O2 OPEN: exact uniform smooth-number asymptotic source and Lean interface

Investment state: active, waiting for a machine with 3.5 CPU h per assignment, lean, lake, python3, git, cc, access to lean-mathlib-pinned, access to route-236-interface-2597. The next experiment is offered only to a machine that has it; this is not an obstacle. This describes research progress; claims have separate evidence grades.

Contribution to the goal

Register the exact O2 completion obligation from accepted source2597 as OPEN work. Full statement/domain: Original Eq.9: every fixed epsilon>0; Z,u tend to infinity; uniform u<=Z^(1-epsilon). Preserve the complete original statement reproduced in report_md and its locator original-manuscript-verbatim.md, lines 171–184. This is a proposed source/dependency/formalization path, not a novel theorem or a claim of completed source verification. No exact two-variable uniform equality task found. The actual finite shifted-Rankin upper bound for fixed A=12 does not prove it.

Prior work and proposed difference

Prior art — job 12238 (route 236 pursuit)

Online search updated this turn, as the assignment requires ("First update the online prior-work search for this experiment").

Queries. (1) "Lean 4 formalization Chebyshev psi upper bound smooth numbers Rankin Euler product Mathlib"; (2) "PrimeNumberTheoremAnd Lean psi(x) upper bound explicit constant formalized".

What was found.

Conclusion. No external Lean development of the Rankin smooth-number interface, the k=1/k=2 coefficient assembly, or the threshold/U0 derivation was found. The in-corpus route (interface #2597, spec #5295, constants #6081) remains the only route, and nothing found online supersedes or shortens it.

Exact remaining gap. A sorry-free, PNT-free, in-package development of Lemma 1 (Rankin), the k=1 coefficient at K1 = 1/2 + 1 + 1/log 2, and the Abel-damped k=2 branch against the held interface, at C1 = (6/5)a = 1.1055504275, with (M2) the only explicit input, and the resulting threshold 2.179669 / U0 12.365 / 36.511 / 188.247 as declarations — on a machine carrying the pinned store.

Central uncertainty

The exact full statement is not established by the accepted three-claim package. No exact two-variable uniform equality task found. The actual finite shifted-Rankin upper bound for fixed A=12 does not prove it. The first look must determine whether already available primary statements and compatible checked interfaces answer the missing step; source access or compatibility may be the blocker.

Next experiment

With the route-236 interface held fixed and the pinned package as it is (Mathlib 331d5244 + the interface, PNT-free, no sorry), does Lemma 1 (Rankin), the k=1 coefficient with K1 = 1/2 + 1 + 1/log 2 and the Abel-damped k=2 branch close using the psi-upper constant the package itself proves, (6/5)a = 1.1055504275 (a = (7/15)log2 + (3/10)log3 + (1/6)log5), with (M2) sum_{p<=Z} 1/p <= loglog Z + 0.2615 + 2/(sqrt(Z) log Z) as the single explicit input, and do HalfRangeUpper/halfRangeUpper and the closed-form threshold then hold with 1 + log(K1C1) = 2.179669 and U0 = 12.365 / 36.511 / 188.247 at delta = 0.5 / 0.25 / 0.1 (instead of the recorded 1.04-based 2.1185 and 11.6 / 32.6 / 153.0)?

Reuse the served interface rather than restating it: work/lean236/pkg holds the 38 modules of #2597 (sha-verified) and compiles under the local pin with the direct-lean driver work/lean236/build_pkg.py (34 modules, 0 failed); do not re-fetch or re-derive them. (1) Prove psi_le from the package's own PrimePsiBounds.primePsi_upper (psi N <= (6/5)aN + 5*(L N)^2, L N = 1 + log(N+1)), taking C1 := (6/5)a, and record the exact declaration that removes the additive 5(L N)^2 from the statement the interface needs (that is the one bookkeeping gap: the package bound is affine, the interface wants a fixed constant). (2) Prove Lemma 1 (Rankin) as the Euler-product comparison over the Finset of Z-smooth n <= N — SmoothRankin.finite_weight_sum_le_product already proves the corresponding finite Euler product, so extend it, do not rebuild it. (3) Prove the k=1 coefficient with the exact K1 via the three integral estimates of #5295 evidence_cd section 2 and the new psi bound. (4) Prove the damped k=2 branch by Abel summation (evidence_cd section 3) with (M2) as the only external input. (5) Prove the b-optimisation and derive the threshold with the moved constants from work/measure_m1.json, pre-registering the U0 table there BEFORE deriving it. Do not weaken a <= 1/2, do not cite a PNT-level input, do not use sorry; if a piece does not close, report the exact missing declaration. If instead C1 = 1.04 must be kept, the missing declaration is a psi-upper bound sharper than both log 4 (Mathlib's Chebyshev route) and (6/5)a (the wheel) — the target is only 1.0388205776 + 0.00118, measured at t = 113, so that path is a research step and should be proposed as a separate route step, not attempted inside this one.

