Erdős problem · #742 · Murty–Simon conjecture · AMS 5 · open
Erdős 742
A diameter-2-critical graph on n vertices has at most ⌊n²/4⌋ edges. A repository claims a complete Lean proof; SISO's review lane got the statement to match Formal Conjectures and the dependency audit to pass, and the theorem did not compile. No kernel or axiom receipt exists, so the state here is unverified.
01
Problem statement
from the Formal Conjectures file fetched at review time (sha256 25f7c5cd…)Murty–Simon Conjecture. Let G be a graph on n vertices with diameter 2 such that deleting any edge increases the diameter. Is it true that G has at most ⌊n²/4⌋ edges? Equality is conjectured to hold for the complete balanced bipartite graph K⌈n/2⌉,⌊n/2⌋.
- formal statement
theorem erdos_742 : answer(sorry) ↔ ∀ (V : Type*) [Fintype V] [DecidableEq V] (G : SimpleGraph V) [DecidableRel G.Adj], IsDiameter2Critical G → G.edgeFinset.card ≤ (Fintype.card V) ^ 2 / 4- definition
IsDiameter2Critical G := G.diam = 2 ∧ ∀ e ∈ G.edgeSet, (G.deleteEdges {e}).diam ≠ 2- category
- research open · AMS 5 (Formal Conjectures tag)
- resolved up to
- a finite check: Fan [Fa87] n ≤ 24 and n = 26; Füredi [Fü92] all sufficiently large n
- lane brief
- targets/742/README.md: "audit and extend proof-logging SAT work at order 25. Preserve CNF identity, solver version, proof log, independent replay, and any decoded graph."
we need one for each ERDOS question we need one for each god question we need one for each agent zero, we need one for each research doc, we need one for each project one pagerShaan → AGENT-ZERO · 00_AGENT_ZERO/source/2026-09-06-1650-… · 6 Sep 16:50
02
Approach · reductions
what the fleet chose to do and why- 1Admit 742 first as a bounded falsification/certification job admittedaudits/easiest-targets.md: "the practical frontier is one finite order (25), the SAT input and candidate checker are public, and a SAT witness would be a genuine counterexample"
- 2Then a repository claimed a complete Lean proof claimaudits/solver-repositories.md §742: Erdos742.lean 173 KB, Equality.lean 300 KB, Erdos742Full.lean 474 KB, two papers; "no lakefile.lean, lean-toolchain, workflow, build script, certificate manifest, or pinned dependency file"
- 3Goal /review-742-proof goal"Wrap the pinned Erdos742/Erdos742 sources in the smallest reproducible Lean project, compile the headline theorem, print its transitive axioms, and determine whether any sorryAx or unstated hypothesis reaches the result." PASS requires a reproducible build plus statement-fidelity and axiom receipts.
- 4Reusable SAT path kept availableleansolving/leansms (Apache-2.0) MurtySimon encoder,
lake exe murty-simon-generate <n>: verified CNF/LRAT impossibility proofs for finite orders; restricted LRAT classes exist upstream (edisonymy/erdos-lean-research experiments/erdos742)
Pindetached checkout at 814289f2, source sha256 recorded
Matchstatement against Formal Conjectures 742 (fetched, sha256 recorded)
Buildsmallest wrapper project: lakefile, toolchain v4.33.0-rc1, Main.lean with
#print axiomsAudittransitive axioms for sorryAx, custom axioms, native_decide
03
Certificate · replay status
proof state only from a receipt; the claim ceiling of this page proof stateunverifiedno kernel or axiom receipt; "classify the claim as candidate-proof/blocked, not kernel-checked or solved" (HANDOFF.md)
statement fidelitypassedIsDiameter2Critical and erdos_742 agree textually with Formal Conjectures after removing answer(sorry)
static sorry auditpassed · 1 literal sorry, statically unreachabledependency-audit.txt; MAC-MINI handoff: "no kernel/axiom receipt exists"
buildexit 124 · timed out in dependency acquisitionlake update tried to clone Mathlib; a path-dependency retry (647's mathlib) was stopped
replayable certificatenoneno LRAT/DRAT for order 25 in this workspace; upstream restricted classes are not this claim
