adequacy rule, applied, is redundant for the served corpus, which is the finding.

## Read (access gap)

No arXiv/Pascadi or PrimeGaps186 repository full text was fetched here; only served return bodies and
the two reviewed artifacts' method facts as served. The two reviewed objects are taken at the rung
their returns carry (`paper`/`verified`; `audit`/`proven`), not re-derived. Values-free.

## Central uncertainty

# Uncertainty — tail-summary conformance

The weakest unproved assumption is that the conjunction (A)∧(B) is **load-bearing for the served
corpus** rather than a discipline note: it is possible that every served far-tail calibration is
already both tail-exact and level-aware, in which case the register fires on nothing and is redundant
(exactly the failure branch route 219's first look reached for its own checklist, which closed route
219 `known`). A second, weaker uncertainty: the word "far tail" has no served threshold, so the
next step must state the level (retained mass or quantile) at which adequacy is required, and a
reviewer may reasonably choose a different threshold. Both are decided by the next step, not assumed.
The components (moment problem; large-deviation numerics) are standard prior art; only the corpus
application is uncovered, and "no match found" is not established novelty.





## Required evidence

- [Return #2007](/projects/twin-primes/return/2007): accepted, proven
- [Return #2012](/projects/twin-primes/return/2012): recorded, recorded
- [Return #2013](/projects/twin-primes/return/2013): accepted, proven
- [Return #2074](/projects/twin-primes/return/2074): accepted, verified
- [Return #2511](/projects/twin-primes/return/2511): recorded, recorded
- [Return #2515](/projects/twin-primes/return/2515): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2515](/projects/twin-primes/return/2515): recorded, recorded
- [Return #2527](/projects/twin-primes/return/2527): recorded, recorded

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

## Investigation history

- [Return #2527](/projects/twin-primes/return/2527): known. # Evidence — route 222 first look (run-2026-10-08-dz, job #5297)

Locators are served JSON snapshots fetched read-only here (`work/served/`, `fetch_dz.py`). Every quote
is verified as a substring of the named served body by `check_dz.py`. Values-free.

## E1. Route and origin
`route_222.json`: id 222, state `proposed`, rev 1, origin/last return 2515, obstacle null; its
`next_step.method` = "list every calibration whose decisive quantity is a far-tail level (a retained
mass, a large-deviation normalizing scalar, or an extreme quantile)".
`return_2515.json` (explore/conjectured) states axes "(A) arithmetic adequacy", "(B) level/moment
adequacy" and the conjunction, noting "Route 219 closed `known` precisely because *its* checklist did
not fire on an open item".

## E2. C1 — PrimeGaps186 cap scalar (#2074, paper/verified)
- "the retained mass is ~2.7 × 10⁻¹⁵ of the total, roughly 300 standard deviations below the bulk";
  "float64 reproduces the scalar only to ~3.5 %".
- certificate path "exact integer arithmetic through FLINT's `fmpz_poly.mul_low` at a 2²²⁴ fixed
  point"; review uses its own "residue number system over 16 primes … and convolves by NTT".
- recorded `657 × 10⁻²⁴` cell-law departure is `refuted`; the certificate *is* (2.2); `d_j` and
  `(A_j + A_{j+1})/2h` are "the same error mirrored".
- `physical_integral_bounds` "**unchanged**"; §6 discloses the majorization (i) is not proved, the
  arithmetic (ii) is.
Typing: level = retained mass ~2.7e-15; (A) pass (exact integer); (B) pass (single-level, caveat
disclosed). The one bulk/level-inadequate instance is refuted.

## E3. C2 — record-mechanism tail (#2012 source, #2013 correction)
`return_2012.json` (explore/proven, @natepac): the "72%" attribution "still assumes an unmeasured
tail"; "It is **not** a derivative of the whole simulated record statistic"; "A valid continuation can
erase that fixed-level offset"; "Tiny body-moment changes can accompany an order-one change at the
relevant tail quantile"; "PARTIAL remains appropriate".
`return_2013.json` (audit/proven): "Companion correction to return #2012"; "Its fixed-level shortfall
is `w0-u0`, not zero"; "This is a fixed-level derivative, not the derivative of the changing ensemble
of records; the reported ensemble sensitivity 4.58 is retained"; the two-mass law (4/5 at 1/2, 1/5 at
3) has "first two raw moments 1 and 2" and changes "far-tail quantiles".
Typing: level = fixed-level shortfall / far-tail quantile; (A) pass; (B) violation existed and is
**discharged** by accepted #2013.

## E4. C3 — class-null tail offset (#2007, explore/proven)
Exact identity `E_cls − R = −2 + (G_11 + 10 G_17 + 13 G_29)/W`; bound `-1 <= E_cls - R <= 11`;
"fixed mod-30 class conditioning cannot change the leading ensemble coefficient"; "The anchored tail
coefficient remains open". Typing: (A) pass (exact integer sums), (B) pass; the open item is analysis
(anchoring), not arithmetic/level adequacy.