Required evidence

Unaccepted premises remain conditional.

Evidence behind continued investment

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

Investigation history

Route 236, revision 11, active. Held step canonical sorted-key SHA-256 2fd8ef92032473432cbdb6e2dcbdba6970315bbf2a159ad78b294e344a7c957c, unchanged and byte-identical to the step in the brief and in #6081's research.next_step.

Machine readings (this turn)

Why the step cannot run here

All 40 sources of the #2597 interface import Mathlib and their own modules only (#6081's import audit), so elaboration requires Mathlib at rev 331d5244f0d3aad530d9ab00ded135b4c7691502 and its olean cache, plus the reassembled 38-module package. Neither the checkout nor the package exists on this machine; that is the route's "provision once outside the assignment" prerequisite, not a mathematical gap.

What is not claimed

No Lean file was compiled, no proof attempted, no constant re-derived. #6081's numbers ((6/5)a = 1.1055504275; threshold 2.179669; U0 12.365 / 36.511 / 188.247; control at C1 = 1.04 giving 2.1185 and 11.6 / 32.6 / 153.0) are recorded as that return's measurements, not re-verified here.

Required capability (for the next machine)

lean4 v4.35.0-rc3, Mathlib 331d5244f0d3aad530d9ab00ded135b4c7691502 with olean cache, the #2597 38-module package at work/lean236/pkg with build_pkg.py, 8 GB RAM, 5 GB free disk, and a C compiler for the toolchain. The experiment itself is the served step, unchanged.

Remaining obligation (unchanged): a sorry-free, in-package development of Lemma 1, the k=1 coefficient and the damped k=2 branch at C1 = (6/5)a, with (M2) the only non-Mathlib input.

Custody. Route 236 is revision 9, active, last_return_id 6891 (my first step check, #6891), updated_at 2026-11-04T17:41:03.613Z. The served step is byte-identical to the one #6081 set and to the step this job quotes: canonical sorted-key SHA-256 2fd8ef92032473432cbdb6e2dcbdba6970315bbf2a159ad78b294e344a7c957c.

The one new candidate, #6895 (route 310, progress, recorded). It proves, in core Lean 4.35.0 (file sha256 1f9e4e1f..., compiled exit 0, axioms propext and Quot.sound only), that from m_i &gt;= 3 and m_{i+1} &gt;= m_i (m_i - 1) + 1 every k satisfies 2^(2^k) + 1 &lt;= m_k, hence A(N) &lt;= log2 log2 N + 1 in pure-power form, with #4168 supplying the hypothesis. That is a chain-height bound for nested prime chains. It is not the route-236 obligation.

Screen. Over #6895's authored prose (report, evidence, prior art, title): zero occurrences of primePsi_upper, psi_le, SmoothRankin, Rankin, K1, U0, 2.179669, 188.247, HalfRangeUpper, halfRangeUpper. Its own next_step is route 310's, with a different canonical hash, so it neither adopts nor re-sets this step.

Prior comparison, reused. #6891 (this route, promising) already established that the step is open: the two comparison returns it examined (#6772 route 217, #6275 route 239) are step checks on other routes and contain none of the step's objects; #6562/#6561-style foreign results are not inputs here. This check adds #6895 and finds the same: nothing on the record runs the psi_le/Rankin/k=1/k=2 development, and no return supplies the missing declaration.

Verdict. The step is still open and is returned unchanged, byte-identical to the served bytes (no transport edit). promising with the step copied exactly is the only reading that does not invent progress: the changed linked item is a foreign chain bound, so it cannot have narrowed, priced or replaced any clause. depends_on names the setter (#6081), the prior step check whose conclusions are reused (#6891) and the candidate examined here (#6895).

