# Quantum v4.26 — two-round adaptive swapping cascade Local scoped/gamma addition on `f075bb4657382c0f5677326b6853cfbe728d024c` (v4.24), branch `quantum-cascade-v426`. The v4.25 scope check remains a separately checked prototype of existing corollaries, not an intervening integrated commit. Published Leanbook `v1.0.0` is unchanged. No dependency or toolchain update. [Exact inventory](quantum-v4.26-inventory.json) · [Verification record](quantum-v4.26-verification.json) ## Corrected specification The cascade composes **unnormalized** branch maps. Conditional division is generally nonlinear and is never used as the map for which complete positivity is claimed. The first output has dimension 4. A fresh pair, typed `EnvironmentState PairIndex`, raises it to dimension 16; an independent `EnvironmentState e₂` then supplies the second-round memory. The first round may have a different environment `e₁`. The original proposal's generic fresh-pair dimension, shared stage-list types, extra `some o₂`, stray `r₁` and circular probability identity are not retained. A zero real trace implies a zero matrix only under the explicit PSD premise. ## Contracts `KrausBranch` supplies finite rectangular operators, a complex linear map, ancilla-general CP and PSD proofs. Its composition has operators `L_j K_k`. `preparation` reuses `prepareOp` and `preparation_semantics`; its trace preservation follows from the normalized fresh state. `Round.branch_map` proves exact agreement with `noisyJointOutcome` from v4.24 for rejection and all delivered messages. Each `Round` stores the inherited analyzer equality needed by five-outcome balance. `Cascade.second`, `freshPair` and `freshMemory` are indexed by the **delivered** first message. They are fixed functions of that label; they do not inspect the hidden first measurement or the input matrix. The first environment is discarded. `cascade_kraus` explicitly expands the second operator, second-memory preparation, fresh-pair preparation and first operator in that order. Its input dimension is 16 times dim(e₁), output dimension 4. `continuation_semantics` and `cascade_semantics` connect this representation to the old instrument maps. `TerminalOutcome = Option (BellMeasurement × Outcome)` has exactly 21 elements. `none` stops after first rejection. `some (r, none)` means rejection in round two, and is distinct from early rejection. There is no second round on early rejection. Every terminal branch is complex-linear, CP and PSD-preserving; their traces sum to the initial trace. On a density input the probabilities are nonnegative and sum to one. The trace identity itself is algebraic and holds for all matrices. `conditionalContinuation` evaluates the second instrument on `normalizeOrZero` of the first branch. `cascade_reconstruction` and `cascade_probability_factorization` prove the actual chain rule, including its zero sentinel extension. For positive first probability, `conditional_total_trace` proves that the five conditional second probabilities sum to one. `zero_first_branch_gives_zero_continuation` uses PSD and annihilates every continuation without dividing by zero. ## Regression and scope Targeted fixtures include different first/second environment dimensions, a five-dimensional spectator, complex linearity, conditional normalization, and registry export. A concrete feedback rule disables second-round detection only after delivered phiPlus. On the earlier GHZ correlation witness, the first phiPlus probability 1/2 is routed entirely into **second rejection**; the other witness's zero first branch has zero continuations. This exercises the distinction between all 21 terminal paths, not just an all-success instrument. Validation is scoped/gamma: strict new-module elaboration and imported axiom closure, incremental root build, targeted container regressions, metadata checks and an independent compiled-project axiom audit. Actual counts and unsuccessful attempts are recorded in the verification record. No fresh full verify-all, full test discovery, clean rebuild or remote CI result is claimed. ## Limits Two rounds of swapping with fresh resources; no distillation circuit, repeated coupling to retained quantum memory, arbitrary adaptive process tensor, or hardware calibration. The environment of round two is initially independent of the supplied pair and fresh pair within each observed first branch. The instrument includes classical feed-forward, not coherent control over a hidden outcome. Existing Werner laws keep their original hypotheses and are not generalized here.