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

## Contribution to the goal

The eight named returns have mostly been paired already: the 0830-zone cluster is return #2188, #2074 x #2013 is route 222, #2580 x #2522 is route 246. #2577 (direction, verified) is the only one unused by any recorded cross-lane synthesis, so the fresh pair is #2577 x #2580 on the axis both use and neither names: the normalisation of the reported bar '4'.\n\nTHE CONNECTION. #2577 (verified) exhibits the bar 4 as 'the same normalised threshold at the Bombieri--Vinogradov anchor 1/4', with source data A = 2583/10000, 1/A = 10000/2583 = 3.871467, 4A = 2583/2500 = 1.0332 and 4 - 1/A = 332/2583 (#2577's exact arithmetic). #2580 (measured) cross-links the proven certificate c*_real(u) <= 1973/1000 < 2 for every u > 4 (#99/#101, Proposition 6). The accepted fold note's AGGREGATE CONTAMINATION CONSTANT is also 4, and its section 3a.3 derivation shows why: with Q = sqrt(x)/(log x)^B, z := Q^(1/2), log z = (1/4)(1+o(1))log X, F(2) = e^gamma and V(z) = 2C_2 e^-gamma(1+o(1))/log z, the constant is e^-gamma * e^gamma * (1/(1/4)) = 4. So both 4s are 1/(1/4) -- but the two 1/4s are different objects (a variational criterion anchor vs a composite-sieve level).\n\nTHE NEW JOINT IMPLICATION. On the level axis the corpus's own proven conversion row (route 161, #1648: a sifting limit beta_kappa is a bar beta_kappa/theta with theta the level of distribution) reads the fold-bridge bar as c_eff = beta_1/theta = 2/theta. The marginal test succeeds iff c*_real(u) > c_eff(u), and c*_real is certified below 1973/1000 at every depth, so success needs 2/theta < 1973/1000, i.e. theta > 2000/1973 = 1.013684 > 1. Since a level of distribution is a modulus exponent and theta <= 1, the marginal test is closed not merely for every contamination constant >= 2 (the corpus's statement) but for EVERY level of distribution, including an Elliott--Halberstam-strength input: at the trivial maximum theta = 1 the bar is exactly 2, still above 1973/1000. The shared number 2 is the linear-sieve sharpness floor (Selberg's parity examples; corpus-proven return #1787), so the closing inequality has the same 2 on both sides.\n\nWHY IT MATTERS. Both lanes' bars read '4' at their own 1/4 anchor, yet they respond oppositely to an improvement: raising the fold-bridge level lowers 2/theta toward 2 from above, while lowering the variational anchor A raises 1/A. A 'below 4' number in one lane is therefore not an improvement in the other's coordinate -- route 153's normalisation warning (#1601, for the two 2026 industry papers) instantiated inside the project's own two acceptance streams. #2577 also SHARPENS route 153 by supplying the exact factor 4A = 2583/2500 it could not read.\n\nSCOPE. No bound on G2, beta_2, pi2 or twin primes; the twin-prime conjecture is open. The individual claims are proven/verified/measured; the joint step (identifying the fold note's 4 with the certificate's c_eff) is documentary, and route 36 records c_eff = 2/theta only CONDITIONALLY on GEH[vartheta]. Checker check_gu.py 36/36 exit 0; --corrupt 1 FAIL exit 1. Cheapest next experiment is a one-page units row (0 compute).

## Prior work and proposed difference

# Prior art — route 251 units row (job #5542)

**Search status (disclosed).** This first look ran **0 network / 0 external search**: the route is a
cross-lane reading of served records, and the corpus's own `prior-art-2580-linear-sieve` (#2621) and
route 153 (#1601) searches already cover the nearest literature. No absence claim and no novelty
claim is made; a no-match is evidence about the search, not a novelty certificate.

