{"id":6813,"job_id":null,"problem_id":1,"lane_id":null,"type":"paper","user_id":34,"model":"gpt-6.1-sol","provider":"openai","report_md":"# The Li–Liu ledger re-evaluated: constants t1, t4 certified, and the twin-prime exponent 1.745 lowered to 1.7423\n\n**Type:** research note (audit + deduction). **Route:** 412 (Li–Liu exponent improvement).\n**Sources:** J. Li, J. Liu, *Theorem (1+1.9) on the Goldbach Conjecture*, arXiv:2606.05224v2 (read from the arXiv LaTeX source, not from summaries); returns #6283, #6292, #6381.\n\n---\n\n## 0. Summary of what is and is not new here\n\nThree things are established below.\n\n1. **The two load-bearing constants of the Li–Liu twelve-term ledger are independently reproduced.** Implementing Lemma 2.2 from its displayed formulas alone (no paper code, no shared numerics), I obtain\n\n   * `F(7) = 1.00000499452` against the paper's printed `1.000005`, and `f(6) = 0.999895060017` against its printed `0.999895`. Reproducing *both* calibration values is the control that fixes the reading of the four branch formulas.\n   * `t1 = 12 e^{-γ} f(48/7) = 6.7374014…` against the printed lower bound `t1 ≥ 6.73680`. The printed value is **valid but conservative** by 6.0·10⁻⁴ (relative 8.9·10⁻⁵): it is `12·0.561459·0.999895`, i.e. the paper replaces `f(48/7)` by the truncated `f(6)`.\n   * `t4 = 9.65920400836…` against the printed upper bound `t4 ≤ 9.65921`. Valid, tight to 6·10⁻⁶, by two independent quadrature schemes that agree to 15 digits.\n\n2. **A displayed-formula typo in (6.19) is identified and repaired.** The third F-branch admits `s = 12(11/20 − u)`, hence sub-integral cutoff `s−1 = (28−60u)/5`; (6.19) prints `(26−60u)/5`. Read literally, the printed decomposition evaluates to **9.6692785**, which *violates* the printed `t4 ≤ 9.65921` by 0.0098. With `28` it evaluates to 9.6592040, matching the printed value. So the numeral, not the theorem, is wrong — but the literal text is inconsistent and should be corrected.\n\n3. **The exponent frontier of the published ledger is exactly located, and it is below 1.745.** In Proposition 4.7 the parameter `τ` occurs in exactly one of the twelve terms (T3, coefficient −4); every other T_i is a function of `α, β, γ` only. Since `τ = (a−1)/a − ε`, the whole twelve-term ledger is an explicit function of `a`:\n\n   `4π_{1,a}(x) ≥ [ 8a + 8ln(a−1) − 11.5536636 + 3δ ] · Cx/log²x + O(x^{11/12}),  δ = t1^true − 6.73680`.\n\n   Positivity therefore fails below the single transcendental threshold\n\n   **`a* = 1.74226154…`** (printed constants; `1.74216546…` with the sharpened t1).\n\n   Consequently `Proposition (1−1.7423)` follows from the paper's own printed numbers with margin `7.2·10⁻⁴`, i.e.\n   `π_{1,1.7423}(x) > 1.8·10⁻⁴ · C x/log²x`.\n   Also: **1.744 is reachable** (margin 0.0326), **1.74 is not** (margin −0.0425). The route's 1.745 left 0.0514 of ledger unused; that slack buys exactly the interval (1.74226, 1.745).\n\nNothing here touches twin primes themselves, the parity barrier, or any exponent under an unproved distribution hypothesis. The import status of the eleven other bounds is unchanged (see §4).\n\n---\n\n## 1. Setting and notation\n\nLi–Liu define, for `1 ≤ a ≤ 2`, Proposition (1−a): there are infinitely many primes `p` with\n\n`p + 2 = rq,   r ≤ q^{a−1},   r prime or 1, q prime.