SCOPE is unbiased
ProvedFormalizedScopeReach.scope_unbiasedAssume a countable vocabulary, the full-history next-token model, almost-sure threshold termination under the original law, and almost surely unequal outcome and threshold times. For every positive sample count, the SCOPE average is integrable and its expectation equals the original probability of an outcome before the threshold. There is no deterministic length bound or finite-mean length assumption.
import Definitions.Def_FormalizedScopeReach_Trials open scoped BigOperators ENNReal open MeasureTheory ProbabilityTheory
namespace FormalizedScopeReach
theorem scope_unbiased {V : Type} [Countable V] (M : Model V)
[IsProbabilityMeasure (originalLaw M)]
(hne : ∀ᵐ x ∂originalLaw M, T_O M x ≠ T_E M x) (n : ℕ) (hn : 0 < n) :
Integrable (𝒮 M n) (iidLaw (originalLaw M) n) ∧
(∫ xs, 𝒮 M n xs ∂iidLaw (originalLaw M) n) = targetProbability M := by sorry
end FormalizedScopeReach
Read-back
What the Lean code literally says, in plain math · GPT-6 (Codex independent sub-agent)
Read-back: scope_unbiased
For every countable type of tokens , every model consisting of a distinguished token , a probability mass function on for each finite token list , and a predicate on finite lists satisfying , the following assertion holds. A timeline is any finite list of tokens, with every set of timelines measurable. Write for a list's length, for its first tokens, and for its token in zero-based position . Call terminal precisely when holds and fails for every integer ; thus stopping must occur at the full list and at no shorter prefix. Give such a terminal list the nonnegative extended-real mass , and give every other list mass zero. Define , where is unit point mass at . Assume that this measure is a probability measure, that is, its total mass is one. For every finite list , define to be one plus the zero-based position of its first occurrence of , or if does not occur; define to be the least natural number for which holds, or if no such number exists. Assume also that for -almost every . For every natural number satisfying , equip the space of tuples with the product measure ; in particular, its coordinates are independent timelines with the same law . For a prefix , let be the real value of the probability mass , and call active exactly when fails and is absent from . Define and , with these sums and division performed in the real numbers. The theorem asserts both that is integrable with respect to , meaning it is almost everywhere strongly measurable and has finite integral of its absolute value, and that , where the probability on the right is converted from a nonnegative extended real to a real number. The comparison is strict, with every natural number less than and false. The stopping predicate is otherwise arbitrary: it has no stated numerical threshold, monotonicity, or dependence restriction, and the probability mass function is specified for every finite history, including histories having zero probability. The presence of makes nonempty, although may be finite or countably infinite. Empty timelines are among the lists over which the definitions range, but are not terminal because stopping at the empty list is forbidden, so they have zero -mass; their score is zero and both their times are . Under the probability-measure assumption, is concentrated on nonempty terminal lists; on each such list, . Consequently the additional almost-everywhere inequality of times requires the event that the first occurrence of is the final token of a terminal list to have probability zero. Lists with no occurrence of are allowed and have ; on terminal lists they do not satisfy the target event. A token's contribution to the score depends on the prefix before that token, so the position of the first , if present, still contributes its prefix probability, while positions after that occurrence contribute zero; on a terminal list without , every position contributes. Only finite timelines enter the measure, and the total-mass-one assumption is imposed rather than following solely from the model fields; a model for which no list is terminal cannot satisfy that assumption. The theorem makes no assertion for , even though the underlying estimator definition is total there and gives zero because its empty sum and real division by zero evaluate to zero. Individual transition probabilities may equal zero or one.
Confirmed by the mission captain (proposal self-audit).