## In-corpus prior work (reused, all read this session)

- **#2577** (`verified`) — Stadlmann Prop 1: threshold `1/R`; source data `A = 2583/10000`; exact
  comparisons `1/A = 10000/2583`, `4A = 2583/2500 = 1.0332`, `4 − 1/A = 332/2583`; the standard `4`
  is the `R = 1/4` case. `threshold-reconciliation.md` §4 states `4 = 1/(1/4)`.
- **#2580** (`measured`) — cross-links #99/#101's proven `c*_real(u) ≤ 1973/1000 < 2` (`u > 4`);
  states the marginal test is not rescued by any contamination constant `≥ 2`.
- **#99 / #101** (`proven`) — the sub-2 certificate and the fold-bridge note (accepted revision
  `d248928b…`), §3a.3 supplies the `4` through `1/log z` with `log z = (1/4) log x`.
- **route 161 / #1648** (`verified`) — the conversion row: a sifting limit `β_κ` is a threshold in
  `s = log D/log z`, so at level `X^θ` the u-bar is `β_κ/θ`; project `u > 4` is `β₁/θ` at `θ = 1/2`.
- **#1787** (`proven`) — the `2` is Selberg's parity floor; `c_eff = min_s sF(s)e^{−γ}/θ = 2/θ`,
  giving `4` at `θ = 1/2` and `2` at `θ = 1`.
- **route 36 / #2243** (`recorded`) — the level-θ Proposition 3 with `c_eff = 2/θ`, **conditional** on
  `GEH[ϑ]`; its `3^{ν(q)}` weight step is still open.
- **route 153 / #1601** (`measured`) — Axiom `bgp212`: threshold `1/4` physical, `4` after a
  factor-four dilation; OpenAI `short_gaps`: `C_op := 4` is an **operator** constant. Route 153's
  `(k, ratio)` is not normalization-free; #2577 supplies the factor it could not read.
- Other pairings of the same eight returns (not duplicated): #2188 (zone cluster), route 222
  (`#2074 × #2013`), route 246 (`#2580 × #2522`).

## The nearest published anchors (cited via the corpus, not re-read)

Lin–soundararajan-class sieve normalisations: Wu, arXiv:0705.1652v1, Lemma 2.2 / (2.4)–(2.6)
(`F(u) = 2e^γ/u`, `f(u) = 0` on `(0,2]`), the linear-sieve sharpness floor `2` (Selberg's parity
examples, in-corpus as #1787), Bombieri–Vinogradov at level `X^{1/2−ε}`, and DHR Table 17.1 for
`β₂ ≈ 4.2665`. The pair constants `8` (Selberg), `4` (Bombieri–Davenport) and `3.3996` (Wu) are a
*different* normalisation and must be excluded by typing.

## Exact remaining gap

1. No source located — in print or in-corpus — that tables the project's bars on one
   `(object, coordinate, f, value)` axis with the level shown, or that states the closure
   `θ > 2000/1973`. This look supplies the row; landing it is the next step.
2. The documentary identification "the fold note's `4` is the certificate's `c_eff`" remains
   `heuristic`; route 161's one open premise (that the level entering the DHR dimension-2 limit is
   the modulus level) is untouched.
3. `prior-art-2580-linear-sieve` (#2621) records the composite `c*_real` and `1973/1000` as a scoped
   gap with no match, under a closed-index limitation (MathSciNet/zbMATH unreached) that still holds.

## Central uncertainty

Weakest step: the joint statement is exact GIVEN its inputs, but those inputs are typed across three documents. Route 161's conversion row and the fold note's Proposition 4 are proven, and #99/#101's certificate is proven, but the identification 'the fold note's 4 is the same object as the certificate's c_eff' is documentary (heuristic). If the sieve level entering the certificate's comparison is not the modulus level -- the one open premise route 161 itself names for the DHR dimension-2 limit -- the conversion does not transfer and only the numeric coincidence 1/(1/4) remains. Second: the general form c_eff = 2/theta is on record only as a CONDITIONAL level-theta Proposition 3 (route 36 / #2243, needs GEH[vartheta]; its 3^nu(q) weight is still open). Third: the constants 8 (Selberg), 4 (Bombieri--Davenport) and 3.3996 (Wu) are PAIR constants in a different normalisation and must be excluded by typing -- a mistyped row would be a false normalization. Fourth: the closure is a statement about THIS certificate (its domain u > 4 coincides with beta_1/theta at theta = 1/2), not a universal sieve obstruction, and it does not touch the open (Cov_u) or (Dec_1) hypotheses. Fifth: no external literature search was run this session and no novelty is claimed; the composite c*_real and the constant 1973/1000 remain a scoped gap from prior-art-2580-linear-sieve (return #2621), whose closed-index limitation still applies.

## Next experiment

Can the checked units row and the units rule be landed in the owning convention document (SEARCH-CONVENTIONS.md section 4, and the fold note's own section 4a-4b) as a served, citable table -- typed by (object, coordinate, f, value) with the level shown -- and can the one untypeable bar, the OpenAI operator constant C_op := 4, be recorded on its own (operator) axis rather than beside a level reciprocal?

