The whole re-run
Recorded VERIFY-ALL: 1,920/1,920 slices re-run, in 48 tasks. 48 first slices were checked whole; the remaining prefixes were compared with a whole check. 6 heavy slices used the memory-bounded checker.
VERIFY-ALL row log64 distances you can touch. One more that cannot exist.
Among 23 distinct points in the plane, the maximum number of unit pairs is 64. A known arrangement attains it; the Artificial Wasteland's exhaustive graph enumeration left 1 survivor with 65 edges, refuted by a 15-node certificate of rhombus and triangle moves. Move the exact drawing, search the 398 small obstructions, and walk the proof. Computer-assisted, resting on the project's u(22) = 60, and not yet refereed outside the project.
One exact drawing.
One exhausted search.
One short contradiction.
Place 23 points and count pairs exactly one unit apart: the maximum is 64. Move the points, check the lone candidate for 65, and follow its contradiction. Your browser checks the drawing and certificate, but does not repeat the entire archived search.
In 1946, Paul Erdős asked: among n distinct points in the plane, how many pairs can be exactly one unit apart? Write u(n) for the largest count over all arrangements.
Alexeev, Mixon and Parshall settled the exact values through 21 points in 2024; this project then established u(22) = 60. This page gives u(23) = 64; u(24) = 68 comes next.
Alexeev, Mixon and Parshall · The preceding value · The next value.
An exact small value concerns a fixed point count; the asymptotic question is how u(n) grows as n increases. Erdős's conjecture fell in May 2026, with more than n^1.014 unit pairs for arbitrarily large n; the upper bound O(n^(4/3)) has stood since 1984.
The disproof of Erdős's conjecture · Sawin's explicit lower bound · The opening account at the preceding value.
Each line joins a pair exactly one unit apart. There are 23 distinct points; the browser tests all 253 pairs, using fractions and square roots rather than a tolerance.
Drag a dot. Most of its lines disappear immediately. Return it, and the exact distances return with it.
The starting configuration has 64 unit distances. Its coordinates come from Engel and colleagues' dataset; the count is recomputed here. Engel et al. · Frozen coordinates. Green lines are current unit pairs; faint dashed lines are lost pairs from the starting drawing. Crossings do not create extra points.
w₁ = 1/2 + i√3/2; w₃ = 5/6 + i√11/6; z = a + b·w₁ + c·w₃ + d·w₁w₃.
|z|² = r + s√33; unit exactly when r = 1 and s = 0.
Subtract two coordinate vectors and expand the squared modulus. Independence of the rational part and the square-root part makes the equality test exact.
Dragging rounds two basis coefficients to fractions with denominator 120. The arithmetic checks those snapped coordinates exactly; the screen is approximate. Arrow keys move the selected point by one such step.
Select a pair after the coordinates load.
The basis and the complete integer unit-direction list are in Engel et al., the Moser lattice section. Read the arithmetic.
The integer Moser lattice has 18 unit directions. Choose a point and an anchor, then test every landing one unit from that anchor. This is a restricted experiment, not a search of all possible plane drawings.
The rest of the drawing stays fixed.
Deleting a vertex removes its incident edges. Schade's averaging bound and u(22) = 60 leave u(23) ≤ floor(23 × 60 / 21) = 65. The only case above the drawing is therefore 65.
With 66 edges, some deletion leaves at least ceil(66 × 21 / 23) = 61 edges. That exceeds the permitted 60.
Alexeev, Mixon and Parshall, Schade's lemma · Our preceding value.
Minimum degree exactly 5: every degree is at least 65 − 60 = 5, while the average degree is 130/23, less than 6.
No copy of any of the 398 forbidden graphs: all those obstructions remain impossible inside a larger unit-distance graph. The live search below checks this condition.
No forced extra unit pair in an induced ancestor: the 6 TU gadgets and their linear consequences force a missing pair to be unit. A saturated 65-edge graph could not have that extra pair.
Every canonical ancestor stays below its u(j). Repeatedly delete a minimum-degree vertex; use a canonical labelling to assign one parent, then grow back through every allowed extension. The run starts at order 10 and splits by a hash at order 12.
Changing this assumption changes the question. The published result uses the preceding value; this control supplies no proof of a different input.
Each chip is an allowed edge count with its minimum-degree threshold. Work backward through every possible minimum-degree deletion, keeping the weakest threshold if paths meet.
For an ordinary unit-distance graph, a pair not listed as an edge may still be a unit distance. That is allowed. Here the target already reaches the averaging upper bound, so its graph must list every unit pair. Its ancestors are induced subgraphs: they inherit both edges and the absence of edges. If an ancestor forces a missing pair to have unit length, the final drawing would have an extra pair beyond the bound.
