Restated as the explicit refutation of "the ratio grows without bound":
ProvedBridges.AttentionKneeFlatness.geoRow_ratio_blowup_refutedaether-catalogbridges
Restated as the explicit refutation of "the ratio grows without bound": there is a single constant, depending on the gate alone, that dominates the knee-to-floor ratio for every geometric row.
theorem Bridges.AttentionKneeFlatness.geoRow_ratio_blowup_refuted{g : ℝ} (hg0 : 0 < g) (hg1 : g < 1) :
∃ C : ℝ, 0 < C ∧ ∀ a : ℝ, 0 < a → a < 1 →
(knee (geoRow a) g : ℝ) ≤ C * (g ^ 2 / geoEnergy a) := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Bridges/AttentionKneeFlatness.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Bridges/AttentionKneeFlatness.lean
import Mathlib
import Definitions.Def_Bridges_AttentionKneeEntropyBound
import Definitions.Def_Bridges_AttentionKneeFlatness
import Definitions.Def_Bridges_AttentionKneeGeometry
/-
# Cycle 3: the flatness dichotomy for the Cauchy–Schwarz knee floor
`Bridges.AttentionKneeGeometry` gives order/grid control of the retention knee
`k*(g)`, and `Bridges.AttentionKneeEntropyBound` gives the information-theoretic
floor `k*(g) ≥ g² / E`, where `E` is the attention energy (collision
probability). The obvious question left open by cycle 2 — and recorded as the
fifth direction of `FUTURE_DIRECTIONS.md` ("Flatness Dichotomy for the
Cauchy–Schwarz Knee Floor") — is *how lossy* that floor is on a genuinely
decaying row. The conjecture recorded there was that on a geometric row
`w i = (1 - a) aⁱ` the ratio
(true knee) / (energy floor)
**grows without bound** as `a → 1⁻`, since the true knee is logarithmic in
`1/(1-g)` while the floor is `g²(1+a)/(1-a)`.
This module settles it, and the recorded conjecture is **refuted**:
* `geoRow_knee_le_log_bound`: `k*(g) ≤ 1 + log(1/(1-g)) / (1-a)`;
* `geoRow_knee_ge_floor`: `g²(1+a)/(1-a) ≤ k*(g)` (Cauchy–Schwarz, via
the exact energy `E(a) = (1-a)/(1+a)`);
* `geoRow_flatness_ratio_bounded` / `geoRow_ratio_blowup_refuted`: the two
sides differ by at most the factor `(1 + log(1/(1-g)))/g²`, which depends on
the gate **only** — uniformly in `a ∈ (0,1)`. Both quantities diverge like
`1/(1-a)`, so their ratio stays bounded and the floor is tight up to a
gate-only constant.
For the NET-63 gate `g = 0.98` the constant is explicit and small:
`net63_flatness_constant_lt_six` shows six keys of slack suffice,
`k* ≤ 6 · (g²/E)` for every geometric row.
Consequences for the experimental thread. The energy floor is *not* a weak
bound that only bites on flat rows: on the entire geometric family it is
within a factor `≈5` of the truth at gate `0.98`. Hence a measured
collision entropy really does predict the key budget up to a constant, which
is what the deployment table needs; and the "flatness diagnostic" proposed in
cycle 2 cannot be read off the ratio, because that ratio is bounded.
All statements below are proved from scratch, with complete proofs.
-/
open Bridges.AttentionKneeFlatness
open Finset Bridges.AttentionKneeGeometry Bridges.AttentionKneeEntropyBound
/-! ## 1. The geometric attention row -/
/-! ## 2. The knee of a geometric row -/
/-! ## 3. The dichotomy: the floor is tight up to a gate-only constant -/Formal statement
theorem Bridges.AttentionKneeFlatness.geoRow_ratio_blowup_refuted{g : ℝ} (hg0 : 0 < g) (hg1 : g < 1) :
∃ C : ℝ, 0 < C ∧ ∀ a : ℝ, 0 < a → a < 1 →
(knee (geoRow a) g : ℝ) ≤ C * (g ^ 2 / geoEnergy a) := by sorrySource