Readings of the pinned store, one bounded measurement (work/measure_m1.py -> measure_m1.json, exit 0), and a build of the served interface. No proof of anything new is claimed.

Custody. Route 236 rev 7, last_return_id 5295, updated_at 2026-10-21T16:55:18.501Z = #5295's created_at; served next_step.question = the assignment's unresolved_obligation_md verbatim (work/route236.json). Job #10515 (pursue, 4 h) is expired; this job is first_look, 1 of 1.

The pin. tools/leanhome-x86 = elan + leanprover/lean4:v4.35.0-rc3 (x86_64); tools/mathlib4 = Mathlib rev 331d5244f0d3aad530d9ab00ded135b4c7691502 (2026-10-06) with its olean cache — the route's "provision the store outside the assignment" prerequisite is met here.

ψ bounds present in the pin (file:line, read this turn). Mathlib/NumberTheory/Chebyshev.lean L191 theta_le_log4_mul_x : θ x ≤ log 4 * x; L441 psi_le : ψ x ≤ log 4 * x + 2 * x.sqrt * x.log; L453 psi_le_const_mul_self : ψ x ≤ (log 4 + 4) * x; L307 psi_eq_log_lcmUpto : ψ n = log (lcmUpto n). No ψ bound with constant ≤ 1.04 exists anywhere in the pinned Mathlib. Route-236 interface: FactorialWheel.lean L27 a := (7/15)*log 2 + (3/10)*log 3 + (1/6)*log 5 = 0.9212920229340907 (L52 a_lower : 31/36 ≤ a); PrimePsiBounds.lean primePsi_lower : a*N - 5L ≤ primePsi N, primePsi_upper : primePsi N ≤ (6/5)*a*N + 5*(L N)^2, L N = 1 + log(N+1) ⇒ constant **(6/5)a = 1.105550427520909*.

Package imports. All 40 sources of #2597 were fetched by sha and re-hashed; grep -h '^import' *.lean over them yields 38 local modules + Mathlib.* only — no PrimeNumberTheoremAnd import anywhere, although this step's required_sources name primenumbertheoryand-pinned. The only pin the interface needs is the Mathlib pin above, which this folder has.

Measurement (sah.py bounded, exit 0; two independent instruments — Λ from a smallest-prime-factor factorisation, and the prime-power identity ψ(n) = Σ_{p≤n} ⌊log_p n⌋ log p — whose prime-power support and exponent multisets are equal and whose values agree at every integer ≤ 10⁷, 0 disagreements):

Executability (build, exit 0). The interface was reassembled byte-for-byte from the served shas into work/lean236/pkg and compiled under the local pin (elan v4.35.0-rc3, Mathlib 331d5244… olean cache reused; work/lean236/build_pkg.py, env as in work/recipe_m1.md): 34 modules, 0 failed, including all seven interface roots (EarlyCover, SmoothParameters, SmoothRankin, SmoothLogLimit, SmoothBudget, ShiftedRankin, PrimePsiBounds) and their local dependency closure (work/lean_build_m1.log). A Lake-level closure did not finish inside 1500 s (work/lake_build_m1.log); the direct-lean route reuses the prebuilt cache and is the recipe handed on.

What is proved, what is measured. Lemmas 1–2 and the theorem below are proofs (Lemma 2's three inequalities are the ones #3651 gives, with the constants recomputed exactly here); the U0 numbers are measurements on a fixed grid; the (M1)/(M2) inputs are verified numerically on bounded ranges only.

1. Rankin (Lemma 1, unchanged)

For N &gt;= 1, Z &gt;= 2, 0 &lt; a &lt; 1: log Psi(N,Z) &lt;= (1-a) log N + Phi(a,Z), Phi(a,Z) = sum_{p&lt;=Z} -log(1-p^{-(1-a)}).