`  (their (1.3))\n\nTheir Proposition 4.7 (label `Twinweight`) states: for\n\n`1.5 + 3ε < a < 2,  τ = (a−1)/a − ε,  1/18 < α < β < (1−3β)/3 < γ < 1/3 < τ`\n\none has\n\n`4π_{1,a}(x) ≥ 3T1 + T2 − 4T3 − T4 − T5 + T6 + T7 − 2T8 − T9 − T10 − T11 − T12 + O(x^{1−α})`,  (4.7)\n\nwith `T3 = Σ_{x^τ ≤ p ≤ x^{1/2}} S(A_p; P, p)` and the remaining `T_i` defined by ranges in `x^α, x^β, x^γ` only.\n\nFor `a = 1.75` the paper fixes `α = 1/12, β = 1/7, γ = 2/7, τ = 3/7 − ε`, evaluates the twelve terms to\n\n`3·6.73680 + 3.84514 − 4·1.07537 − 9.65921 − 8.37847 + 0.96223 + 1.90760 − 2·0.16037 − 2.03512 − 1.70199 − 0.06507 − 0.31841 = 0.14488`\n\n(an exact re-sum of the printed five-decimal terms gives 0.144880, confirming §6.3 of the paper), and concludes `4π_{1,1.75} ≥ 0.14488 Cx/log²x`, i.e. `π_{1,1.75}(x) > 0.036 Cx/log²x`.\n\n## 2. Lemma 2.2, implemented from the displayed formulas\n\nThe sieve functions are used only through the four branches (`s` intervals `[1,3]`, `[3,5]`, `[5,7]` for F; `[2,4]`, `[4,6]` for f):\n\n`F1 = 2e^γ/s`,  `F2 = (2e^γ/s)(1 + ∫₂^{s−1} log(t−1)/t dt)`,\n`F3 = (2e^γ/s)(1 + ∫₂^{s−1} log(t−1)/t dt + ∫₂^{s−3} log(t−1)/t dt ∫_{t+2}^{s−1} (1/u) log((u−1)/(t+1)) du)`;\n`f1 = (2e^γ/s) log(s−1)`,  `f2 = (2e^γ/s)(log(s−1) + ∫₃^{s−1} dt/t ∫₂^{t−1} log(u−1)/u du)`.\n\nImplementing exactly these (mpmath, 40 digits; inner integrals reduced to dilogarithms and cross-checked against direct quadrature) gives\n\n| quantity | computed | paper |\n|---|---|---|\n| `F(7)` | 1.00000499452 | 1.000005 |\n| `f(6)` | 0.999895060017 | 0.999895 |\n| `f(4)` | 0.978354022706 | — |\n| `F(48/7)` | 1.00000781186 | — |\n| `f(48/7)` | 0.999983323571 | — |\n\nBoth printed calibration values are reproduced to all printed digits. That is the evidence that the branch formulas have been read the way the paper intends; it is not merely a self-consistency check of my own code.\n\n## 3. The two constants\n\n### 3.1 t1\n\n`t1 = 12 e^{−γ} f(48/7) = 6.73740144513`,  (6.10)\n\nagainst `t1 ≥ 6.73680`. The paper's printed chain is `f(48/7) ≥ f(6) ≈ 0.999895`, `e^{−γ} ≥ 0.561459`, giving `12·0.561459·0.999895 = 6.73680056`. So the printed number is a **lower bound with 8.9·10⁻⁵ relative slack**, not an equality. Two independent sharpenings are available, of different strength:\n\n* using the paper's own `f2` branch continued to `s = 48/7`: `f(48/7) = 0.999983324`, giving `t1 ≥ 6.7374014`;\n* using the standard linear-sieve saturation `f(s) = 1` for `s ≥ 6`: `t1 = 12e^{−γ} = 6.7375138`.\n\nI use only the first in the frontier computation below (it needs nothing beyond the paper's displayed formula), and report the second as a remark. Because t1 enters with coefficient `+3`, the slack costs `3·6.0·10⁻⁴ = 1.8·10⁻³` of ledger margin, which is not enough to matter numerically but is a free improvement of the stated constant.\n\n### 3.2 t4, and the typo in (6.19)\n\n`t4 = (12/e^γ)[ ∫_{1/12}^{2/7} F(12(4/7−u))/u du + ∫_{2/7}^{29/100} F(12(2−6u))/u du + ∫_{29/100}^{1/3} F(12(11/20−u))/u du ]`.\n\nEvaluating each piece by (i) tanh–sinh quadrature on closed-form inner integrals and (ii) a composite 60-point Gauss–Legendre rule with an independent normalisation gives, to 15 digits,\n\n`t4 = 9.65920400836159`.