Erdős problem 472 open updated 16 July 2026

Can a greedy sequence of primes run forever?

I’m Syn. This is my working notebook for an attack on Erdős 472—models proposing ideas, code trying to break them, Lean checking the semantic core, and me deciding what survives.

I have not solved it. The best trajectory has passed one million verified steps, but a finite run is still finite.

live object

The beginning of the singleton [7] orbit

q = 6x + 1

origin / x = 1 P(x) = 7 seed coordinate
01

The rule

Start with a finite increasing list of primes. At each step, take the current last term qn and inspect

qn + qi − 1

for every earlier term qi. Append the least prime candidate. If they are all composite, stop.

What we can prove∀L ∃ seedis not∃ seed ∀LWhat the problem asks
02

What has survived scrutiny

Ideas the project has killed

  • 01

    Static certificates. CRT can assign fresh blockers to any fixed finite shell.

  • 02

    Scheduled winners. They hide a thin infinite-prime conjecture.

  • 03

    Unsigned factor counts. They do not cross the sieve parity barrier.

  • 04

    Finite memory. Rich finite behaviour can still be programmed to terminate.

03

Where the proof stops

The current candidate is a growing-age selector built from the actual birth history of the singleton [7] orbit. After filtering small prime factors, the remaining obligation is one centered rough/semiprime estimate.

F280 / not yet proved
(R(bW) − R(cW))+ + (S(cW) − S(bW))+ < Π(bW)

R cube-root-rough mass

S rough semiprime mass

Π prime mass

The right side is a positive flat prime reserve. If the two errors on the left stay smaller, selected prime mass is positive and a hypothetical terminal row cannot exist. That uniform inequality is the missing mathematics.

finite evidence

Centered budget against failure threshold 1

38,692 rows audited · ranks 1,309–40,000

1.0 / failure
selected0.923228708

rank 1,755 · margin 0.076771292

Finite means finite. The audit found no failure in this interval. It does not prove the next row.

04

How I’m working

AI is useful here when it behaves like a restless collaborator: propose a precise claim, write the falsifier, search for the smallest counterexample, and keep the analytic debt visible.

Code checks finite worlds. Lean checks deterministic implications. Primary papers supply named analytic inputs. I remain responsible for deciding what the evidence supports and what gets published.

proposeattackformalizepublish honestly
References and artifacts

Primary sources first.