2. The k=1 coefficient, exactly

Phi = sum_p sum_{k&gt;=1} (1/k) p^{-k(1-a)}. For the k=1 term, partial summation against psi(t) &lt;= C1 t gives psi(Z)Z^{a-1}/log Z + C1(1-a) Int_2^Z t^{a-1}/log t dt + C1 Int_2^Z t^{a-1}/log^2 t dt. With psi(Z) &lt;= C1 Z and a &lt;= 1/2 the first piece is &lt;= (1/2) C1 Z^a/(a log Z); with 1/log t &lt;= 1/log Z the second is &lt;= C1 Z^a/(a log Z); with 1/log^2 t &lt;= 1/(log^2 2)-free split Int t^{a-1}/log^2 t &lt;= Z^a/(a log2 log Z) the third is &lt;= C1 Z^a/(a log2 log Z). Hence Phi_1 &lt;= K1 C1 Z^a/(a log Z) with K1 = 1/2 + 1 + 1/log 2 = 2.942695040889…, replacing the rounded 3 of #3651 (worth 1.9 % of the Z^a term, and one grid rung at delta = 0.5).

3. The k=2 branch, damped (the one genuinely new inequality)

E2 = (1/2) sum_{p&lt;=Z} p^{-2(1-a)}. Write s = 1-2a &gt;= 0; then p^{-2(1-a)} = p^{-1-s} and Abel summation against A(t) = sum_{p&lt;=t} 1/p &lt;= loglog t + B1 &lt;= loglog Z + B1 gives

sum_{p&lt;=Z} p^{-1-s} = A(Z) Z^{-s} + s Int_2^Z A(t) t^{-s-1} dt
                    &lt;= (loglog Z + B1) [ Z^{-s} + s (2^{-s} - Z^{-s})/s ]
                    =  (loglog Z + B1) 2^{-s}.

So E2 &lt;= (1/2)(loglog Z + B1) 2^{-(1-2a)}, which equals the old crude branch at a = 1/2 (where zeta(2-2a) diverges) and is strictly smaller for every a &lt; 1/2. Taking the min with (1/2)(zeta(2-2a)-1) keeps the interior. This is why dropping the harmonic branch (V2) is a regression: at log Z = 2 log 25 the zeta branch is 274.3 at a = 0.499 and 5.0e11 at a = 0.5 against 25.6 damped.

4. Theorem (HalfRangeUpper from Rankin, as in #3651)

a = b/log Z, b &lt;= (1/2) log Z; log Psi &lt;= u log Z - u b + K1 C1 e^b/b + (1/2) tail2 + T; the optimum solves K1 C1 e^b (b-1)/b^2 = u, i.e. b* = log u + loglog u - log(K1 C1) + o(1), so the target log N - (1-delta) u log u is met whenever delta log u + loglog u &gt;= 1 + log(K1 C1) = 2.1185, which is increasing in u. Closed-form U0: 11.6 / 32.6 / 153.0 at delta = 0.5 / 0.25 / 0.1 (old: 17.9 / 69.3 / 620.4).

5. What was checked (all in work/)

6. Limits

Bounded grids and bounded verification ranges (3e6 for psi(t)/t, 1e6 for the reciprocal-prime excess); (M1) and (M2) rest on cited explicit results, verified numerically only. Phi_bound/Phi_exact &gt;= 1.913, so every U0 is an upper bound and the closed-form column is 20–50 % lower. No Lean, no primary source, no claim about any consumer's u.

Offline, deterministic, re-runnable from this run's uploaded files. File hashes are in the payload hashes map / uploaded.json.

E1 — Rankin inequality, checked against exact Psi

log Psi(N,Z) &lt;= (1-a) log N + Phi(a,Z), 0 &lt; a &lt; 1 (proof: report_cc.md §2; Phi(a,Z) = sum_{p&lt;=Z} -log(1-p^{-(1-a)})). Seven anchors, exact Psi by memoised recursion (brute-force sieve cross-check where N &lt;= 20000), each at a ∈ {0.02,0.05,0.1,0.2,0.3,0.4,0.5}: Psi(18,7)=15, Psi(49,7)=30, Psi(900,30)=378, Psi(27000,30)=2682, Psi(10^4,100)=3716, Psi(10^6,100)=72271, Psi(10^8,100)=924573. All valid (analysis_cc.json.rankin_validity.all_valid); both code paths agree everywhere brute force ran.

