The replication gate implemented beside design()¶

replication() and replication_gate() in sphragis/experiment/decomposition.py read the GitHub family as registered on 2026-10-04. The family takes only Apache, LLVM, .NET and Grafana, in that order; a Gerrit organization is refused there, as design() refuses a GitHub one, and a registered organization outside the members is reported as not collected.

The members are a constant, REPLICATION_MEMBERS, set by the commit that freezes them at Stage 1, never an argument; until then the gate refuses to read the family, and afterwards it refuses a cell or a bound for any organization outside them, so none joins later. Each member's bound comes by level from its own partition simulation (sensitivity_bounds), must be a positive finite number, and must equal the bound its cell was computed under, which h1_over_partitions now records in the cell. The gate also refuses fewer than 1,000 resamples, a resample count that puts the 0.0125 bound between draws, a missing field, and a p-value that disagrees with its interval (below 0.0125 exactly when the lower bound is above zero). It reads each verdict from the cell's two-sided 97.5% interval and r from the partial conjunction at one-sided 0.0125. The 97.5% level is registered as fixed, not as a Holm step. An empty family has no r.

Each cell now records the organization it was computed for. The checks every reading of an H1 cell needs are one function, require_readable in across.py, run by both this gate and across_organizations, now in cells.py beside the level-key helpers and the registered test-read constants, so the confirmatory gate does not import its checks from the readings beside it: filed under the organization it was computed for; its fields; a whole resample count of at least 1,000, whole at the one-sided level; a p-value in [0, 1]; a finite interval at the level, low at most high; and a p-value that agrees with the interval. Level keys in both gates, h1_over_partitions, sensitivity_bounds and detectable_effects pass through one helper, so a cell or bound read back from JSON (string keys) gives the same reading, and two keys naming one level are refused. The Gerrit gate now refuses a bound that is not a positive finite number too.

scripts/partition_pilot.py takes --replication for a GitHub member, reading it at 97.5% alone whatever the Gerrit Holm family's size, and refuses the flag for any other organization and its absence for a member, and, once members are frozen, refuses an organization outside them. It refuses --hypotheses beside it, and, before any run is loaded, a resample count the gate would refuse; a bound the gate would refuse is refused before the cell is computed. Any test-window read, on either platform, now needs --sensitivity, --sizing and the registered 10,000 resamples at bootstrap seed 7, as it needed --planted, and a member's only after the freeze (registered here as "Reading the test window"). Each run's window is read from every half's source with the parser check 5 uses, so a run with one half on the test window is never taken for development. pilot_sizing refuses a test-window reading or another organization's pilot as the source of K, and on a test-window read a pilot that does not name its organization: the committed OpenStack and Wikimedia pilots predate the field, so they are rerun before their test reads. The pilot refuses another organization's simulation as the source of bounds. The report records the window, the pilot K came from (k_from), the simulation and the K it ran at (sensitivity), and the bootstrap seed; require_test_read in cells.py checks all of them, and that both sources are at the read's K, on the report the pilot is about to write (before any verdict is printed) and again in the gate, so the two hold one rule. The gate adds what is the family's own: a one-cell simulation, 97.5% alone. Decided here: a member's bound comes from a simulation of one cell at 97.5% (--h1-cells 1), since it is read alone, not as one of an intersection; registered-decisions.md says so, and the gate checks the report's simulation is of one cell. Without --replication the pilot takes only the four registered Gerrit organizations. The partition simulation reads its levels from the pilot, so the member's bound is at the same level. The pilot's report no longer carries its own bounds, which the cell now holds. A test runs the script on synthetic runs and reads its report through the gate, and another writes a cell to JSON and reads it back for each of the three verdicts.

Still chosen when the test window is read, and so open until a committed read plan fixes them (IMPL): the Holm family's size (--hypotheses), the spread target and cell count the bound is read at, which development pilot calibrated the simulation, and the H2 reading's resamples and seed. None can be used before in-principle acceptance, since the test window stays sealed until then. Each run's halves must now all be built from its admissible partition, not only the first.