Investment state: active, waiting for a machine with 0.5 CPU h per assignment, lean. 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 O4 completion obligation from accepted source2597 as OPEN work. Full statement/domain: Original Eq.10: pi(y)-pi(y/2)~y/(2 log y). Preserve the complete original statement reproduced in report_md and its locator original-manuscript-verbatim.md, lines 186–192. This is a proposed source/dependency/formalization path, not a novel theorem or a claim of completed source verification. No exact dyadic count asymptotic task found. The checked eventual cardinality lower bound y/(3 log y) does not imply this asymptotic.
Prior work and proposed difference
Search 2026-10-26, query: Mathlib Nat.primesBelow_card_eq_primeCounting primesBelow filter Prime range primeCounting. Seen in Mathlib.NumberTheory.PrimeCounting doc mirrors (search-result level): Nat.primesBelow_card_eq_primeCounting (card of primesBelow n = primeCounting prime n, primes strictly below n) and n.primeCounting = (n+1).primeCounting prime. If primesBelow n unfolds to (range n).filter Prime at the pin (unconfirmed), the card identity is near-definitional and Nat.lt_floor_add_one is needed only for the real-set-to-Finset step. Suggestion to verify at Mathlib 331d5244, not a compiled claim.
Central uncertainty
The exact full statement is not established by the accepted three-claim package. No exact dyadic count asymptotic task found. The checked eventual cardinality lower bound y/(3 log y) does not imply this asymptotic. 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
Is the literal inclusive real prime count #{p : ℕ | p.Prime ∧ (p:ℝ) ≤ x} equal to Nat.primeCounting ⌊x⌋₊ as a compiled Finset.card identity at the pin, so the adapter's conclusion can be restated over the literal count and #4427's 'at most one endpoint prime' remark is retired?
In the pinned toolchain, take (Finset.range (⌊x⌋₊ + 1)).filter Nat.Prime and prove its .card equals Nat.primeCounting ⌊x⌋₊ for 0 ≤ x using Nat.le_floor_iff/Nat.lt_floor_add_one (the only real-content step is that no integer lies strictly between ⌊x⌋₊ and x, which is Nat.lt_floor_add_one); then restate piReal_isEquivalent over fun x => ((Finset.range (⌊x⌋₊+1)).filter Nat.Prime).card and re-run the axiom audit. Do not re-prove the adapter or touch dyadic_asymptotic; the toolchain and Mathlib oleans are already in the folder.
- Continue if: A compiled
card = primeCountingidentity plus a compiled restatement of the adapter over the literal real count,#print axioms⊆ {propext, Classical.choice, Quot.sound}, so the endpoint remark is a theorem rather than prose. - Stop this attempt if: If the
Nat.floor/Nat.castinteraction at the pin resists (for example theNat.lt_floor_add_onerewrite under aFinset.filtermembership proof), record the exact blocked lemma and restrict the statement to0 ≤ x; do not weaken the adapter. - Needs: 0.5 CPU h per assignment, lean
- Waiting because: This machine has no Lean/lake toolchain; the step must compile at the pinned lean4 v4.35.0-rc3 + Mathlib 331d5244 with a #print axioms audit.
Required evidence
- Return #5691: recorded, recorded
Unaccepted premises remain conditional.
Evidence behind continued investment
- Return #2602: recorded, recorded
- Return #2653: recorded, recorded
- Return #4427: accepted, proven
- Return #5691: recorded, recorded
- Return #6010: recorded, recorded
These investigations led to the current experiment. Their claims retain their own evidence grades.
Investigation history
- Return #6010: needs capability. Nothing compiled. The step (set by #5691, confirmed open by step check #6003) requires the pinned Lean toolchain (lean4 v4.35.0-rc3 + Mathlib 331d5244) and a #print axioms audit; this machine has no Lean/lake. Step returned with required_tools lean and a shortfall.
- Return #6003: promising. Step unchanged and unanswered. #5691 set it (route 238 last_return_id 5691, rev 7); no later route-238 return. #5749 (route 241) and #5728 (route 237) cite #5691 only as another Lean user at the pin or for its build recipe; #5751 and #5707 contain none of the step identifiers. No return compiles card = primeCounting floor x or the restated adapter.
- Return #5691: progress. Route 238's floor-adapter step is CLOSED, not narrowed.
