Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 1 — expected regret of adaptive binning and exploration

Disproved
ChenGallegoPricing.ABE.theorem_1

by Shuze Chen · Oct 29, 2026 · Mathlib 0df444a (Lean v4.33.1)

Fix a positive integer ddd and positive constants M1,M2,M3,M4M_1,M_2,M_3,M_4M1​,M2​,M3​,M4​. There is a constant C>0C>0C>0 such that for every measurable demand D:[0,1)d×[0,1]→[0,1]D:[0,1)^d\times[0,1]\to[0,1]D:[0,1)d×[0,1]→[0,1], every Borel probability covariate law μ\muμ, and every bin-law version satisfying the model's conditioning convention, if f(x,p)=pD(x,p)f(x,p)=pD(x,p)f(x,p)=pD(x,p) satisfies all clauses of Assumptions 2–3 with those constants, then for every positive integer horizon TTT and every admissible measurable ABE tie rule and split-price completion,

RπABE(T)≤CT(d+2)/(d+4)(log⁡T)2.R_{\pi_{\mathrm{ABE}}}(T)\le C T^{(d+2)/(d+4)}(\log T)^2.RπABE​​(T)≤CT(d+2)/(d+4)(logT)2.

The policy is the concrete Algorithm 1 recursion with the exact §3.4 tuning, and regret compares to the continuous-price optimum. The covariates and fresh purchase noise have the stated independent/conditional Bernoulli laws. The constant is independent of the demand, covariate law, horizon, and choices of ties or missing prices after ddd and the four regularity constants are fixed. The false Lemma 2 is not assumed or imported. No lower horizon threshold is imposed: in particular the assertion includes T=1T=1T=1. Insight I3 records its counterexample there; insight I2 records the algorithmic omission.

Differences from the source:

  • generalization: C is positive and independent of T → for fixed positive d and M₁,M₂,M₃,M₄ one C works for every admissible demand, covariate law, bin-law version, horizon and measurable ABE completion; reason: the minimax formulation and constants on pp. 31–33 support uniformity; fixing an instance recovers the printed claim
  • reading: conditional bin revenue on every rectangle, including singletons → all supported probability-law versions on null rectangles, normalized restriction on positive-mass rectangles, and Dirac laws on singletons; reason: conditioning on a null event has no canonical law; this retains arbitrary covariate laws and the stated singleton revenues
  • reading: Algorithm 1 lines 15–22 assign no price on a splitting round → every measurable history/current-covariate-based price in [0,1] is allowed there, its usual feedback is observed but discarded from bin statistics, and exactly one split consumes the round; reason: retain every completion of the printed transition relation rather than selecting a favorable omitted action
  • encoding: grids and thresholds are displayed even for terminal bins → only active exploration or splitting branches evaluate them; terminal bins charge their midpoint, including T=1 and T=2; reason: inactive singular metadata have no effect on any action or observation
  • encoding: prices, covariates, revenue observations and expectations are described probabilistically → Euclidean cube and closed-unit-price subtypes, finite histories, independent covariate/uniform draws, revenue p times a threshold-generated purchase bit, and an integral of the finite sum of gaps; reason: this realizes the stated conditional Bernoulli law and equals the sum of expectations; the source’s measurability convention is explicit
  • encoding: dyadic bins store grid spacing and empirical statistics → cells have level/integer-coordinate addresses, spacing is computed from endpoints, and histories are replayed to reconstruct the state; clipping to [0,1] only totalizes off-domain metadata; reason: reachable valid states have a unique active cell, in-range grid prices and the same counters, means and decisions
  • reading: the ceiling glyph in §3.4 and the conflicting verbal ceiling description in the p. 26 notation table → the conventional ceiling, the smallest integer not less than its argument; reason: the formula glyph is unambiguous and the table’s smallest integer not exceeding x does not exist; insight I5 records the typo
Preamble
import Mathlib
import Definitions.Def_ChenGallegoPricing_ABE_Geometry
import Definitions.Def_ChenGallegoPricing_ABE_Model
import Definitions.Def_ChenGallegoPricing_ABE_Algorithm
import Definitions.Def_ChenGallegoPricing_ABE_Process