independent review (Sol)not runSol is reserved for a genuine proof boundary; none reached
| Gate | Command | Result | Receipt |
|---|---|---|---|
| checkout pinned | git -C Erdos742 rev-parse HEAD | 814289f2… | HANDOFF.md verify line; status --short empty |
| source hash | sha256 Erdos742/Erdos742.lean | 355a8539… | HANDOFF.md evidence line |
| reference statement | curl … formal-conjectures/…/742.lean | exit 0 | formal-conjectures-fetch.exit = 0; sha256 25f7c5cd… |
| bounded build | timeout 10s … lake build Erdos742 | exit 124 | build-timeout.exit = 124; build.log 15 lines, cloning aesop/Qq/batteries |
| axiom print | lake env lean Main.lean → #print axioms Erdos742.erdos_742 | not reached | Main.lean present; never executed |
replay command · what a certificate would take
cd SISO_Research/erdos/targets/742/runs/proof-review # obtain Mathlib at exactly ae0d973 without touching Erdos742/, point lake-manifest at it, then elan run leanprover/lean4:v4.33.0-rc1 lake build Erdos742 elan run leanprover/lean4:v4.33.0-rc1 lake env lean Main.lean # must print axioms of Erdos742.erdos_742From HANDOFF.md "Smallest next action". Until the second command completes and its transitive axioms are inspected for
sorryAx, custom axioms and native_decide, the chip above stays amber.04
Compute spent · host
what is measured, what is not
1review run (proof-review)
10 sbounded build budget · exit 124
0kernel-checked theorems
2historical ChatGPT turns recovered · 240 m + 33 m—tokens · not measured
- host
- laptop (MacBook), frozen 30 Aug 16:35: "Freeze all laptop proof/Lean compute and move the programme to a dedicated Mac Mini workspace"
- next host
- Mac Mini M4 · 16 GB (ESTATE.md);
ERDOS-LEADseat present in atlas/now.json, node mini, idle - toolchain
leanprover/lean4:v4.33.0-rc1installed; Mathlibae0d973named by upstream comments, not acquired- what is not measured
- wall time of the lane, tokens, CPU; the workspace has no run ledger for 742 (only the exit-code receipts above)
- historical
- audits/connector-recovery-2026-09-06/erdos-742-historical-chat.md: a recovered ChatGPT thread ("Worked for 240m 4s", "I did not finish a full proof"); prose, not a certificate
05
Decision log
DECISION-LOG.html and CHECKPOINT-2026-08-30.html, verbatim lines as factsfact · has a receiptproposal · not yet
WIN: #23 exact replay passed; #647 bounded C slice passed; #742 and #848 release blockers were localized.DECISION-LOG.html · 30 Aug 2026
DECISION: Freeze all laptop proof/Lean compute and move the programme to a dedicated Mac Mini workspace.DECISION-LOG.html · 30 Aug 2026
TRAP: A successful arithmetic replay is not a theorem-interface proof; a repository claiming kernel closure is not reproducible when its own manifest gate fails.DECISION-LOG.html · 30 Aug 2026
LEARN: Run release-integrity and dependency-lock gates before expensive proof review; keep intentional axioms separate from certified finite evidence.DECISION-LOG.html · 30 Aug 2026
Hard-won traps: Do not accept certificate arithmetic as semantic theorem validation. Do not accept a Lean claim without a pinned dependency lock and executed axiom audit. Do not let a returned worker leave an orphaned Lean process.CHECKPOINT-2026-08-30.html · 30 Aug 2026 16:35
Exactly one resume step: a fresh Mini Codex reads the handoff, then resumes the interrupted #23 Sol interface review before admitting any new proof compute.MAC-MINI-HANDOFF-2026-08-30.md §"Exactly one resume step"; 742 waits behind it
Limits
- Proof state on this page comes only from targets/742/runs/proof-review receipts and the 30 Aug audits; no handoff sentence upgrades it.
- The upstream repository's own claim ("only ordinary Lean axioms, no sorryAx") is quoted by audits/solver-repositories.md and is not a receipt.
- The erdos workspace sits inside SISO_Workspace (outer repo untracked); no commit hash pins these files.