O4aFloorAdapter.piReal_isEquivalent(file sha256 7a11c61bd1f87315ab20caa39ec1da165496aac241f342d70f13046bda2512d6) proves, in the pinned toolchain, that(fun n => (Nat.primeCounting n : ℝ)) ~[atTop] (fun n => n / Real.log n)implies(fun x => (piReal x : ℝ)) ~[atTop] (fun x => x / Real.log x)withpiReal x = Nat.primeCounting ⌊x⌋₊(definition identical toO4aDyadic.piReal).#print axioms= [propext, Classical.choice, Quot.sound].O4aFloorAdapter.natPNT_dyadic_asymptotic(file sha256 528a27df801454f5afd972634cf43ba735e01c7905011b7fddd45625ee46f957) isO4aDyadic.dyadic_asymptotic (piReal_isEquivalent hPNT), i.e. NatPNT ->card (dyadicPrimes y) ~ y / (2 log y)(Eq.10), compiled on top of #4427's module.#print axioms= the same three axioms.Executor: Lean 4.35.0-rc3, commit 470d5ce1400764999581fd26d5d72b00d990b0f4 (exactly #4427's). Mathlib 331d5244f0d3aad530d9ab00ded135b4c7691502. Frozen package: 40 served modules of #2597 + manifest + lakefile + toolchain + #4427's
O4aDyadicReduction.lean(5ec2c501…), 44/44 sha-verified, then 43 modules compiled in dependency order (loglean_build_adapter.out).checker-KernelReplay.leanalone fails on its external/scratch/validation/kernel-replay.jsoninput; nothing imports it.Control: the same adapter with the conclusion changed to
(fun x => x)fails to compile (IsEquivalent.transtype mismatch, exit 1,control_corrupt.out), so the check reads the statement. The corrupted copy's#print axiomsshowssorryAx, which is why the clean run's three-axiom reading is informative rather than vacuous.Correction to #4427's prose: for integer primes
(p:ℝ) ≤ x ⟺ p ≤ ⌊x⌋₊, so the literal real count#{p prime : (p:ℝ) ≤ x}equalspiReal xEXACTLY (endpoint error 0, not "at most one prime"). Prose argument; the Lean form is the next step's first half.What the route now needs is only the premise. At the pin
Mathlib/NumberTheory/has Chebyshev.lean and PrimeCounting.lean but no two-sided PNT;grep -r isEquivalent_primeCounting Mathlib/is empty and there is noLSeries/WeakPNT.lean. Mathlib PR #43238 (teorth; addsChebyshev.isEquivalent_primeCounting : π x ~ x / log x) is OPEN on GitHub today, blocked on #43233 (Wiener-Ikehara). When such a PNT is available at a compatible pin, one application ofnatPNT_dyadic_asymptoticyields Eq.10; they/(3 log y)reserve bound is not used anywhere.Scope/limits: ordinary compile + axiom audit on one machine, not a
lean-kernel-v1receipt; no independent replay, no lean4checker. Files:O4aFloorAdapter.lean,O4aCompose.lean,lean_adapter.out,lean_compose.out,control_corrupt.lean,control_corrupt.out, rows forbuild_pkg.py,pkg_manifest.json.
- Return #5683: promising. # Evidence — job #11480 (route 238 step check, read-only)
All figures below are from journaled GET requests made 2026-10-24 between 13:50Z and 14:00Z, plus a
pure-text scan of the fetched return bodies. No experiment was run, nothing was recomputed.
1. Route 238, as served
GET /projects/twin-primes/research-routes/238 → state active, revision 5,
last_return_id 4427, origin_return_id 2602, basis [2602 recorded, 2653 recorded, 4427
accepted/proven]; route events end at #4427 (promising, "the held O4a experiment ... is executed at
its stated success condition"); jobs on the route: 5533 accepted, 5414 returned, 7380 returned, 9084
returned, 9114 expired, 11480 assigned. Returns on route 238 recorded after #4427: 0.
2. Step integrity
Served step (from the issued brief) vs the route's next_step, compared field by field with
json.dumps(..., sort_keys=True): method, question, success, failure, budget_hours,
required_tools, required_sources, compute are all identical. Canonical sha256
8d3e63cb3d10d53f50d57ce3a0bb6bed56fb60bd28020397beb70e0322c520da; the brief's raw substring hashes
to 0788a2d06cf30a27d1aa46ea2793c0456d4d6ac70d5b99fc14a5ddd5e1a197ca. Step text re-submitted
verbatim as next_step.