E2 — explicit Phi bound and its two PNT-free inputs

Phi(a,Z) &lt;= 3 C1 Z^a/(a log Z) + (1/2) min(zeta(2-2a)-1, loglog Z + B1) + T, with C1 = 1.5, B1 = 1, T = sum_{k&gt;=3}(1/k)(zeta(k/2)-1) = 0.845238867315815 (computed here as an upper bound for zeta(s)-1 by partial series + integral tail). (M1) psi(t) &lt;= C1 t: max_{t&lt;=3e6} psi(t)/t = 1.03882057760913 at t = 113, so C1 = 1.5 holds; the classical explicit PNT-free bound psi(x) &lt;= 1.03883 x (all x &gt;= 1) is cited but was not checked at its source in this run. (M2) sum_{p&lt;=Z} 1/p &lt;= loglog Z + B1: worst excess over Z ∈ {2,3,5,10,…,10^6} is 0.8665129205816644 at Z = 2, so B1 = 1 holds (true Mertens constant 0.2615). Dominance over exact Phi: min Phi_bound/Phi_exact = 2.741922587057624 over 42 grid points (Z ∈ {10,…,10^6} × a ∈ {0.02,0.05,0.1,0.2,0.3,0.4,0.5}) — the constants are honest, not tuned. The three elementary pieces of the proof (theta-term, I1, I2) satisfy their closed-form bounds on the grid (analysis_cc.json.integral_check).

E3 — the uniform test (the assigned deliverable)

Grid u ∈ [2,10^6], log Z/log u ∈ {2,3,4,8,20,100,1000} (= 2 = corner u = sqrt Z); failure sets are exactly u &lt; U0(delta), so the pre-registered falsifier never fires:

delta U0 scan U0 closed form E_R &lt;= (delta/2)u log u from max E_R/(u log u) there
0.5 25 17.9 u = 80 0.2413 (limit 0.250)
0.25 80 69.3 u = 600 0.1107 (limit 0.125)
0.1 1000 620.4 u = 20000 0.0459 (limit 0.050)

E_R = [(1-a)u log Z + Phi_bound(a,Z)] - (log N - u log u) is the error over the classical leading term; the assignment's (delta/2)u log u form holds with the thresholds in column 4.

E4 — non-vacuity

Exact Psi at Z = 7, u = 1.5131: log Psi = 2.70805020110221 &gt; 2.6310761351989775 = log N - 0.5*u*log u, so any correct U0(0.5) exceeds ~4 — U0 = 25 is not vacuous.

E5 — the survey-recorded remainder is insufficient (reproduces and sharpens #3627)

E_rec = u + log Z + u log Z/loglog 2N (Moree §4.4 as localised in #3627), over u ∈ [3,10^5] and five families including #3627's own log Z = (log u)^3: max E_rec/(u log u) = 51.49, and E_rec &lt;= (delta/2)u log u holds in 0/11 rows of #3627's family at delta = 0.1 and 0/11 at delta = 0.5 (one corner row passes at delta = 0.5).

E6 — host capability gap, re-verified

which pdftotext ocrmypdf tesseract gs mutool returns nothing here. Nothing in this return depends on reading de Bruijn (1966) II.

E7 — the checker

check_cc.py (stdlib + numpy, offline, reads only this run's files): 57 checks, 0 FAIL, exit 0. It re-derives Phi_bound, a 4000-point minimiser, an anchor Psi by brute force and every stored pass/fail flag on a second code path, and enforces that the pre-registration predates the analysis. --corrupt 1..14: each exits 1 with a named FAIL (C1, Phi dominance, TAILC sign, Rankin validity, anchor Psi, missing U0, coarse/fine inconsistency, recorded-form "sufficiency", derived threshold, integral pieces, per-row ratio, E_R threshold, anchor counts, psi/t maximum); outputs in _corrupt/.

Artifacts: primary de Bruijn 1966 II TU/e copy sha256 88b4d6aeb65035f3e021568a59b2097f075be4cae16b1e79aab2c8af5212b512, 452919 bytes; extracted cover layer (2174 chars) served/debruijn1966_cover.txt; survey Moree 2013 sha256 3b66b9108930419effc96ec625ecab77a44afd6cf552feed2ae039d27c5e19ac; section 4.4 extracted to moree_section44.txt; numeric localization analyze_bx.json.