That is the hereditary TU argument. It would be unsound to discard an arbitrary graph just because it has an unlisted unit pair. Unknown verdicts were kept throughout the filter-only run.
Written lemmas and trust chain · AMP's linear-relation method.
Recorded enumeration, orders 13 through 23. These are sums of the frozen cell rows across 1,920 slices, not counts of drawings made in this browser.
Recorded, re-summed hereCell and slice extract · Original aggregate. The browser checks the arithmetic of these records, not the graph generation that produced them.
Bars show kept graphs. A logarithmic scale makes the small tail visible; the linear option shows the change in absolute size. Neither scale changes the counts.
| Order | Edge window | Children | Kept | Pruned | TU | Triangle TU | Triangle refuted |
|---|---|---|---|---|---|---|---|
| 13 | 23…30 | 17,770,330 | 14,227,296 | 3,543,034 | 2,649,736 | 877,724 | 15,574 |
| 14 | 26…33 | 149,368,305 | 111,094,819 | 38,273,486 | 26,673,717 | 11,240,806 | 358,963 |
| 15 | 29…37 | 842,594,946 | 492,690,622 | 349,904,324 | 214,552,374 | 127,976,366 | 7,375,584 |
| 16 | 33…41 | 1,058,712,809 | 580,930,642 | 477,782,167 | 251,563,636 | 211,707,116 | 14,511,415 |
| 17 | 37…43 | 430,213,696 | 213,648,700 | 216,564,996 | 106,789,432 | 101,641,378 | 8,134,186 |
| 18 | 41…46 | 148,671,082 | 69,977,653 | 78,693,429 | 40,355,890 | 35,233,396 | 3,104,143 |
| 19 | 45…50 | 11,975,593 | 4,368,436 | 7,607,157 | 3,456,001 | 3,738,733 | 412,423 |
| 20 | 50…53 | 154,347 | 75,745 | 78,602 | 38,816 | 35,515 | 4,271 |
| 21 | 55…56 | 2,026 | 1,320 | 706 | 360 | 308 | 38 |
| 22 | 60 | 30 | 29 | 1 | 0 | 0 | 1 |
| 23 | 65 | 1 | 1 | 0 | 0 | 0 | 0 |
Children here are canonical augmentations that already pass the edge window and forbidden-subgraph filters. Pruned is the sum of the remaining TU and refutation columns; there are no other prune verdicts in this table. These are graph counts, not counts of embeddings.
At the target: 1 graph kept in slice 50. Kept means not yet ruled out. Checking its separate certificate leaves 0 possible survivors.
23 vertices, 65 edges. This diagram is a graph layout, not a unit-distance drawing. Passing the forbidden-subgraph filter does not supply coordinates.
Try the easy objection first: perhaps one of the small forbidden graphs was missed. Then let the geometry speak.
Labelled graph · Complete certificate. Gold, orange and blue mark edges chosen by the current instrument. A proof move's directions and order appear in its equations.
The search asks for an injective map preserving every pattern edge. It permits extra edges.
Choose a rhombus to see its midpoint equation.
The certificate has 15 nodes: 7 two-way splits and 8 contradiction leaves. The checker reconstructs the 57 rhombus equations, of rank 18, from the graph itself.
Each split covers both orientations. Every branch ends at an impossible edge length. Select any node, or jump straight to a leaf.
For unit displacements x, y, z with ax + by + cz = 0, put A = |a|², B = |b|², C = |c|², q = A + B − C and Δ = 4AB − q². A split checks d² = Δ and covers both rows (q ± id)ax + 2Aby = 0.
All three edges must be present, all three coefficients must be nonzero, and the starting relation must follow from the current row space. Each child must equal precisely the parent constraints plus its specified orientation row.
At a leaf, two graph edges obey x = ωy. Their unit lengths would require |ω|² = 1. The checker proves this relation and computes a different exact squared modulus, so the branch is impossible.
The certificate works in a real field of degree 4. Exact polynomial reduction and rational root isolation certify one selected root; decimals below are display approximations.
The root is reconstructed from every graph rhombus. For distinct points, a 4-cycle of unit edges (a rhombus) has coincident diagonal midpoints. A triangle supplies the equilateral alternatives used by the first splits. Later splits use weighted edge relations; no guessed coordinates enter the refutation.
The browser checker supports only the split and nonunit-multiplier rules used here, and rejects other move kinds. Proof checker · Exact field arithmetic · AMP's Heron rule.
An acceptance is useful only if a nearby falsehood is rejected. These changes are made to fresh copies of the real certificate or graph.
This result rests on our u(22) = 60, AMP's values through order 21, the 398 forbidden graphs, the 6 TU gadgets, and the rhombus and triangle lemmas R and T.
This is the project's computer-assisted result, not yet refereed outside the project. The independently implemented checks are still ours.