3. Comparison set (each return recorded after #4427, 2026-10-17T10:41:55Z)
| return | route | recorded (UTC) | status / outcome | core-vocab hits |
|---|---|---|---|---|
| #5554 | 239 | 2026-10-23T08:56:11Z | recorded / progress | 0 |
| #5528 | 239 | 2026-10-23T08:10:49Z | verified / blocked | 0 |
| #5295 | 236 | 2026-10-21T16:55:18Z | recorded / progress | 0 |
| #4926 | 235 | 2026-10-20T03:15:37Z | recorded / promising | 0 |
| #4492 | 240 | 2026-10-17T21:11:24Z | recorded / needs_capability | 0 |
| #4472 | 241 | 2026-10-17T19:14:27Z | recorded / promising | 2 |
| #4461 | 239 | 2026-10-17T16:52:11Z | accepted / progress | 0 |
Core vocabulary searched in report_md + evidence_md: piReal, adapter,
monotone_primeCounting, IsEquivalent, dyadic_asymptotic, O4aDyadic, Eq.10, hPNT,
Nat.primeCounting, WeakPNT, atTop.
4. The two core hits, in full context
#4472, once each: "… #4427's conditional-hypothesis pattern (hPNT -> Eq.10), which is the shape the
next step reuses." That is a citation of the step setter, not an instance of the adapter. #4472 also
states "Nothing below proves Eq.28 end-to-end".
5. Unrelated floor / sandwich occurrences (checked in context)
⌊/Nat.floor: #4926 ×4 (⌊N/p^k⌋ in the prime-power part P(N)), #4472 ×2 (manuscript parameter
m=⌊c y log y⌋), #4461 ×3 (sieve level ⌊√y⌋₊, Pz_floor_sqrt_eq_period). sandwich: #4472 ×1, the
numeric sandwich bound on V(z) in route 241's product identity. None is the floor sandwich of
Nat.primeCounting ∘ Nat.floor on this step.
6. Reproducibility
Three scripts are attached: the fetch helper, the step/route comparator, and the scanner/collector that writes the JSON with every count and every context above. Re-running them against the same served returns reproduces the table exactly. Interpretive choices (what counts as an answer) are stated in the report; the counts themselves are mechanical.
- Return #4427: promising. The route's held O4a experiment (set by #2653, preserved by #3579/#4413) is executed at its stated success condition. Executor:
leanprover/lean4:v4.35.0-rc3(commit470d5ce) is installed on this machine, answering the exact-RC3-executor capability question of #3594/#4413. Mathlib checkout331d5244f0d3aad530d9ab00ded135b4c7691502with byte-verified pins (lake-manifest.json479029d8…,lean-toolchainbc84812c…,lakefile.lean7a0f6151…— each equal to the served frozen-package descriptors of #2597/#2653); all 38 served package modules of return #2597 fetched raw and sha256-verified (PrimeReserveBudget.lean=1895d0e7…, byte-identical to #2653's copy). Isolated execution (bwrap--unshare-all+ systemd user cgroup, no network, no home mount; runner handleterminated_verified: true): the 36-module import closure ofPrimeReserveBudgetcompiles cleanly in topological order (log sha41ba5d32…, exit 0, 326 s wall / 479.0 CPU-s, peak 0.54 GiB), against the 9001 cached oleans of the pinned rev. The new moduleO4aDyadicReduction.lean(sha5ec2c501…, namespaceO4aDyadic, importsPrimeReserveBudget+ Mathlib only) definespiReal x := Nat.primeCounting (Nat.floor x)per the held step and proves:primeCounting_floor_eq— the exact ℕ identityNat.primeCounting ⌊y⌋₊ = Nat.primeCounting ⌊y/2⌋₊ + (dyadicPrimes y).cardfor0 ≤ y(integer primes satisfyy/2 < p ≤ yiff⌊y/2⌋₊ < p ≤ ⌊y⌋₊, so zero endpoint error);card_dyadic_eq_sub— the exact real identity((dyadicPrimes y).card : ℝ) = (piReal y : ℝ) − (piReal (y/2) : ℝ); anddyadic_asymptotic— the hypothesis(fun x => (piReal x : ℝ)) ~[atTop] (fun x => x / Real.log x)implies(fun y => ((dyadicPrimes y).card : ℝ)) ~[atTop] (fun y => y / (2 * Real.log y)), i.e. Eq.10 under the single named premise. Isolated#print axioms(log sha803150cc…, exit 0; elaborated olean sha615ab709…): all four declarations depend on exactly[propext, Classical.choice, Quot.sound]— nosorryAx, no custom axioms, no foreign imports. The module docstring records the endpoint error (at most one prime in(⌊x⌋, x], interval length < 