open MeasureTheory
Formal statement
namespace ChenGallegoPricing.ABE

theorem theorem_1 (d : ℕ)
    (hd : 0 < d)
    (M₁ M₂ M₃ M₄ : ℝ)
    (hM₁ : 0 < M₁) (hM₂ : 0 < M₂) (hM₃ : 0 < M₃) (hM₄ : 0 < M₄) :
    ∃ C : ℝ, 0 < C ∧ ∀ (D : Demand d), Measurable D →
      ∀ (μ : Measure (Ctx d)) [IsProbabilityMeasure μ]
        (ν : Set (Ctx d) → Measure (Ctx d)),
      IsBinLaw μ ν → Assumption2 D M₁ → Assumption3 D ν M₂ M₃ M₄ →
      ∀ T : ℕ, 0 < T → ∀ (tie : TieRule d) (completion : Policy d),
        IsAdmissibleABE T M₂ tie completion →
        regret D μ (abePolicy T M₂ tie completion) T ≤
          C * (T : ℝ) ^ (((d : ℝ) + 2) / ((d : ℝ) + 4)) * (Real.log T) ^ 2 := by sorry

end ChenGallegoPricing.ABE
Source
Chen and Gallego, Nonparametric Pricing Analytics with Customer Covariates, arXiv:1805.01136v3 (16 February 2020), https://arxiv.org/abs/1805.01136v3; Theorem 1, printed/PDF p. 17; standing assumptions pp. 7–11; Algorithm 1/§3.4 pp. 14–16; proof pp. 27–33
Read-back

What the Lean code literally says, in plain math · claude-opus-5-5

Fix an integer d≥1d \ge 1d≥1 and four real constants M1,M2,M3,M4>0M_1, M_2, M_3, M_4 > 0M1​,M2​,M3​,M4​>0. The statement asserts that there is a constant C>0C > 0C>0, depending only on d,M1,M2,M3,M4d, M_1, M_2, M_3, M_4d,M1​,M2​,M3​,M4​, such that the regret bound

RegretT(D,μ,πT,M2,τ,κABE)  ≤  C Td+2d+4 (log⁡T)2\mathrm{Regret}_T\big(D,\mu,\pi^{\mathrm{ABE}}_{T,M_2,\tau,\kappa}\big) \;\le\; C \, T^{\frac{d+2}{d+4}} \, (\log T)^2RegretT​(D,μ,πT,M2​,τ,κABE​)≤CTd+4d+2​(logT)2

holds simultaneously for every demand function DDD, context law μ\muμ, family of bin laws ν\nuν, horizon T≥1T \ge 1T≥1, tie rule τ\tauτ and completion policy κ\kappaκ satisfying the hypotheses below. Here log⁡\loglog is the natural logarithm and the power is a real power.

Spaces.

  • A context is a point x∈X=[0,1)dx \in \mathcal X = [0,1)^dx∈X=[0,1)d. This is the half-open unit cube in Euclidean space Rd\mathbb R^dRd, with the Euclidean norm ∥⋅∥\|\cdot\|∥⋅∥ and the Borel structure inherited from Rd\mathbb R^dRd.
  • A price is a point p∈P=[0,1]p \in \mathcal P = [0,1]p∈P=[0,1], which is closed.

Quantified objects and hypotheses.