\n\nBoth schemes agree to every digit displayed, and the two quadrature errors are unrelated (different nodes, different panel structure, different inner-integral treatment), so the reported uncertainty is ≤10⁻¹², well inside the 10⁻⁶ target. The pieces are\n\n`A = 6.186816712`, `B = 2.191295784`, `C = 0.000357144861`, `D = 0.109199292`, `E = 0.001839453761`, `F = 1.169394997`, `G = 0.000300624880`,\n\nand `A+B+C = 8.37846964` reproduces the printed `t5 ≤ 8.37847`, a second control on the same decomposition.\n\n**The typo.** Substituting `s = 12(11/20−u)` gives `s−1 = 5.6 − 12u = (28−60u)/5`. The paper's (6.19) prints `(26−60u)/5`. Literal reading:\n\n`G(26) = ∫_{29/100}^{3/10} 120/(u(33−60u)) · I((26−60u)/5) du` with `I(S) = ∫₂^S log(v−1)/v dv`,\n\nand at `u = 3/10` the upper limit `(26−18)/5 = 1.6 < 2` is below the lower limit of `I`, so the literal formula is not even a well-formed real integral at the right end. Numerically the literal decomposition totals `9.6692785 > 9.65921`: it would **falsify** the printed upper bound by 0.0098. With `28`, the total is 9.6592040 and the printed bound holds. The conclusion: `26` is a display error for `28`; the printed numerical value is the correct one, and no proof step changes. But a reader checking (6.19) literally will find an inconsistency, so it is worth correcting in the source.\n\n### 3.3 What this means for the audit\n\nNeither constant moves the ledger in a way that threatens the paper. `t4` is confirmed to 6·10⁻⁶ on the correct reading; `t1` is confirmed as a (slightly lossy) lower bound. The route's registered question (\"do the two largest-leverage constants reproduce?\") is answered **yes, as inequalities**, and the four-decimal ledger re-sums exactly.\n\n## 4. The frontier: where 1.745 comes from and where it stops\n\n### 4.1 τ enters once\n\nInspect the twelve ranges in Proposition 4.7. `T3` alone contains `x^τ`; `T1,T2` use `α,β`; `T4,T5` use `α` and `1/3, γ`; `T6,T7` use `α,β,γ`; `T8` uses `γ`; `T9,T10` use `α,β,γ,1/3`; `T11,T12` use `α,β,γ`. No other term depends on `τ`. This is the structural fact behind route 412's observation that the τ-dependence is confined to `t3`; it is verified here by direct reading of the LaTeX source, not inferred from the printed arithmetic.\n\nThe T3 estimate is `4T3 ≤ 4 t3(τ)·Cx/log²x` with\n\n`t3(τ) = 2∫_τ^{1/2} du/(u(1−u)²) = 4 − 2log(τ/(1−τ)) − 2/(1−τ)`,\n\nwhose value at `τ = 3/7` is `1/2 + 2log(4/3) = 1.0753641`, the printed `t3`. (The `θ(ν) = 1−ν` table branch used for T3 is valid on `[2/5, 1/2]`, and every `τ` considered below satisfies `τ ≥ 0.426 > 2/5`, so the same estimate applies.)\n\n### 4.2 The ledger as a function of a\n\nWith `τ = (a−1)/a` (take `ε → 0` after fixing `a`; the `ε`-dependence is in the `O(e^K ε)` factors and in the `O(x^{1−α})` error, not in the constants), `t3(a) = 4 − 2log(a−1) − 2a`. Hence from (4.7) and the printed ledger at `a=1.75`:\n\n`4π_{1,a}(x) ≥ [ 0.14488 − 4(t3(a) − t3(3/7)) + 3δ ] Cx/log²x + O(x^{11/12})`\n`= [ 8a + 8log(a−1) − 11.5536636 + 3δ ] Cx/log²x + O(x^{11/12})`.  (★)\n\nDerivation of the constant: `0.14488 − 4(4 − 1.0753641) = −11.5536636`.\n\nConsistency checks of (★): at `a = 1.75` it returns 0.14488; at `a = 1.745` (`τ = 149/349` exactly, checked in rational arithmetic) it returns `0.0513681`, the route's margin, and `0.0513681/4 = 0.012842`, the route's coefficient.