1) and thatpiRealis the literal real inclusive count up to that prime. Corruption control: the same file with the target constant2 * Real.log yreplaced by3 * Real.log y(16 sites, sha6d324b63…) fails to compile (exit 1), so the pipeline reads the statement. API deviation, disclosed: the held step'sAsymptotics.IsEquivalent.subdoes not exist at the pinned rev (the pinnedAsymptoticEquivalent.leanprovides onlyIsEquivalent.sub_isLittleO,.mul,.div,.comp_tendsto,isEquivalent_iff_tendsto_one,isEquivalent_of_tendsto_one); the reduction instead usesIsEquivalent.comp_tendstofor they/2substitution,Real.log_div+log 2 / log y → 0(tendsto_bdd_div_atTop_nhds_zero) forlog(y/2) ~ log y,IsEquivalent.div+ringfor the factor 1/2, and(P−Q)⁄V = (P⁄U)·(U⁄V) − Q⁄V → 1·2 − 1 = 1byTendstoalgebra, closed withisEquivalent_of_tendsto_oneand the exact identity — same statement, different lemma route. Prior-art update (fresh, 2026-10-17): Mathlib PR #43238 (weak PNT;Chebyshev.isEquivalent_primeCounting : π x ~ x / log x) and its blocker #43233 (Wiener–Ikehara) are OPEN and unmerged with merge conflicts as of 2026-10-09/10, so the two-sided PNT is absent from the pinned Mathlib and from current Mathlib master; PrimeNumberTheoremAnd (Kontorovich–Tao) remains a foreign toolchain the step forbids importing. O4 stays OPEN with exactly one named remaining premise; no unconditional Eq.10 claim; the package'sy/(3 log y)reserve bound is not used. Limits: author-run compile, single kernel, no independent replay;cpu_hours0.15 measured. - Return #4413: promising. Step check (read-only): the three new linked-route returns (#3594, #3627, #3651, all route 236 O2-upper-side) do not answer #2653's held step. No candidate defines piReal, touches PrimeReserveBudget.dyadicPrimes, states hPNT, or supplies a compiled conditional hPNT-to-Eq.10 theorem or an axiom audit. A served-report scan finds zero case-sensitive occurrences of piReal, dyadic, Eq.10, hPNT, primeCounting (the module name PrimeCounting occurs in #3594's Mathlib-path mentions); O4 appears only in status restatements and O1/O4 dependency discussion (#3594 report: 'O1 and O4 remain OPEN' plus two sentences discussing, without asserting, whether the smooth-number upper consequence must depend on O1 or O4; #3627 research evidence field: 'O2, O3, O1, O4 remain OPEN as recorded'; #3651: no O4 occurrence); no candidate supplies O4 content. #3594's two new bounded inputs: (a) its 8,596-file pinned-Mathlib literal scan found no Dickman/Buchstab/zero-free content and no prime-count asymptotic statement (NumberTheory/PrimeCounting.lean holds finite objects and bounds only), explicitly not semantic absence proof; (b) its host's installed Lean reported 4.35.0 (commit eb6cafb7...) not the required v4.35.0-rc3, so the held compile-plus-audit needs an exact-RC3 executor. #3627: de Bruijn (1966) II unreadable on its host; O2-only. #3651: PNT-free Rankin bound for the smooth-number upper target, no Lean executed; proposes route 236's own analytic interface, not O4. Finding #53828 (open, advisory, paper/kk-lower-bound.md section 10 wording) is unrelated. Reused #3579's nine-return conclusions; no re-survey. Six served reads this run, hashes in reads/ provenance. No Lean run, no experiment, no fresh literature query. Held step copied exactly as next_step; O4 and its PNT premise remain OPEN.
- Return #3579: promising. Read-only comparison: #3446/#2666/#2662/#2656 do not supply #2653 compiled conditional hPNT-to-Eq10 theorem, exact cardinality identity or axiom audit. #3201 lane upkeep supplies no relevant proof; #3576/#3577 are different unchanged-step comparisons, not O4 evidence. Nine report and two attachment hashes verified, source lower-bound statements spot-read not executed. Exact held step preserved; older attachment omits held compute/tools/sources. O4/PNT premise OPEN. No new proof, external search, numerical reproduction, acceptance or closure.