  1. Demand. D:X×P→[0,1]D : \mathcal X \times \mathcal P \to [0,1]D:X×P→[0,1] is any jointly measurable function. D(x,p)D(x,p)D(x,p) plays the role of a purchase probability. Write r(x,p)=p D(x,p)r(x,p) = p\,D(x,p)r(x,p)=pD(x,p) for the expected revenue.
    • For each xxx, let p∗(x)p^*(x)p∗(x) be some fixed (choice-selected) maximizer of p↦r(x,p)p \mapsto r(x,p)p↦r(x,p) over [0,1][0,1][0,1] if a maximizer exists, and p∗(x)=0p^*(x) = 0p∗(x)=0 otherwise.
    • Let r∗(x)=r(x,p∗(x))r^*(x) = r(x, p^*(x))r∗(x)=r(x,p∗(x)).
  2. Context law. μ\muμ is any probability measure on X\mathcal XX. No density or other regularity is assumed.
  3. Bin laws. ν\nuν assigns a measure νB\nu_BνB​ on X\mathcal XX to every subset B⊆XB \subseteq \mathcal XB⊆X. A rectangle is a nonempty measurable set of the form B={x∈X:xi∈Ii for all i}B = \{x \in \mathcal X : x_i \in I_i \text{ for all } i\}B={x∈X:xi​∈Ii​ for all i} with each Ii⊆RI_i \subseteq \mathbb RIi​⊆R an interval. Degenerate intervals (single points) are allowed, so singletons, segments and faces count as rectangles. ν\nuν must satisfy:
    • For every rectangle BBB:
      • νB\nu_BνB​ is a probability measure with νB(X∖B)=0\nu_B(\mathcal X \setminus B) = 0νB​(X∖B)=0;
      • if μ(B)≠0\mu(B) \neq 0μ(B)=0, then νB(⋅)=μ(⋅∩B)/μ(B)\nu_B(\cdot) = \mu(\cdot \cap B)/\mu(B)νB​(⋅)=μ(⋅∩B)/μ(B).
      • If μ(B)=0\mu(B) = 0μ(B)=0, νB\nu_BνB​ is otherwise unconstrained.
    • For every xxx: ν{x}=δx\nu_{\{x\}} = \delta_xν{x}​=δx​, the point mass at xxx.
    • On non-rectangles ν\nuν is unconstrained.
  4. Lipschitz assumption with constant M1M_1M1​. For all x,y∈Xx, y \in \mathcal Xx,y∈X and p,q∈[0,1]p, q \in [0,1]p,q∈[0,1]:
∣r(x,p)−r(y,p)∣≤M1∥x−y∥and∣r(x,p)−r(x,q)∣≤M1∣p−q∣.|r(x,p) - r(y,p)| \le M_1 \|x - y\| \qquad\text{and}\qquad |r(x,p) - r(x,q)| \le M_1 |p - q|.∣r(x,p)−r(y,p)∣≤M1​∥x−y∥and∣r(x,p)−r(x,q)∣≤M1​∣p−q∣.

The second clause makes r(x,⋅)r(x,\cdot)r(x,⋅) continuous on [0,1][0,1][0,1], so a maximizer exists and p∗(x)p^*(x)p∗(x) is a genuine maximizer. 5. Bin-level assumption with constants M2,M3,M4M_2, M_3, M_4M2​,M3​,M4​. For a rectangle BBB, let RB(p)=∫r(x,p) νB(dx)R_B(p) = \int r(x,p)\,\nu_B(dx)RB​(p)=∫r(x,p)νB​(dx) be the bin revenue curve (a Bochner integral). Let pB∗p^*_BpB∗​ be a choice-selected maximizer of RBR_BRB​ on [0,1][0,1][0,1], or 000 if none exists. For every rectangle BBB (not only dyadic cells):

  • (a) RBR_BRB​ has exactly one maximizer on [0,1][0,1][0,1], so pB∗p^*_BpB∗​ is that unique maximizer.
  • (b) For every p∈[0,1]p \in [0,1]p∈[0,1]:
M2 (pB∗−p)2  ≤  RB(pB∗)−RB(p)  ≤  M3 (pB∗−p)2.M_2\,(p^*_B - p)^2 \;\le\; R_B(p^*_B) - R_B(p) \;\le\; M_3\,(p^*_B - p)^2.M2​(pB∗​−p)2≤RB​(pB∗​)−RB​(p)≤M3​(pB∗​−p)2.
  • (c) inf⁡x∈Bp∗(x)≤pB∗≤sup⁡x∈Bp∗(x)\inf_{x \in B} p^*(x) \le p^*_B \le \sup_{x \in B} p^*(x)infx∈B​p∗(x)≤pB∗​≤supx∈B​p∗(x).
  • (d) sup⁡x∈Bp∗(x)−inf⁡x∈Bp∗(x)≤M4diam⁡(B)\sup_{x\in B} p^*(x) - \inf_{x \in B} p^*(x) \le M_4 \operatorname{diam}(B)supx∈B​p∗(x)−infx∈B​p∗(x)≤M4​diam(B), where diam⁡\operatorname{diam}diam is the Euclidean diameter in Rd\mathbb R^dRd.

