# FINDING-relay-population — Relay Membrane / Stream 1.0.0 Preregistration `PREREG-relay-population.md` verified at `a72305eea2d5df504b6dda05892f2359a765278f2f597d844a74520ab1710dbe`. Thresholds were sealed before any measurement; this run applied them once. ``` prereg verified: a72305eea2d5df50… (thresholds locked) sample: 882 observed · 240 repeated-airing works R1 silhouette near-duplicates : 98.07% threshold >5% FIRES median NN contour distance : 4.04% of own radius R2 combined NN duplicates : 0.00% threshold >2% clear median NN / median pairwise: 0.506 R3 hue-family entropy : 2.321 bits of 2.322 threshold <1.8 clear realized-hue entropy : 3.954 bits of 5.170 family counts : [180, 177, 181, 168, 176] R4 pressure-field coverage : 0.0365 threshold <0.1 FIRES IQR px 0.210 of 1.08 · IQR py 0.161 of 0.86 R5 within-record / between : 1.778 threshold <0.25 clear records with ≥2 passes : 40 AMI(duration, 4-means) : 0.012 DATA-INERT FIRED: R1, R4 (2 of 5) VERDICT: REPETITION DEMONSTRATED — version 1.1 is authorised, scoped to data-derived morphological regimes (PREREG §4). by stratum (median NN distance, share of sample): duration brief n=115 medNN=0.537 H_fam=2.23 extended n=18 medNN=0.476 H_fam=2.26 long n=147 medNN=0.498 H_fam=2.32 short n=602 medNN=0.489 H_fam=2.32 completeness complete n=223 medNN=0.488 H_fam=2.30 partial n=553 medNN=0.491 H_fam=2.32 sparse n=106 medNN=0.549 H_fam=2.23 topology 5 n=188 medNN=0.498 H_fam=2.31 6 n=178 medNN=0.489 H_fam=2.31 7 n=181 medNN=0.489 H_fam=2.30 8 n=168 medNN=0.501 H_fam=2.32 9 n=167 medNN=0.502 H_fam=2.30 pass position body n=864 medNN=0.496 H_fam=2.32 closing n=9 medNN=0.502 H_fam=1.84 opening n=9 medNN=0.536 H_fam=2.20 ``` --- ## Post-run validity analysis The preregistration fired **R1 and R4**, which reads as REPETITION DEMONSTRATED. Before acting on that, each firing criterion was tested for whether it measures what it claimed to. This is a **validity** question, not a threshold question: no line below was moved, and the run above is reported exactly as it computed. ### R1 is REFUTED as constructed R1 counts works whose nearest neighbour's contour lies within 5% per coordinate. Nearest-neighbour distance shrinks as a sample grows, so the statistic is **not scale-free in N**. Re-measured on the *same, unchanged* population at increasing sample sizes: | N | R1 rate | median NN | |---:|---:|---:| | 60 | 46.7% | 0.405 | | 120 | 65.8% | 0.375 | | 250 | 80.0% | 0.359 | | 500 | 94.2% | 0.334 | | 882 | 98.1% | 0.323 | The population is identical in every row. R1's rate more than doubles purely because more of it is being looked at. **R1 therefore measures sample density, not repetition**, and its firing carries no information about the work. Corroborating, from the same descriptors: - the *closest* pair in 882 works still differs by **7.0% maximum radial deviation** — a visibly different contour, not a duplicate; - **0% of records** have their own re-airings inside R1's duplicate radius; - same-record re-airings average **0.718** silhouette distance, which is *larger* than the population's median pairwise distance of **0.701**. Two airings of one record differ more than two random works do. That last figure is the direct refutation of the repetition worry, and it is the same quantity R5 measured independently and cleared at 1.778. ### R4 stands Pressure-field coverage uses an interquartile range, which is not a nearest-neighbour statistic and does not drift with N: | N | 60 | 120 | 250 | 500 | 882 | |---|---|---|---|---|---| | coverage | 0.0477 | 0.0450 | 0.0448 | 0.0440 | 0.0365 | **R4 is a real observation.** The middle half of works places its pressure centre within about 19% of the available width and 19% of the available height — roughly 3.7% of the compositional field. The pressure centre is the work's main compositional variable and it is clustered near the middle. ### The AMI diagnostic is not supported as constructed `AMI = 0.012` triggered the `DATA-INERT` label, which would be the more serious finding. It is not supportable here. The feature vector is 64 silhouette dimensions against 6 others, so the clustering is essentially a clustering of contours — and contour derives from `structureSeed`, i.e. from the *airing instant and record identity*, which is not duration. The test therefore shows only that **duration does not predict contour**, which is true, expected, and not evidence that the data fails to shape the work. The channels demonstrably do shape it: the acceptance gate scores the four load-bearing mappings at 0.990–0.998. **`DATA-INERT` is withdrawn as unsupported by this instrument.** ## Outcome With R1 refuted, **one** criterion of five stands. PREREG §4: > One firing is published but does not authorise 1.1. A single marginal > statistic is how a programme talks itself into work it already wanted to do. **VERDICT: REPETITION NOT DEMONSTRATED. Relay Membrane / Stream 1.0.0 stands unchanged and the question is closed.** No version 1.1 is authorised. R4 is recorded as a real, measured property of the work — pressure centres cluster centrally — available as evidence if a future version is ever opened for another reason. It is not, on its own, grounds to open one. ## What this cost, and what it bought Two instruments were built for this review and one of them was wrong. R1 was a plausible-sounding criterion that would have authorised a rewrite of an accepted work on the strength of a sample-size artifact. It was caught because the thresholds were sealed first, the run was reported before interpretation, and the criterion was then tested against the only thing that can refute a measurement — the same population measured again under a different condition. The v1 preregistration was also voided before any run, for a ceiling model that could not be built. Both failures are recorded at the same weight as the result, per the station's standing practice.