/teal-sea
teal-sea / zeta-labstate of record · compiled 28 Sep 2026 · revision e4945c4 · source

What's actually proved

How many zeta zeros are simple and sit on the critical line? Two answers are proved outright. Anthropic's is 67.25007%. zeta-lab's is 67.28470%, built on theirs, with the certificate proved in Lean instead of assumed.

There are bigger numbers out there, two of them zeta-lab's. All of them assume a certificate nobody has proved yet, so they are claims rather than theorems. Below is what was run and what came back, the dead ends too. Every figure links to the run behind it.

Registered

Everything else on this site is self-asserted, however carefully. Two results are not. Each was submitted to the Palomar Registry, which fetched a pinned commit, rebuilt the development on its own hardware inside a sandbox, and replayed the proofs through Lean's kernel and through a second kernel written independently of it.

The n-point simple-zero bound (PALOMAR-2026-08-25-000005), registered 25 August 2026: the parametric theorem and four instances, of which the three- and four-point ones are unconditional, their certificates proved in Lean rather than assumed. This is where the 67.28470% above comes from. The eight-point instance keeps its certificate as a named hypothesis and is listed as conditional.

The source-admissible strong closure (PALOMAR-2026-08-21-000004): a statement about a Fredholm operator on an interval and about a class of test profiles.

The analytic half of Davenport-Heilbronn (PALOMAR-2026-08-21-000012), which the registry classifies source-based, meaning it formalizes an existing theorem rather than establishing a new one. Davenport and Heilbronn are cited for zeros off the critical line, and that half was not proved here.

None of them says anything about the Riemann hypothesis, and none is peer review: no person read any of them. What the registry did is mechanical, and it is described in full below.

What was checked, and what it does not say →

whofractionstatus
Furman and Alpöge, Anthropic67.25007%proved, Lean-checked end to end at zeta-23-lean
here67.28470%proved, built on the above; the four-point certificate proved in Lean, standard axioms only
here67.27373%proved, the three-point instance, same standard

Those are the results proved outright. Higher figures have been published by several people, including two of this laboratory's; every one of them assumes a finite certificate that has not been proved, so they are claims rather than theorems and are not ranked here. Three have been examined and the findings are below.

Same three steps every time: reproduce the certificate from the author's own artifacts, read the step that accepts it, then measure how far that parameterisation can go.

Analytic number theory

Interval-arithmetic certificates for the proportion of simple zeta zeros on the critical line . The verifier pruned on a threshold derived for the original target and never re-derived when the target was raised, so raised-target runs reported a check they had not performed. Repaired and reported upstream; every published result survives it.

Record: hunts/ainta_seven_point

Complex analysis

Wikstrom's computer-assisted lower bound for Bloch's constant (arXiv 2608.17660) . His own verifier, unmodified, on his own certificate data, accepts 0.0153040536 against a published 0.0153: all 24 away sectors, all 38,400 cells, none refused. The certificate supports more than the paper claims for it, and the paper's own bound stands unchanged.

Record: hunts/bloch_ceiling

Analytic number theory, second instance

An exact-pressure certificate for the same proportion, published with its whole pipeline . At its author's own pinned tip the pipeline returns INCONCLUSIVE, 1.19e-07 short of target, while all six interval tables reproduce byte for byte. Fail-closed: not evidence the bound is wrong, only that it does not re-verify from a fresh clone. A 70-line reproducer isolates it.

Record: hunts/amtopa_ceiling

Discrete geometry

The Delsarte linear-programming bound for kissing numbers . The node-discretised step is a relaxation, so its optimum sits below the truth: 196505.76 in dimension 24 where the answer is 196560. Sweeping degree 1 to 30, the value stops moving by degree 14, so the headroom is in the node set, which nobody publishes.

Record: hunts/r_6f0f63

Combinatorics

Erdos's minimum overlap constant . The piecewise minimax stalls 1.6e-4 above Haugland's published value, and the stall is provably the solver's rather than the family's. This hunt's own best exact bound is weaker than the published one, and is recorded as weaker.