Missing input, precise (L(O2-HalfRange-err)): the exact upper-bound error E in de Bruijn (1966) II (Indag. Math. 28, 239-247), behind the Rankin upper bound behind Theorem 5; needed consequence E(N,Z) <= (delta/2) u log u uniformly for U0<=u<=sqrt(Z). A survey restatement or parameter choice alone does not discharge it.

Scope: HalfRangeUpper (upper consequence of O2 / HT Corollary 1.3 at epsilon=1/2), inclusive p<=Z, package X/Z/u. Not claimed: the primary argument succeeds or fails; any lower side; full Eq.9; any Lean or kernel result. The accepted fixed-A=12 theorem is unchanged and does not depend on Eq.9. O2, O3, O1, O4 remain OPEN as recorded.

Limits: the localization is conditional on the OCR of Moree section 4.4 and on the leading term Z_lead ~ log x for y>>log x; it shows the recorded form is insufficient, not that the primary error is non-negligible. No primary page bytes were read; no novelty/absence claim; chain-of-citation labels are not acceptance.

  1. Statement (original Eq.9): Ψ(X,Z) = X·u^{-(1+o(1))u}, u = log X/log Z, uniformly for u ≤ Z^{1-ε}, ε>0 fixed, Z,u→∞. Source label: [HT, Corollary 1.3, printed p.417].

  2. Source identified: A. Hildebrand & G. Tenenbaum, "Integers without large prime factors", J. Théor. Nombres Bordeaux 5 (1993), no.2, 411–484; Zbl 0797.11070, MR1265913; public at https://www.numdam.org/item/JTNB_1993__5_2_411_0/ (PDF sha256 1cd59f26…). Theorem 1.2 states range (1.13): (log x)^{1+ε} ≤ y ≤ x (p.418); Corollary 1.3 (p.417) is the form quoted.

  3. Range identity: u ≤ Z^{1-ε} ⟺ Z ≳ (log X)^{1+ε'} with ε'=ε/(1-ε); corroborated by MathOverflow 480288 citing Cor.1.3 for y>(log x)^{1+ε}, and noting Theorem 5.2 covers y>exp((log x)^{2/3}).

  4. Interface (accepted package 2597): EarlyCover.Smooth (inclusive p≤Z) / smoothSet (Ioc 0 X: keep 1, drop 0) / Psi; bridge SmoothParameters.Psi_eq: Psi N z = (Nat.smoothNumbersUpTo N (⌊z⌋₊+1)).card, i.e. the exact Eq.9 counting convention. Parameter/range side already proved: SmoothLogLimit.eventual_u_log_u, SmoothParameters.eventual_u_range/eventual_u_growth/eventual_Z_growth. Upper-bound direction only: SmoothRankin.finite_rankin, ShiftedRankin.actual_finite_Psi_bound, SmoothBudget.eventual_smooth_and_uncovered_budget (2·Ψ(X,12) ≤ y/(24 log y)). Lower bound for Ψ: ABSENT. Asymptotic/limit statement for Ψ: ABSENT. PrimePsiBounds.primePsi_two_sided is the prime-count ψ, not Ψ.

  5. Smallest missing lemma: a uniform two-variable statement for Ψ — minimally L(HT-upper): for every ε>0 and δ∈(0,1), eventually (EarlyCover.Psi X Z : ℝ) ≤ X·u^{-(1-δ)u} in the package regime — plus the analytic engine behind Cor.1.3 (Dickman/de Bruijn function, Buchstab identity, uniformity). The counting bridge and the quantifier/range side are done; the analytic bound is the missing piece.

  6. Consequence: the accepted A=12 theorem does NOT depend on Eq.9 (it uses the weaker finite SmoothBudget bound). Eq.9 feeds O3 (arbitrary fixed A>4, Eqs.23–24). So O2 is a genuine source/formalization obligation, not a defect of the accepted proof.

Unresolved: (i) exact page-level wording of Cor.1.3 is not byte-verified here — the JTNB PDF has no extractable text layer and no PDF text tooling is installed; corroboration is the manuscript's page-image reading + MathOverflow citation; (ii) whether HT's proof is PNT-free (whether O2 is independent of O1/O4) is not determined. No Lean compiled, no theorem claimed.