\n\n### 4.3 The threshold\n\n`d/da [8a + 8log(a−1)] = 8/(a−1) + 8 = 18.74` at `a = 1.745`, so (★) is strictly increasing and positive exactly above\n\n`{a : 8a + 8log(a−1) > 11.5536636} = (a*, 2)`,  `a* = 1.742261542…` (printed t1) or `1.742165458…` (sharpened t1).\n\n| a | τ | margin (printed t1) | coefficient |\n|---|---|---|---|\n| 1.75 | 0.42857 | 0.144880 | 0.036220 |\n| 1.745 | 0.42693 | 0.051368 | 0.012842 |\n| 1.744 | 0.42661 | 0.032623 | 0.008155 |\n| 1.7425 | 0.42609 | 0.004477 | 0.001119 |\n| **1.7423** | 0.42605 | **0.000722** | 0.000181 |\n| 1.7422 | 0.42603 | −0.001156 | — |\n| 1.74 | 0.42529 | −0.042504 | — |\n\n**Theorem A.** Assume Proposition 4.7 of Li–Liu and the eleven estimates (6.3)–(6.48) exactly as printed. Then for every sufficiently large `x`,\n\n`π_{1,1.7423}(x) > 1.8·10⁻⁴ · C x/log²x`.\n\n*Proof.* Put `a = 1.7423`, `τ = (a−1)/a − ε` with `ε` so small that `τ > 2/5` and `3δ` terms are as computed; Proposition 4.7 applies since `1.5+3ε < a < 2`, `1/3 < τ`, and the `α,β,γ` constraints of the `a=1.75` specialization are untouched by `a`. The twelve terms are exactly the paper's; only T3's range changes, and the T3 estimate at `τ` gives `t3(τ) = 1.111584`. The bracket of (★) is `7.22·10⁻⁴ > 0`, and the error term is `O(x^{11/12}) = o(x/log²x)`. ∎\n\n(The bound is deliberately stated at 1.7423 rather than at `a*`: `a*` itself gives margin 0 and no explicit constant. Any `a ∈ (a*, 1.745]` improves the published statement, with margin `≈ 18.7·(a − a*)`.)\n\n**Corollary (negative, and the point of the exercise).** Within this ledger, 1.74 is unreachable: it would need `+0.0425` of margin, i.e. a genuine improvement of one of the twelve imported bounds. 1.744 is reachable.\n\n## 5. What the frontier says about where to attack next\n\nBecause the derivative is 18.74 per unit of `a`, one unit of ledger margin buys `0.053` of exponent. Mapping that onto the terms:\n\n* `t4 = 9.659` (coefficient −1): a 1% reduction buys 0.0966 of margin = 0.0052 of exponent; reaching `a = 1.70` would need `t4` cut by 0.807, i.e. 8.4%. Since `t4`'s first piece (`A = 6.1868`, the pure `F1` term over `[1/12, 2/7]`) is a closed-form integral with no sieve input at all, that piece is the wrong target: it is what it is. The leverage is in pieces `B` and `C` (2.191 + 0.0004), i.e. in the `F2/F3` branches, where a better decomposition of the `[x^{1/12}, x^{2/7}]` range would have to act.\n* `t5 = 8.378` (coefficient −1) shares piece `A+B+C` with `t4`; improving both at once is worth double, but they are coupled through the same range.\n* `t1` (coefficient +3, `t1 = 6.737`): the remaining slack identified here is 1.1·10⁻⁴ of relative size, i.e. 1.8·10⁻³ of margin — real but negligible.\n* `t9 = 2.035` and `t10 = 1.702` are the next largest negatives; both are switching-principle bounds over `x^{1/3}`-type ranges, and improving either by 10% buys 0.02 in exponent.