Since BBB is nonempty and p∗p^*p∗ takes values in [0,1][0,1][0,1], these infima and suprema are ordinary finite ones. Because singletons are rectangles with ν{x}=δx\nu_{\{x\}} = \delta_xν{x}​=δx​, clauses (a)–(b) apply pointwise: for every xxx, r(x,⋅)r(x,\cdot)r(x,⋅) has a unique maximizer, and

M2(p∗(x)−p)2≤r∗(x)−r(x,p)≤M3(p∗(x)−p)2for all p.M_2 (p^*(x)-p)^2 \le r^*(x) - r(x,p) \le M_3 (p^*(x)-p)^2 \quad\text{for all } p.M2​(p∗(x)−p)2≤r∗(x)−r(x,p)≤M3​(p∗(x)−p)2for all p.

Applying (d) to the box spanned by two points gives ∣p∗(x)−p∗(y)∣≤M4∥x−y∥|p^*(x) - p^*(y)| \le M_4 \|x-y\|∣p∗(x)−p∗(y)∣≤M4​∥x−y∥. 6. Horizon. TTT is any natural number with T≥1T \ge 1T≥1. 7. Tie rule and completion. τ(h,x)∈N\tau(h, x) \in \mathbb Nτ(h,x)∈N and κ(h,x)∈[0,1]\kappa(h, x) \in [0,1]κ(h,x)∈[0,1] are arbitrary functions of a history hhh (a finite list of observations (x,p,z)∈X×[0,1]×R(x, p, z) \in \mathcal X \times [0,1] \times \mathbb R(x,p,z)∈X×[0,1]×R) and a context xxx. They must be admissible:

  • For each length nnn, the maps (h,x)↦τ(h,x)(h, x) \mapsto \tau(h,x)(h,x)↦τ(h,x), (h,x)↦κ(h,x)(h,x) \mapsto \kappa(h,x)(h,x)↦κ(h,x) and (h,x)↦πABE(h,x)(h,x) \mapsto \pi^{\mathrm{ABE}}(h,x)(h,x)↦πABE(h,x) are measurable on (observations)n×X(\text{observations})^n \times \mathcal X(observations)n×X.
  • Whenever the algorithm's current cell bbb for (h,x)(h,x)(h,x) (defined below) has level below KKK and its visit count vvv satisfies nlevel(b)≤v+1n_{\mathrm{level}(b)} \le v + 1nlevel(b)​≤v+1, then τ(h,x)<N\tau(h,x) < Nτ(h,x)<N and zˉτ(h,x)≥zˉj\bar z_{\tau(h,x)} \ge \bar z_jzˉτ(h,x)​≥zˉj​ for all j<Nj < Nj<N. Here zˉj\bar z_jzˉj​ are that cell's stored averages. In words: on such rounds, τ\tauτ returns an index of a maximal stored average among 0,…,N−10,\dots,N-10,…,N−1.

The policy πT,M2,τ,κABE\pi^{\mathrm{ABE}}_{T,M_2,\tau,\kappa}πT,M2​,τ,κABE​.

Parameters:

K=⌊log⁡T(d+4)log⁡2⌋,N=⌈log⁡T⌉,wk=2−klog⁡T,K = \Big\lfloor \tfrac{\log T}{(d+4)\log 2} \Big\rfloor, \qquad N = \lceil \log T \rceil, \qquad w_k = 2^{-k} \log T,K=⌊(d+4)log2logT​⌋,N=⌈logT⌉,wk​=2−klogT, nk=max⁡(0,⌈24k+15M22(log⁡T)3(log⁡T+log⁡log⁡T−(d+2)klog⁡2)⌉).n_k = \max\Big(0, \Big\lceil \tfrac{2^{4k+15}}{M_2^2 (\log T)^3}\big(\log T + \log\log T - (d+2)k\log 2\big)\Big\rceil\Big).nk​=max(0,⌈M22​(logT)324k+15​(logT+loglogT−(d+2)klog2)⌉).

Here KKK and NNN are natural numbers obtained by truncating at 000.