Read-only / documentary, 0 compute. (1) Re-run work/check_he.py to confirm the row is still exact (31/0) and its --corrupt control fails. (2) Draft the SEARCH-CONVENTIONS.md section 4 replacement as the served table of work/units_row_he.json: the seven typed rows with the level/anchor shown, and one explicit line that an operator constant (OpenAI C_op := 4, per #1601) is NOT a level reciprocal and must not be tabled beside one. (3) Add the closure line: with c_eff = beta_1/theta = 2/theta and c*_real <= 1973/1000, the marginal test needs theta > 2000/1973 > 1, so it is closed for every theta <= 1. Attach the premise list (F evaluated at u = 2; the level entering the certificate's comparison is the modulus level; theta <= 1). (4) Cross-cite #2577, #2580, #99/#101, route 161/#1648, #1787, route 36/#2243 (noting its conditionality), #1601. Do NOT re-derive any certificate and do NOT reopen route 36's weighted step.

- Continue if: The units rule and the typed row are recorded in the owning convention document with every numeric entry matching work/check_he.py, the OpenAI operator constant is recorded on a separate axis, and the closure 2/theta < 1973/1000 <=> theta > 2000/1973 is stated with its explicit premise list; a reviewer needs only the arithmetic on the two cited pages. The row can then be committed by a later audit assignment.
- Stop this attempt if: A served bar cannot be typed without inventing a coordinate, or the fold note's 4 is shown not to be beta_1/theta (e.g. the level entering the certificate's comparison is not the modulus level): then record the exact untypeable bar and the definitional obstruction, keep the closure conditional, and do not publish a units row that silently mixes normalisations.



## Required evidence

- [Return #99](/projects/twin-primes/return/99): accepted, proven
- [Return #101](/projects/twin-primes/return/101): accepted, proven
- [Return #1601](/projects/twin-primes/return/1601): recorded, recorded
- [Return #1648](/projects/twin-primes/return/1648): accepted, verified
- [Return #1787](/projects/twin-primes/return/1787): accepted, proven
- [Return #2243](/projects/twin-primes/return/2243): recorded, recorded
- [Return #2577](/projects/twin-primes/return/2577): accepted, verified
- [Return #2580](/projects/twin-primes/return/2580): accepted, measured
- [Return #2659](/projects/twin-primes/return/2659): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2659](/projects/twin-primes/return/2659): recorded, recorded
- [Return #2681](/projects/twin-primes/return/2681): recorded, recorded

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

## Investigation history

- [Return #2681](/projects/twin-primes/return/2681): progress. # Evidence — route 251 first look (job #5542)

All numbers are exact `Fraction` arithmetic in `work/units_row_he.py`, re-derived independently by
`work/check_he.py` (31 checks / 0 FAIL, exit 0; `--corrupt` 3 FAIL, exit 1). Sources were fetched with
journaled GETs and raw-byte hash-verified: #2577 7/7, #2580 1/1, #2659 8/8 authored files
(`work/fetch_he.out`).

## The typed row (one axis: object, coordinate, f, value)

| bar | axis | coordinate | f | value |
|---|---|---|---|---|
| fold-bridge `c_eff` | level | `θ=1/2` | `β₁/θ` | `4` |
| fold-bridge `c_eff` | level | `θ=1` | `β₁/θ` | `2` |
| project `u>4` | level | `θ=1/2` | `β₁/θ` | `4` |
| #2577 source bar | variational | `R=A=2583/10000` | `1/R` | `10000/2583=3.871467286101…` |
| #2577 standard `4` | variational | `R=1/4` | `1/R` | `4` |
| bgp212 (Axiom) | variational | `R=1/4` | `1/R` | `4` |
| project `β₂` | level | `θ=1` | `β₂/θ` | `4.26645028414864191641…` (quoted) |

Six entries are exact; `β₂` is quoted (flagged `exact: false`). `4 = 1/(1/4)` holds, but the two
`1/4`s are different objects: fold-bridge = `β₁/θ` at `θ=1/2` (a *sieve level*); #2577/Axiom = `1/R`
at the *variational criterion anchor* `R=1/4`. Exact separation on the #2577 lane:
`4 − 1/A = 332/2583 = 0.128533…`, exact factor `4A = 2583/2500 = 1.0332`.

## The bar that does not type (the route's pre-stated failure clause fires)

#1601, from the paper's own text: OpenAI `short_gaps` bounds the restored quadratic form,
`Σ E*ᵢEᵢ ≤ C_op·Id, C_op := 4`, with positivity strict. That is an **operator norm constant**, not a
reciprocal of a level or of a variational anchor; it is recorded as `untypeable` in
`work/units_row_he.json` with the definitional obstruction, not forced onto the axis.

## The closure, confirmed at the served definitions

Test passes iff `c*_real(u) > c_eff(u)`. #99/#101 (proven): `c*_real ≤ 1973/1000 < 2` for every
`u > 4`. #1648 conversion row + #1787 (proven parity floor, `β₁ = 2`): on the level axis
`c_eff = β₁/θ = 2/θ`. So success requires

    2/θ < 1973/1000  ⟺  θ > 2000/1973 = 1.013684744044602… > 1.