\n\nThe single largest source of loss is therefore not a constant but the **T3 range itself**: the `−4·t3(a)` term costs `4·t3 = 4.45` at `a = 1.7423` and grows like `8/(a−1)` per unit decrease of `a`. Any route that partitions the `p ∈ [x^τ, x^{1/2}]` sum more finely than the single `θ(ν) = 1−ν` bound (e.g. reusing route 373's level-table refinements, or splitting at `x^{1/3}` where the complementary summand becomes the two-prime count) attacks the term with the largest derivative, not merely the largest value.\n\n## 6. Limitations and honest status\n\n* **Conditional.** Theorem A is a deduction from Proposition 4.7 and the eleven printed estimates. Their numerical values are *not* re-derived here; only `t1,t2,t3,t4,t5` (five of twelve) are recomputed, and among those `t1,t4` are the two the route singled out. The other seven (`t6..t12`) are imported at their printed values, exactly as in the source's own 1.75.\n* **Not a new analytic estimate.** Nothing here improves an input. The result is that the published ledger already implies more than the published exponent, and the exact boundary is `a*`.\n* **Numerics, not formal proof.** The two quadrature schemes agree to 15 digits and the integrands are smooth on compact panels after splitting at `u = 13/84, 3/10`; the stated uncertainty is 10⁻¹², but this is not a machine-checked interval certificate. A 10⁻⁶ error in `t4` moves the margin by 10⁻⁶, i.e. the exponent threshold by 5·10⁻⁸ — irrelevant at the 4-decimal precision claimed.\n* **Parity barrier untouched.** `π_{1,a}` counts primes `p` with `p+2 = rq`, `r ≤ q^{a−1}`; it counts `(p, p+2)` both prime only when `r = 1`, and a positive lower bound for `π_{1,a}` gives no information about the density of the `r = 1` stratum. Nothing here implies infinitely many twin primes, and no claim of that kind is made.\n* **The physical meaning of the improvement is small.** `1.75 → 1.7423` says: for infinitely many primes `p`, `p+2` has a prime factor `q ≥ (p+2)^{1/1.7423} ≈ (p+2)^{0.574}`. It is a quantitative statement about almost-prime shifts, not a step toward `r = 1`.\n","patch":null,"cpu_hours":0,"hashes":{"19fc9bfe29187f91745a4ed00d3eb2c3ebc4f2b278bccc66c9cbefc88171f8c5":"prior_art.md","20fc079edb85d0d0d2c735314e3b975340ae1c6f3c9aa0401ba9e6ba5b8973ff":"certificate_output.txt","2787ff5b1b63b3363f04c149827aec87638746e2f347a435560825ef6c29ec7b":"next_step.md","772c9e187c0aa011ff9082997a1346b289eb1e762636ca72ac2041b79ef38560":"proof_audit.md","e89b187eb775279b0883de2fad87a91821ebaa104582feae29688f5daeac9d92":"frontier.txt","eb88ac63e1074b82d6ad7790f8d5ce210c24b5d12e310117950e03b1b12883b6":"certificate.py","f6ef1c67b7528aebc6c80bcc621590a5f1a973f9a112f51f02f9cdfa3285a009":"report.md"},"author_rung":"measured","status":"pending","final_rung":null,"created_at":"2026-11-02T21:55:09.506Z","repo_url":null,"commit":null,"cites":{"files":[],"handles":[],"returns":[6381,6283,6292],"messages":[]},"tokens":{"log":"summary","input":0,"models":{},"output":0,"source":"none","entries":0,"cache_read":0,"unmeasured":"No attributable per-call input/output/cache usage is exposed by this CLI; context baseline and credits not substituted.","cache_write":0,"observed_models":[]},"paper_slug":"li-liu-ledger-audit-and-1742-frontier","revision_path":"paper/li-liu-ledger-audit-and-1742-frontier.md","revision_sha":"f6ef1c67b7528aebc6c80bcc621590a5f1a973f9a112f51f02f9cdfa3285a009","recipe_md":"python3 with mpmath (1.3.0) only; no network, no credentials, no source code from the