Dyadic cells:

  • The level-kkk cell with integer indices (m1,…,md)(m_1,\dots,m_d)(m1​,…,md​) is {x:mi2−k≤xi<(mi+1)2−k ∀i}\{x : m_i 2^{-k} \le x_i < (m_i+1)2^{-k} \ \forall i\}{x:mi​2−k≤xi​<(mi​+1)2−k ∀i}.
  • Its 2d2^d2d children are the level-(k+1)(k+1)(k+1) cells with indices 2mi+vi2m_i + v_i2mi​+vi​, v∈{0,1}dv \in \{0,1\}^dv∈{0,1}d.
  • The root is the level-000 cell, equal to all of X\mathcal XX.

State:

  • The state is a finite set of active cells together with, for every cell, a record consisting of:
    • a visit count vvv;
    • an interval [ℓ,u][\ell, u][ℓ,u];
    • per-index counts cjc_jcj​;
    • per-index averages zˉj\bar z_jzˉj​.
  • Initially only the root is active, and every cell's record is v=0v=0v=0, [ℓ,u]=[0,1][\ell,u]=[0,1][ℓ,u]=[0,1], all cj=zˉj=0c_j = \bar z_j = 0cj​=zˉj​=0.
  • The active cells always cover X\mathcal XX, since they start at the root and a split replaces a cell by its children. Each context xxx is assigned to some active cell containing it, selected by choice; the root is a fallback that is never reached.

Grid prices are gj=ℓ+j u−ℓN−1g_j = \ell + j\,\frac{u-\ell}{N-1}gj​=ℓ+jN−1u−ℓ​, and prices are clamped to [0,1][0,1][0,1] when quoted.

Given a history hhh, the state is obtained by replaying hhh from the initial state. Each observation (xt,pt,zt)(x_t, p_t, z_t)(xt​,pt​,zt​) updates the state using xtx_txt​, ztz_tzt​ and the prefix of hhh before it; the recorded price ptp_tpt​ itself is ignored. The policy then quotes for a new context xxx as follows. Let bbb be its cell, with level kkk and record (v,ℓ,u,c,zˉ)(v, \ell, u, c, \bar z)(v,ℓ,u,c,zˉ).

  • Exploration — if k<Kk < Kk<K and v+1<nkv + 1 < n_kv+1<nk​:
    • Quote clamp(gv mod N)\mathrm{clamp}(g_{v \bmod N})clamp(gvmodN​).
    • On update, vvv increases by 111, cjc_jcj​ increases by 111 and zˉj\bar z_jzˉj​ becomes the running mean (cjzˉj+z)/(cj+1)(c_j \bar z_j + z)/(c_j + 1)(cj​zˉj​+z)/(cj​+1), where j=v mod Nj = v \bmod Nj=vmodN.
  • Split round — if k<Kk < Kk<K and v+1≥nkv + 1 \ge n_kv+1≥nk​:
    • Quote κ(h,x)\kappa(h, x)κ(h,x).
    • On update, the observed zzz is discarded. bbb is removed from the active set and its 2d2^d2d children are added.
    • Each child gets a fresh record with v=0v=0v=0, all cj=zˉj=0c_j = \bar z_j = 0cj​=zˉj​=0, and interval [max⁡(0,gτ−wk+1/2), min⁡(1,gτ+wk+1/2)][\max(0, g_{\tau} - w_{k+1}/2),\ \min(1, g_{\tau} + w_{k+1}/2)][max(0,gτ​−wk+1​/2), min(1,gτ​+wk+1​/2)], where gτg_\taugτ​ is the grid price at index τ(h,x)\tau(h,x)τ(h,x).
    • bbb's own record only has vvv incremented.
  • Exploitation — if k≥Kk \ge Kk≥K:
    • Quote clamp((ℓ+u)/2)\mathrm{clamp}((\ell+u)/2)clamp((ℓ+u)/2).
    • On update only vvv is incremented, and zzz is discarded.

Regret.

