Campaign results¶

The full detection-time protocol over the misuse corpus: 300 independent campaigns per (mutant, strategy), censored at a 2000-seed ceiling, strategies random / pct-d2 / pct-d3, false positives measured on every known-correct control at 400 runs per cell. The whole dataset reproduces from one master seed: just dst-campaign-full regenerates these tables, the p̂-versus-PCT-bound section, and the charts (just dst-campaign runs the fast N=100 pilot). The detection-statistics page is the guided reading of this data.

Detection-time campaigns¶

Seeds-to-first-detection per (mutant, strategy): Kaplan–Meier quantiles (a campaign censored at the 2000-seed ceiling stays in the estimate) and the geometric per-seed detection probability with its exact Clopper–Pearson interval. No means — detection times are heavy-tailed and censored.

False positives (negative controls)¶

The harness's violation rate on known-correct code — the gate every external claim stands on. 0 observed violations still carries an exact upper bound, never a bare zero.

p̂ versus the PCT bound (W3)¶

PCT with depth parameter ≥ d guarantees, per trigger-carrying execution, a schedule- detection probability ≥ 1/(n·k^(d−1)). The measured per-seed p̂ is a product: p(workload carries the trigger) × p(schedule realizes it) — so the bound is compared against the conditional p̂_sched = p̂ / p_trigger, with p_trigger taken from the recorded regime structure (the collision pool; the two-rule workload mix). Mutants whose trigger is a fault lottery (crash stream) or an uninstrumented workload-order lottery are excluded — the theorem does not speak about them.

n and k are measured per run (distinct contending tasks; realized ordering-choice ticks), folded per cell as maxima — the largest observed contention gives the lowest, most conservative floor. The formal bound uses the PCT draw range steps=50 for k (the guarantee is over the range the change points are drawn from, not the schedule that happened); the k-tuned floor column restates the same guarantee had steps been set to the measured schedule length — the honest decomposition of any looseness into draw-range slack versus residual conservatism. A cell whose records predate the instrumentation falls back to the structural estimates (workload concurrency; the draw range) and says so.

Multiplicity. The scan checks 31 cells and reports one violation count, so a per-cell 95% interval would not be a 95% claim about the family — under the null that the bound holds everywhere, the chance of at least one spurious flag grows past a coin flip by ~15 cells, and a false alarm here sends a reviewer off to re-derive a correct depth label. Each interval below is therefore computed at 99.8387% per cell (Bonferroni over 31), holding 95% family-wise across the scan.

Flip margin. Uncertainty is propagated through p̂ and through nothing else: p_trigger is a structural constant, several of its values exact combinatorics, but all of them derived from reviewed reasoning rather than measured. Respect holds iff p_trigger ≤ p̂_upper / bound, so each cell carries the exact factor F by which p_trigger would have to be understated for that cell's verdict to flip — no arbitrary perturbation band to calibrate. A cell at F = 40× is immune to any plausible derivation error; one at F = 1.2× is a single reviewed assumption away from a false alarm. Where the flip would need p_trigger > 1 it is unreachable, reported as such rather than as a meaningless factor.

Excluded from the bound comparison (trigger is not a schedule lottery): D4-unmerged-remote-hlc, I3-ack-before-processing, M1-dual-write-shipment, N1-drop-tenant-predicate.

Bound violations: 0 (family-wise 95% over 31 cells).

Reading: for every depth-1 cell the conditional schedule probability sits at ≈ 1 — once the workload carries the trigger, essentially any schedule realizes it, consistent with d=1 meaning zero ordering constraints (for d=1 the bound is 1/n and k drops out). The depth-2 cells are where the bound does real work, and the measured k decomposes their looseness: the formal floor divides by the draw range (steps=50), but the realized schedules are far shorter — the k-tuned floor shows how much of the gap is draw-range slack (recoverable by setting steps to the measured schedule length) versus PCT's residual conservatism. A violation anywhere would have meant a wrong depth label or wrong n/k accounting — the first (unconditioned) pass of this analysis produced exactly such false violations and was corrected to the conditional form above.

n and k carry no interval on purpose. They are per-run measurements folded per cell as maxima — a biased extreme-order statistic, but biased toward the lowest, most conservative floor, which is the direction that cannot manufacture a violation. The absence of an interval there is a decision, not an oversight.