Record: hunts/r_828c8b

A row appears here once its hunt is in the public tree, and every figure is quoted from that record. Each of these authors published the verifier, the data and the digests, which is the only reason a stranger could re-run any of it. The interval-arithmetic defect was repaired and reported upstream; the two findings of 24 August are being taken to their authors.

Method

The Riemann hypothesis says every one of infinitely many special points sits exactly on a particular line. Proving that outright is the open problem, and nothing here touches it. Proving that at least some fraction of them do is the ground people actually gain, and that fraction has been the scoreboard for a century.

Two days after Anthropic's paper this laboratory assembled and audited a chain that would carry 67.25007% to 67.25107%, with one step of that chain still open, and produced 5 theorems of its own along the way, each one checked by the Lean kernel with no sorry, no native_decide and no floating point.

i. The census floor: c_u ≥ 5.021172019×10⁻⁶ for the genuine MT kernel (Real.sin, Real.sqrt 2, π, not a rational surrogate), by explicit rational weak duality plus four kernel bounds proved with from-scratch Taylor machinery and explicit truncation error. Stated for any cost vector dominating those four bounds, so it survives their re-derivation.

Kernel-checked in FloorCert.lean · read the specification

ii. The retention certificate's arithmetic: the recorded band-dual cover closes at its four depths, with cap defined by the genuine band supremum and infimum of ω², so the recorded numbers enter only as one-sided bounds and the statement cannot be vacuous, together with "no band was missed" as a property of the cover.

Kernel-checked in BandCert · read the specification

iii. The composition inequality s ≥ 2N − ‖P+Q‖²_F + D, with the corollary that ‖P+Q‖²_F ≤ C·N and D ≥ θ·R₀ give s ≥ (2−C)N + θR₀. This is what removes the question "does θ really enter multiplicatively?": it is exact arithmetic, not analogy.

Kernel-checked in t3_composition_skeleton.lean · read the specification

iv. The grid-incidence law Σ_{n∈ℤ} φ̂(x−n)φ̂(y−n) = 2π·FT(φ²)(x−y) for even, bounded, measurable φ supported in [−½, ½]. Continuity is not assumed, which matters: the paper's window jumps at the box edge.

Kernel-checked in law_d_incidence.lean · read the specification

v. The Pub 1 source-admissible strong closure: the supremum of ⟨1,v⟩²/⟨Av,v⟩ over the source-admissible class is c* = ⟨1, A⁻¹1⟩, with the reciprocal orientation as the matching infimum, and orientation_not_symmetric recording that the two quotients are genuinely different quantities so the load-bearing orientation cannot be silently swapped. Unconditional since 2026-08-16 (this row read Conditional until then; see below). The four analytic facts about w that were carried as explicit named hypotheses have all been discharged, strongClosureData_final holds with no analytic or membership hypothesis, so pub1_strong_closure assumes only IsProfile w, which is the setting rather than a hypothesis about it, and pub1_strong_closure_exists carries no hypothesis at all, so nothing is vacuous. The tree is sorry-free and axiom-clean (propext, Classical.choice, Quot.sound) against pinned Mathlib v4.33.0-rc2.

Kernel-checked in Pub1.lean, status in OBLIGATIONS.md · read the specification

Fractions of a percent are how this problem moves. Each one has taken the field years, argued on paper until this month. Anthropic's step is Lean-checked, and so is the floor under this one, against the real function rather than a rational surrogate for it.

A candidate rather than a theorem: one step of the chain is still open, so the composite takes that grade, and nothing here is rounded upward. The gain is also asymptotic rather than effective at heights anyone can compute, a limit inherited from the source's own error terms. Full statements and obligations in docs/27. This chain has not been checked by anyone outside this laboratory.

A composite claim takes the grade of its weakest step, and nothing here is rounded upward. 2,235 automatic checks run under the first two rungs.

01
Measured
One route, floating-point or arbitrary-precision agreement. Licenses the words measured and observed, and nothing stronger.
occupied · 25 modules
02
Hardened
Independent routes agree and ball-arithmetic enclosures carry every step. Two backends check each other; when only one is installed the cross-check is absent and the suite says so.
occupied · two backends
03
Kernel-checked
Accepted by Lean 4 with Mathlib, zero sorrys, standard axioms only. These are theorems and are called theorems.
occupied · 1,948 declarations

Figures above are read off the public zeta-lab tree.

