2026-10-01. Lean now checks the following statement for unbounded starting values:
n > 1, k ≤ 1024, FirstLight(n,k) ⇒ T^k(n) < n.
T is the accelerated map. FirstLight(n,k) means that k is the first prefix
with 3^a(n,k)<2^k, where a counts odd steps. The existence of such a prefix is
not asserted. In particular, this is not a universal 1,024-step stopping bound
and does not prove universal descent. No historical priority is claimed.
The generic accumulator theorem divides the proof into two finite checks:
- A coefficient-gap bound for all lengths through 1,024. The compact checker proves this with threshold 99,730 by checking one extremal odd count per length.
- Every exceptional start 1<n<99,730 has a first-crossing descent. These starts are checked at horizon 135, then their witnesses are embedded in the larger horizon. There are 99,728 such starts.
The small-start range is divided into 49 adjacent intervals, at most 2,048
inputs each. Each interval uses kernel decide; a cumulative-prefix theorem
combines it with the preceding interval. Separate compiler processes release
reduction memory between intervals. The final theorem uses the resulting
single prefix certificate. Python generates the scaffolding only; its outputs
are never trusted as mathematical evidence.
A monolithic attempt and a build interrupted by a shared dependency rebuild
were superseded by the completed serial-module build. The complete certificate
target passed all 116 build jobs. The interface also supplies ExactThrough 1024 to the generic first-descent and residue-class classification API.
The independent Python probe checks the full triangular gap table, not just the extremal count, and follows every exceptional trajectory. Its limiting gap pair is (k,a)=(485,306); its latest first crossing is 135, attained at n=35,655. The probe agrees with the kernel certificates and records no failures.
Five proof endpoints, including the final cumulative prefix and the unbounded-n
theorem, have an axiom audit. The gap certificate has no axioms; the combined
proofs use only standard Lean logical axioms. No native_decide or added axiom
is used. These target checks do not by themselves assert that current full
repository CI has completed.
python3 scripts/generate_first_light_1024.py --check
lake build Collatz.Strategy.FirstLight1024Interface
lake env lean scripts/audit_first_light_1024.lean
python3 scripts/probe_first_light_certificates.py --horizon 1024Final certificate, generated intervals, generic interface, axiom audit, probe.