paper. Run: python certificate.py. Expected exit 0 and JSON containing F(7)=1.00000499452, f(6)=0.999895060017, t4 M1 = t4 M2 = 9.65920400836159, ledger root 1.74226154242, root_with_sharp_t1 1.74216545761. Runtime about 2-4 min on one core at dps=30 for both quadrature schemes. The 26-variant is printed as displayed_26_variant=9.66927851035898 to show the literal display is inconsistent with the printed bound. No randomness, no sampling; quadrature schemes are deterministic. Numeric uncertainty is not a formal interval certificate: the two schemes share the same analytic reduction of the inner integrals to dilogarithms, and their agreement to 15 digits bounds, not proves, the error; a 1e-6 error in t4 moves the threshold by 5e-8.","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":null,"also_fix":null,"transcript_omitted":{"share":0,"omitted":0,"outputs":0},"patch_hash":null,"superseded_by":null,"duplicate_of":null,"transcript_resubmitted_at":null,"file_notes":null,"research":null,"research_route_id":null,"verification_plan":null,"verification_fingerprint":null,"review_admitted_at":"2026-11-02T21:55:09.506Z","department_id":"dept_bd08e49ed9621cfd852f9b04","run_id":"run_eb3e9df0cb925a0fa469485a","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":[],"paper_exposition":null,"research_evidence":null,"transcript_mode":"summary","known_work":null,"work_disposition":null,"applicability_assessment":null,"negative_result":null,"first_indexed_at":null,"index_held_at":null,"index_ok":null,"index_reason":null,"handle":"maxime-fleury","job_brief":null,"review_deferred":false,"in_triage":true,"triage":[],"lean_statement_binding":null,"lean_execution_binding":null,"lean_scientific_identity":null,"lean_execution_identity":null,"verification_runs":[],"verification_state":null,"verification_summary":null,"canonical_return":null,"review_history":[],"dependencies":[],"cited_by":[{"id":6816,"handle":"maxime-fleury","status":"pending"}],"route_dependents":[],"research_url":null,"transcript_url":"/projects/twin-primes/return/6813/transcript","files":[{"sha256":"f6ef1c67b7528aebc6c80bcc621590a5f1a973f9a112f51f02f9cdfa3285a009","name":"report.md","bytes":15147},{"sha256":"772c9e187c0aa011ff9082997a1346b289eb1e762636ca72ac2041b79ef38560","name":"proof_audit.md","bytes":5885},{"sha256":"eb88ac63e1074b82d6ad7790f8d5ce210c24b5d12e310117950e03b1b12883b6","name":"certificate.py","bytes":7115},{"sha256":"20fc079edb85d0d0d2c735314e3b975340ae1c6f3c9aa0401ba9e6ba5b8973ff","name":"certificate_output.txt","bytes":1527},{"sha256":"19fc9bfe29187f91745a4ed00d3eb2c3ebc4f2b278bccc66c9cbefc88171f8c5","name":"prior_art.md","bytes":3873},{"sha256":"2787ff5b1b63b3363f04c149827aec87638746e2f347a435560825ef6c29ec7b","name":"next_step.md","bytes":2876},{"sha256":"e89b187eb775279b0883de2fad87a91821ebaa104582feae29688f5daeac9d92","name":"frontier.txt","bytes":2177},{"sha256":"856042ab0bbb1f13196267d5ff1c0773adb4d76c8f9222373a623b620f479488","name":"claim-sources.md","bytes":2959}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"f6ef1c67b7528aebc6c80bcc621590a5f1a973f9a112f51f02f9cdfa3285a009","next_step_sha256":null,"research_authority":{"witness_status":null,"research_status":"pending","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}