what turned green

we kept the real integration boundary and used authored inputs. the context receipt contains 18 fixture rows and 0 live rows. privacy, byte-budget, required-term and non-application gates passed. assist-review eligibility stayed false.

the reliability receipt records 7 observed receipts, a passing summary, no replay failures and no quality-gold violations. duplicate observations could replay without becoming verified-quality labels.

the model path stopped honestly

the blocked smoke receipt says credentials were unavailable and execution was false. it is a blocker receipt, not passing model evidence. we do not replace the missing model comparison with deterministic green checks.

the next sample has to be real

the promotion ladder requires at least 500 real shadow context packets, including at least 100 large-input packets, before context assist review. report and stop decisions need at least 1,000 paired live decisions. authored fixtures cannot satisfy those gates.

we have a completed bounded boundary campaign and an unexecuted model campaign. we do not have an integration-wide promotion.

what this does not prove

these are frozen source receipts, not new reproductions. no customer-quality, live latency or token-economics win was demonstrated. a successful transport call, a completed turn and a safe privacy projection are different things from a useful answer. this is a draft of the negative evidence boundary, not authorization to enable routing.