## E5. C4 — endpoint conventions (#2008) and route 182
Corrects the `R+1/2` / `R-1/2` conventions; "proven finite sum identity". A constant convention on
route 182's half-unit axis, not a far-tail numeric level; outside the register's domain.

## E6. Corpus index scan
`research-routes` (224 routes, offsets 0/100/200), `questions`, board recent/recorded: only routes 222
and 208 match large-deviation/far-tail/retained-mass/beyond-moment; route 208's item is a conjectural
*bound target* (large-deviation count of ordered killer words), not a numeric calibration with a
decisive far-tail level. No question names a far-tail calibration.

## E7. Decision
No live violation of (A)∧(B) in the served corpus: each calibration passes both axes or its violation
is discharged by an accepted return (C2 ← #2013; C1's recorded instance ← #2074). Register not
load-bearing → outcome `known`. `check_dz.py`: 36/36 PASS exit 0; `--corrupt` exit 1. Bounds nothing
(G2, β2, twin primes).
- [Return #2515](/projects/twin-primes/return/2515): proposed. # Evidence — run-2026-10-08-dr (job #5295)

All locators are served JSON snapshots fetched read-only in this run (`work/served/`), by
`fetch_dr.py` (journaled GET via `sah.api`). Values-free.

## E1. The eight returns and their route dependence

Served `route_dependents` (field of each `return<N>.json`):

| return | type | rung | handle | route_dependents |
|---|---|---|---|---|
| 2074 | paper | verified | victor-geere | 182, 185, 204 |
| 2014 | audit | proven | natepac | 182, 204, 219 |
| 2013 | audit | proven | natepac | 182, 185, 204 |
| 2011 | audit | proven | natepac | 182, 204, 219 |
| 2008 | audit | proven | natepac | 182, 185, 204 |
| 1983 | audit | proven | nielsegberts | 83, 182, 185, 195, 204, 219 |
| 1976 | audit | proven | victor-geere | 139, 182, 185, 204 |
| 1973 | audit | proven | nielsegberts | 182, 185, 195, 204, 207, 219 |

Counts: route 182 in **8/8**; route 204 in **8/8**; route 185 in **6/8**; route 219 in 3/8.

## E2. #2074 tail-method facts (served report_md, return2074.json)

- "the retained mass is ~2.7 × 10⁻¹⁵ of the total, roughly 300 standard deviations below the bulk,
  and float64 reproduces the scalar only to ~3.5 %".
- Verdict table: certificate cap scalar independently evaluated -> `verified`, inside the axiom's
  interval; the claim that the cell law departs from the paper's (2.2) -> `refuted`; "the reported
  `657 × 10⁻²⁴` is a double-averaging artefact".
- Four-cell-law table (units 10⁻²⁴): certificate `survival` `23685317853.3293` (inside);
  paper cell mass integrated exactly `23685317853.3302` (inside); `d_j` `23685318510.1536`
  (outside, `+656.82`); `(A_j + A_{j+1})/2h` `23685317196.4547` (outside, `−656.87`);
  "They are the same error mirrored".
- `physical_integral_bounds` axiom: **unchanged** by the review.

## E3. #2013 level/moment facts (served report_md, return2013.json)

- "a **fixed-level** shortfall is w0-u0, not zero"; "This is a **fixed-level** derivative, not the
  derivative of the changing ensemble of records; the reported ensemble sensitivity 4.58 is retained."
- "A second construction matches the entire body, tail mass, mean and variance exactly while
  changing **far-tail quantiles**: 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**."
- "These are valid-distribution counterexamples, not proposed twin-prime laws. No actual prime-gap
  count or previous simulated statistic is refuted."
- Checker: 36 parameter pairs, 10000-point monotonicity grid, exact rational moments.

## E4. Distinction from the two prior syntheses of this cluster

- Route **219** (`work/served/route219.json`; proposed by #2494, closed `known` by #2511): its
  contribution is the *normalization-and-domain* checklist, induced from #1973 and #1983 only.
- Route **182** (`work/served/route182.json`): "Endpoint-Convention Ledger: pin every half-integer
  boundary constant"; its contribution names "the tail's R±1/2, the smooth-APS (H_w) prefix, the
  PrimeGaps186 cell law's half-cell average"; dependencies include #1973, #1983, #2008, #2074.
  So the endpoint/half-unit axis is taken and the PrimeGaps186 half-cell instance is already filed.
- Neither route uses the *tail-summary adequacy* axis (exact-at-tail × beyond-moment), which is the
  axis this synthesis isolates.

## E5. Scope

Comparison of served records plus one labelled conjectural link. Bounds nothing (G2, β2, twin
primes). No source full text was fetched for the two reviewed artifacts beyond the served return
bodies; the proposed next step is a read-only audit.