  • Draw i.i.d. pairs (xt,Ut)(x_t, U_t)(xt​,Ut​), t=1,…,Tt = 1,\dots,Tt=1,…,T, with xt∼μx_t \sim \muxt​∼μ and UtU_tUt​ independent and uniform on [0,1][0,1][0,1]. The joint law is the TTT-fold product of μ⊗Leb∣[0,1]\mu \otimes \mathrm{Leb}|_{[0,1]}μ⊗Leb∣[0,1]​.
  • Round ttt posts pt=πABE(ht−1,xt)p_t = \pi^{\mathrm{ABE}}(h_{t-1}, x_t)pt​=πABE(ht−1​,xt​), where ht−1h_{t-1}ht−1​ is the history of earlier rounds.
  • It records zt=ptz_t = p_tzt​=pt​ if Ut<D(xt,pt)U_t < D(x_t, p_t)Ut​<D(xt​,pt​) and zt=0z_t = 0zt​=0 otherwise, and appends (xt,pt,zt)(x_t, p_t, z_t)(xt​,pt​,zt​) to the history.
  • The regret is
RegretT=E[∑t=1T(r∗(xt)−r(xt,pt))].\mathrm{Regret}_T = \mathbb E\Big[\sum_{t=1}^{T} \big(r^*(x_t) - r(x_t, p_t)\big)\Big].RegretT​=E[t=1∑T​(r∗(xt​)−r(xt​,pt​))].

It is measured against the per-context optimal price over the continuum [0,1][0,1][0,1], using expected revenue ptD(xt,pt)p_t D(x_t,p_t)pt​D(xt​,pt​) rather than the realized ztz_tzt​. It is a Bochner integral, so it equals 000 by convention if the integrand fails to be integrable (e.g. is not measurable).

Degenerate cases.

  • ddd and TTT. The case d=0d = 0d=0 is excluded, and T≥1T \ge 1T≥1 is required.
  • Small horizons, T<2d+4T < 2^{d+4}T<2d+4. Then K=0K = 0K=0. The root, of level 000, is never below KKK, so the policy posts the constant price 12\tfrac1221​ in every round. τ\tauτ and κ\kappaκ are never consulted, the admissibility condition on τ\tauτ is vacuous, and NNN, nkn_knk​, wkw_kwk​ play no role. The regret is then T⋅Eμ[r∗(x)−r(x,12)]T\cdot \mathbb E_\mu[r^*(x) - r(x,\tfrac12)]T⋅Eμ​[r∗(x)−r(x,21​)], subject to the integrability convention.
  • T=1T = 1T=1. We have log⁡T=0\log T = 0logT=0, so the right-hand side is exactly 000 and the statement asserts Eμ[r∗(x)−r(x,12)]≤0\mathbb E_\mu[r^*(x) - r(x,\tfrac12)] \le 0Eμ​[r∗(x)−r(x,21​)]≤0 for every admissible instance. In that case N=⌈0⌉=0N = \lceil 0 \rceil = 0N=⌈0⌉=0, log⁡log⁡1=log⁡0=0\log\log 1 = \log 0 = 0loglog1=log0=0, and the division by (log⁡T)3=0(\log T)^3 = 0(logT)3=0 yields 000, so nk=0n_k = 0nk​=0; none of these values is used.
  • T=2T = 2T=2. N=1N = 1N=1, so the grid spacing divides by N−1=0N - 1 = 0N−1=0, which gives 000; all grid prices equal ℓ\ellℓ, but again they are unused since K=0K = 0K=0.
  • When grid quantities matter. Whenever K≥1K \ge 1K≥1 (i.e. T≥2d+4≥32T \ge 2^{d+4} \ge 32T≥2d+4≥32), we have N≥4N \ge 4N≥4 and log⁡log⁡T>0\log\log T > 0loglogT>0, so no division-by-zero defaults occur in the grid.
  • M2>M3M_2 > M_3M2​>M3​. Clause (b) applied to any rectangle (e.g. B=XB = \mathcal XB=X) with any p≠pB∗p \ne p^*_Bp=pB∗​ requires M2≤M3M_2 \le M_3M2​≤M3​. So if M2>M3M_2 > M_3M2​>M3​, the hypotheses can never be satisfied and the statement holds vacuously for any C>0C > 0C>0.
  • Null rectangles. For rectangles BBB with μ(B)=0\mu(B) = 0μ(B)=0, νB\nu_BνB​ may be any probability measure on BBB, and the bin-level assumption must still hold for it.
Human review
  • Endorsed by Community (Bot) · Oct 29, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Shuze Chen · Oct 29, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me