Provedalphabaguettebaguette-zrange
Baguette-Z widths: 80 blocks with gains ≤ 2¹⁰ fit int64 with 44 fractional bits, int32 with 13
For all , if and , then
What the notation means · or point at any symbol
- RReal numbers
- The real numbers: every point on the number line. (
ℝ) - x ∈ TVariables
- G, x0 range over real numbers: the statement must hold for every choice. (
(G x0 : ℝ)) - ≤At most
- G is less than or equal to 2^10. (
G ≤ 2 ^ 10) - |x|Absolute value
- The size of x0 without its sign. (
|x0|) - <Less than
- (|x0| + 80 · (2 · G)) · 2^44 is strictly smaller than 2^63. (
(|x0| + 80 * (2 * G)) * 2 ^ 44 < 2 ^ 63) - ∧And
- Both hold: (|x0| + 80 · (2 · G)) · 2^44 < 2^63, and (|x0| + 80 · (2 · G)) · 2^13 < 2^31. (
(|x0| + 80 * (2 * G)) * 2 ^ 44 < 2 ^ 63…) - +Addition
- |x0| plus 80 · (2 · G). (
|x0| + 80 * (2 * G)) - aᵇPower
- 2 raised to the power 10. (
2 ^ 10) - ·Multiplication
- 2 times G. (
2 * G)
theorem bz-baguette-width-certificate (G x0 : ℝ) (hG : G ≤ 2 ^ 10) (hx0 : |x0| ≤ 2 ^ 10) :
(|x0| + 80 * (2 * G)) * 2 ^ 44 < 2 ^ 63 ∧ (|x0| + 80 * (2 * G)) * 2 ^ 13 < 2 ^ 31With embedding coordinates ≤ 2¹⁰ and per-layer increments ≤ 2·2¹⁰ (gained normalized outputs, gains in the AdamW box ≤ 730 < 2¹⁰), the residual stays below 161·2¹⁰ < 2¹⁸. That fits int64 at 2⁻⁴⁴ resolution and int32 at 2⁻¹³. Part of the plan 'Baguette-Z: closed-loop certified integer-state training'.
How it was proved
- Proved
- 3 October 2026
- By
- Ajax, with the agent “Baguette math & proofs (Claude Code, 2026-10-03)”
- Checked by
- Lean's kernel (Lean 4.33.1, with Mathlib); the proof is signed and replays offline
- Axioms
- Classical.choice, Quot.sound, propext
- Proof
- 19 lines of Lean
Sign in to see the proofThe proof, its signed report and its history.