Version
proof-review · 30 Aug 16:22HANDOFF.md status: done · result: PARTIAL
06
What would close it
- 1Kernel-checked build of erdos_742 with an executed axiom print closes the claimthe two commands in §03 complete on a clean machine with Mathlib ae0d973; transitive axioms contain no sorryAx, custom axiom or native_decide
- 2Or a finite counterexample at order 25 closes the conjecturea SAT witness decoded to a diameter-2-critical graph with more than ⌊625/4⌋ = 156 edges (easiest-targets: "would settle the conjecture negatively")
- 3Or a replayable LRAT/DRAT for order 25 finiteextends Fan's finite check by one order; a finite result, not the conjecture (README "Completion standard")
- 4Cold Sol review at the boundary gateonly once one of the above exists (AGENTS.md operating rules)
07
Repos · certificate files
the erdos workspace (inside SISO_Workspace, not a git repo of its own) and the pinned upstreamsSISO_Research/erdos
no public locator
AGENTS.md · manager/{CHARTER,launch-board}.md · goals/ · targets/742/runs/proof-review/ (wrapper project, logs, exit receipts, dependency-audit.txt, pinned Erdos742/) · audits/ · DECISION-LOG.html · CHECKPOINT-2026-08-30.html · MAC-MINI-HANDOFF-2026-08-30.md
workspaceSISO
No public payload
Erdos742/Erdos742
https://github.com/Erdos742/Erdos742
The claimed proof: Erdos742.lean (173 KB), Equality.lean, Erdos742Full.lean, two papers. No lakefile, toolchain, lock or manifest.
no declared licence
Open repo
google-deepmind/formal-conjectures
https://github.com/google-deepmind/formal-conjectures/blob/main/FormalConjectures/ErdosProblems/742.lean
The reference statement (fetched at review, sha256 25f7c5cd…).
Apache-2.0
Open repo
leansolving/leansms
https://github.com/leansolving/leansms
MurtySimon encoder; verified CNF/LRAT impossibility proofs for finite orders (the order-25 path).
Apache-2.0
Open repo
08
Fleet · seats
the four problems as the 30 Aug handoff observed them| # | Lane | Observed status (30 Aug) | Durable evidence |
|---|---|---|---|
| 742 | REVIEW-742 · luna-review-742 | PARTIAL statement/static dependency audit passed, theorem did not compile | targets/742/runs/proof-review/HANDOFF.md |
| 23 | VERIFY-23 · luna-verify-23 · SOL-REVIEW-23 | PARTIAL exact arithmetic replay passed; six mathematical interfaces not discharged; Sol review interrupted, no verdict | targets/23/runs/baseline/HANDOFF.md |
| 647 | REPLAY-647 · luna-replay-647 | handoff C validation slice passed (55 closed, 41 open residues, one intentional Stage-1 axiom); Lean replay interrupted | targets/647/runs/baseline/HANDOFF.md |
| 848 | VERIFY-848 · luna-verify-848 | PARTIAL fast verifier passed; kernel-package public gate failed (PDF bytes/hash disagree with manifest) | targets/848/runs/baseline/HANDOFF.md |
root Codex · Erdős manager (laptop, frozen)luna-review-742 · partialsol · reserved, not runERDOS-LEAD · mini · idle
09
Attempts
upstreamreview lanedecisionsdocshis words
10
Mind map
problem → reductions → repos → gatesReductions3
Repos3
Gates4
Hosts2
Each leaf links to its record. Branches are laid out by subtree size, not by hand.
12
Agent entry
cat SISO_Research/erdos/targets/742/runs/proof-review/HANDOFF.mdcat SISO_Research/erdos/targets/742/runs/proof-review/build-timeout.exit # 124jq '.tasks[]|select(.id=="REVIEW-742")' SISO_Research/erdos/tasks/TASKS.jsonshasum -a 256 SISO_Research/erdos/targets/742/runs/proof-review/Erdos742/Erdos742.lean # 355a8539…
Agent entry
- Read first
SISO_Research/erdos/AGENTS.md → MAC-MINI-HANDOFF-2026-08-30.md → tasks/TASKS.json → SISO_Research/erdos/targets/742/runs/proof-review/HANDOFF.md- Owner
- Erdős manager (root Codex, frozen) · ERDOS-LEAD on the Mini, idle · last writeback 30 Aug 16:35 (handoff) · 6 Sep 04:32 (recovery audit)
- Done when
- Open this page and see the problem, the proof-logging status (unverified, amber: statement audit passed, build exit 124, no axiom receipt) and the replay command with the receipt files it would produce.