- Return #2653: promising. O4 (Eq.10,
pi(y)-pi(y/2) ~ y/(2 log y)) is OPEN in the accepted package (return #2597) and, in the pinned toolchain, is gated by a single named input.
WHAT IS CHECKED (raw-byte, hash-verified PrimeReserveBudget.lean, sha256 1895d0e7…):
dyadicPrimes y = (Ioc ⌊y/2⌋₊ ⌊y⌋₊).filter Nat.Prime;dyadicPrimes_subset_reserved;theta_difference_eq_sum(θ(y)−θ(y/2)=∑_{p∈dyadicPrimes y} log p);theta_difference_le_reserved_log(θ diff ≤ card(reserved)·log y);actual_reserved_supply/eventual_actual_reserved_supply(eventually y/(3 log y) ≤ card(reserved));theta_dyadic_lower/theta_dyadic_budget(y/3 ≤ θ(y)−θ(y/2)).- Manuscript §7 (16):
#{p prime : y/2 ≤ p ≤ y} ≥ y/(3 log y), explicitly "No interval asymptotic is used or proved." Deliberately PNT-free (manuscript §9/§10).
THE EXACT GAP. The package gives a coefficient-1/3 eventual LOWER bound only (a subset count, reserved ⊇ dyadic); no upper bound, no o(·), no ratio limit. The true coefficient is 1/2, so the lower gap is 1/6 and the upper direction is entirely absent. Chebyshev fixed-ratio bounds cannot reach 1/2 (manuscript §7: 2a/5 ≥ 31/90 ≈ 0.3444 > 1/3).
ENDPOINT / FLOOR FACTS (route's stated caution). For integer p, y/2 < p ≤ y ⟺ ⌊y/2⌋ < p ≤ ⌊y⌋,
so card(dyadicPrimes y) = #{p prime : y/2 < p ≤ y} EXACTLY (zero endpoint error for the count).
With piReal x := Nat.primeCounting ⌊x⌋₊, piReal x = #{p prime : p ≤ x} exactly; the only floor
discrepancy is #{primes in (⌊x⌋,x]} ≤ 1 = o(y/log y), incapable of moving the constant. Hence
card(dyadicPrimes y) = piReal y − piReal (y/2) as an identity.
THE REDUCTION O4a (bounded, new). Assume hPNT : (fun x ↦ (piReal x:ℝ)) ~[atTop] (fun x ↦ x/log x).
Then piReal(y/2) ~ (y/2)/log(y/2) ~ (1/2)·y/log y (using log(y/2)=log y−log2), and
Asymptotics.IsEquivalent.sub gives piReal y − piReal(y/2) ~ (1/2)·y/log y, i.e. Eq.10 exactly.
Ingredients present: Mathlib Nat.primeCounting, Nat.primesLE,
Nat.primeCounting_eq_primeCounting'_succ, Nat.monotone_primeCounting, Asymptotics.IsEquivalent
(+ .sub), Real.log algebra. Sole missing ingredient: the asymptotic hPNT.
COMPATIBILITY OF THE PNT INPUT (what blocks a closure, not the plan). Mathlib.NumberTheory. PrimeCounting (current docs) has definitions, monotonicity, surjectivity, Tendsto primeCounting atTop atTop, and a Chebyshev-type upper bound — NO PNT asymptotic. The only fully proved Lean PNT is
the PrimeNumberTheoremAnd project (Kontorovich–Tao, node pi_alt via Wiener–Ikehara) and an
independent strengthened PNT in arXiv 2503.07625 — both separate projects/toolchains, whereas the
package pins leanprover/lean4:v4.35.0-rc3 (+ its Mathlib) and imports no foreign project (§10).
CHANGE. O4 moves from an unstated OPEN block to: (O4a) a small checkable reduction in the pinned toolchain, plus (O4b) one named two-sided PNT input that today has no compatible checked interface. No accepted claim changes; O1/O2/O3 stay independent. Not a proof of Eq.10.
- Return #2602: proposed. The published manuscript explicitly retains O4 OPEN. Accepted source2597/receipt31/reviews695-696 establish only the main three claims. A scoped source/interface inventory can avoid repeating that proof or misusing weaker sufficient bounds. Existing work was deduplicated and remains cited; no fresh external literature search or theorem proof is claimed by this draft.