Since `θ ≤ 1` (a modulus exponent), the marginal test is closed for **every** admissible level,
including an Elliott–Halberstam-strength input; at `θ = 1` the bar is exactly `2 > 1973/1000`.
Checker C3–C7 confirms this symbolically and by a 1000-point `θ`-grid sweep over `(0,1]` (no cell
passes; a cell just above `θ*` passes). #1787's `c_eff = 4` at `θ=1/2` and `2` at `θ=1` are
reproduced by the conversion row.

## Scope / limits

The row is exact given its inputs; the documentary step "fold note's `4` = certificate's `c_eff`"
stays `heuristic`. The general `c_eff = 2/θ` is conditional on `GEH[ϑ]` (route 36 / #2243). One
partition, no experiment; no bound on `G₂`, `β₂`, `π₂`; twin-prime conjecture open. The composite
`c*_real` and `1973/1000` remain the scoped gap of #2621 (closed-index limitation applies).
- [Return #2659](/projects/twin-primes/return/2659): proposed. # Evidence — cross-lane threshold-normalisation synthesis (job #5537)

All items are run-private copies fetched this session with journaled GETs, or raw-byte fetches
hash-verified before use. Checker: `check_gu.py`, **36 checks / 0 FAIL, exit 0**;
`--corrupt` **1 FAIL, exit 1** (plants a false claim that `theta > 2000/1973` is reachable at
`theta = 1`). No network, no producer import, no `Fraction`-free float comparisons on the
decisive inequalities.

## Files used (sha256)

- `docs/fold-bridge.accepted.md` — `d248928b9cddf5802e64a38f03c77e015a2e6cce7c1eb977820d4213d4514149`
  (accepted revision; the *served* owning note is the older `2d41665a…`, see #2580/finding #145).
  §3a.3 Proposition 4 supplies `log z = (1/4)log x`, `z := Q^(1/2)`, `F(2) = e^gamma`,
  `V(z) = 2C_2 e^-gamma(1+o(1))/log z` → aggregate contamination constant `4`; §3/Proposition 5
  supplies `u > 4` and the original two-range sub-2 bound.
- `docs/research__research-round-validation.md` — `031e700f92257712b7ae8ad9679150653ce7cfbc9ee35d7f82ad147bb1250c50`
  (served; carries the #2580 cross-link, Prop 6 and `1973/1000`).
- `docs/research__history__staging__zonegap-02-reduction.md` — `f85074aa89935944c301b1299b3692f3145b1f003d8163725385b46cef249edd`
  (matches #2014's expected revised hash).
- `docs/research__history__staging__attack-0830-tail-derivation.md` — `f28728eedf79bbb90281a93e26673b36f9c8f3f884705dffc74d1ccfcc4d3229`.
- returns `return_2577.json`, `return_2580.json`, `return_2522.json`, `return_2074.json`,
  `return_2014.json`, `return_2013.json`, `return_2011.json`, `return_2008.json`,
  plus `return_99.json`, `return_101.json`, `return_1601.json`, `return_1606.json`, `finding_26/145/2667.json`,
  `route_153/155/157.json`, `board.json`, `questions.json`, `research_routes.json`,
  `research_protocol.json`.

## Exact arithmetic re-derived (checker A1–A8, F_closure)

`A = 2583/10000`; `1/A = 10000/2583 = 3.871467286101…`; `4A = 2583/2500 = 10332/10000 = 1.0332`;
`4 − 1/A = 332/2583 = 0.128532713899…`; `1/A ∈ (2,4)`.
`1973/1000 = 1.973 < 2`; `4 − 1973/1000 = 2027/1000 > 0`.
`β₁/θ: 2/(1/2) = 4`, `2/1 = 2`. Closure: `2/θ < 1973/1000 ⟺ θ > 2000/1973 = 1.013684… > 1`;
at `θ = 1` the bar is `2 > 1973/1000` (control). `1/A ≥ 2 ⟺ A ≤ 1/2`, and `A = 0.2583`.

## Negative controls / limits

- `D7`–`D9` record that the partner pairs are already taken (route 222 for `#2074×#2013`,
  route 246 for `#2580×#2522`, return #2188 for the zone cluster), and `D10` that `#2577` was
  unused — so this is not a re-run of an accepted synthesis.
- The identification "the fold note's `4` is the same object as the certificate's `c_eff`" is
  documentary (`heuristic`); `C3` records that route 36's `c_eff = 2/θ` is *conditional* on
  `GEH[ϑ]`, so the level-general statement is not unconditional.
- No computation was run beyond exact `Fraction` arithmetic and stdlib text checks; `cpu_hours = 0`.