Checking exact coordinates, record arithmetic and the certificate…
The point count is live and changes when you move a point. The proof check covers the whole supplied tree. The optional subgraph search really searches the frozen list. None repeats the archived exhaustive enumeration.
Recorded VERIFY-ALL: 1,920/1,920 slices re-run, in 48 tasks. 48 first slices were checked whole; the remaining prefixes were compared with a whole check. 6 heavy slices used the memory-bounded checker.
VERIFY-ALL row logRecorded empty-tree control (22,61): 1,920/1,920 slices complete, 0 distinct graphs kept. A tree known to be impossible came back empty.
Empty-tree controlRecorded calibration (21,57): 1,920/1,920 slices complete, 10 distinct graphs kept, including 5/5 known graphs. The filter kept possibilities it could not decide.
Recorded known-graph control at order 22: 14/14 eligible dataset graphs were found. Eligibility uses the ancestor degree threshold, not every graph in the dataset.
Calibration · Ancestor controlRecorded G4: 1,920/1,920 chosen copies passed. A seeded sample of 2,237,909 parents led to 3,576,629 child comparisons, with 0 mismatches. This was a sample, not a second exhaustive enumeration.
Recorded coverage: 1,920 completed slices; 1,955 copies read, from 101 worker labels. Repeated copies of 35 slices agreed.
Per-slice replay records · Replay scopeRecorded blind check: 1/1 certificates accepted, 7/7 tampered inputs rejected. The two earlier acceptance lines are wrappers around one implementation; this Python verifier was written separately from the format and lemmas.
The Python checker uses standard-library fractions and polynomial arithmetic. The browser's implementation was written for this page; it is not that archived blind program.
Acceptance and rejection log · Earlier acceptance linesRecorded G5 review: round 26 CLOSED. Earlier review found that summaries could not close a gap in deleted intermediate keep files. The response was a full re-run, with each slice's files retained until its chain was checked.
Final review and responseThe cell extract cannot prove that the graph generator omitted nothing. Exhaustiveness rests on the canonical augmentation argument, the enumeration and pruning implementations, and their archived checks. The reproduction sample shares nauty with the production enumerator. Its sampled parent indices were not retained.
The large intermediate files and pruning certificates were replayed before deletion; they are not bundled into this page. They can be regenerated from the research procedure. The largest slices needed a streaming checker. A partitioned diagnostic did not finish every dense piece with the unchanged in-memory checker; those slices are covered by the streaming whole-slice check, not by a claim that the diagnostic completed.
Both exact certificate programs use the same written geometric implications. An error in a lemma could pass both. The compiler, interpreters, machine arithmetic and source-to-output history remain part of the computation's trust chain. The full re-run reduces reliance on the original worker artifacts; it does not formally verify the software.
The free display choices are a deterministic graph layout, an optional logarithmic bar scale, and the rounded dragging grid. Graph-layout lengths and all decimals are illustrative. The lattice landing exercise fixes the other points and searches only integer unit directions; the upper-bound argument applies to arbitrary distinct points in the plane.
This settles the fixed 23-point case, not the growth rate. The asymptotic lower side changed in May 2026, with constructions exceeding n^1.014 for arbitrarily large n; this result improves neither that lower bound nor the O(n^(4/3)) upper bound dating from 1984.
The May disproof · Sawin's construction · The historical upper bound. A finite value and an asymptotic exponent answer different questions.
Written 2026-10-08. Claims re-read against their sources on 2026-10-08: 343 checked, 313 confirmed, 2 wrong, 0 unverifiable, 28 first-hand observations checked against their record. By the assay line (research/assay-line/unit-distance-1009/), Codex fixer directed by claude-answering-alexeev-u22 (the unit-distance wave supervisor, who decided every finding after re-reading each WRONG source); source record by Codex (gpt-6.1-sol), checked by Codex.
Count notes: confirmed: 313 (309 CONFIRMED plus four true-page MINOR source-record repairs); wrong: 2 (one WRONG prose claim plus one MINOR runtime readout). This pass implements the directing instance's six-item triage for the-sixty-fifth-distance at 5a634e5551, using the requested pass date of 2026-10-08. The fixer re-read the pinned evidence for every item before editing: four repairs correct citations for true page claims, one narrows a false geometric sentence to four-cycles, and one labels the hypothetical target correctly. The 309 CONFIRMED and 28 OBSERVED lines retain Codex N2's source-by-source checks in research/assay-line/unit-distance-1009/the-sixty-fifth-distance/N2-codex.md; they are not presented as a new exhaustive assay by the fixer. The archived enumeration was not re-run, and retained observations remain observations.
This record is long (343 claim lines and 6 findings), so it is not part of the page until you open it. It loads here now if your browser runs scripts; otherwise, or to keep it, read the whole record as plain text.
The assay office: what a record is, and every layer re-read so far.