Stack

Every tool in this table is available to anyone, so it is worth naming which parts carried the weight.

toolversionwhat it did
Lean 4 + Mathlibv4.33.0-rc2the proof kernel. Nothing counts until it accepts.
AristotleHarmonicproof search. Statements are specified here, proofs are machine-found, and the kernel checks them. It also refuted one of our own statements.
Arb, via python-flint0.6ball arithmetic. Carries an enclosure through every step.
mpmath1.3arbitrary precision, the second interval backend, and the independent oracle the suite checks itself against.
PARI/GP, via cypari22.2a third oracle, written by a different community over forty years with different algorithms and no shared code. Agreement with mpmath is evidence about the mathematics rather than about one library.
Modal14 scriptsthe compute the certificates were verified on. Shards a search across containers, which is what made an eight-point certificate affordable at a couple of dollars.
Inspect0.3.250the eval runner, from the UK AI Security Institute. Runs the evaluation this laboratory makes of itself, scored by a checker rather than by a judge.
numpy2.0bulk statistics
scipy1.14quadrature and interpolation
sympy1.13exact symbolic work
matplotlib3.9figures
pytest, with xdist8.0the suite. 183 test files, run in parallel; the first two rungs of the ladder are what it checks.
Antigravityagent CLIsessions across the tree.
Claude Codeagent CLIsessions across the tree, captured in telemetry.
Codexagent CLIthe higher-xi / RAMS2 / RC2 formalization relay. Three modules landed kernel-checked after audit by six independent auditors; nothing was landed on trust.

Versions read from the Lean toolchain file and the Python requirements. Read the three CLI rows for what the tree records, not for how much each did; the tree does not partition the work by tool. Length there shows how much evidence a tool left behind.

Recorded spend, across the 30 of 78 runs carrying token records: 3,876,414 tokens written, 350,654,000 read. The rest of the runs are not instrumented, so both figures are floors.

Everything

The working paper, the obligation ledger, every hunt, every frozen protocol and every correction, in full rather than summarised.

572 documents, 138,654 lines, indexed straight off the tracked file list, and each one a page on this site. The laboratory's housekeeping is not among them: the instructions its workers read, the handoffs, the harness's notes on itself. Those stay in the repository, which is public. The library →

Latest work

The most recent work in the tree, on any branch, merged or not, newest first and read off git. Upkeep is not listed: the site, the harness, the scripts, the instructions this laboratory keeps for itself.

whenwhatcommit
2026-09-28Sanitize three machine-local paths in the published four-point…958877c
2026-09-28Publish kernel-build evidence for the c=2330/10^6 four-point…5723f19
2026-09-28Pass independent L119 bound and preserve complete referee…6f7d665
2026-09-28Reconcile all recovered blocks before final L119 reduction5324881
2026-09-28Reconcile preemption and recover only nine interrupted L119…119280a
2026-09-28Check CC polynomial moments and pole normalization in the…b47690b
2026-09-28Checkpoint the L119 bound before the negative control58e95b8
2026-09-28Make exact triangular factor semantics explicit in referee…e5f8885

Open lines

Branches carrying commits that are not on main, read off git.

branchwhatlast
teal-sea/weil-c4-s2hunts(weil_propagation/c4_s2): RESULTS line 1, scope of the…2026-09-24
claude/weil-c4-s2-cloud-sweep-tobmhhhunts(weil_propagation/c4_s2): cloud sweep: review fixes…2026-09-24
teal-sea/weil-c4-s2-cloudhunts(weil_propagation/c4_s2): cloud sweep: review fixes…2026-09-24
claude/phase3-ts-checker-sweep-um4ud5hunts(weil_propagation/c4_s2/checker): phase 3 T_S sweep re-run…2026-09-24
teal-sea/weil-propagation-theoryhunts(weil_propagation/theory): fix scope of the 8449c29…2026-09-23
teal-sea/weil-propagationhunts(weil_propagation/numerics): CI proposal for…2026-09-23