# FINDING-relay-v2-study — Relay Membrane v2 study, measured against frozen v1

DMCF Radio · 2026-08-15

The v2 study archive was received, verified independently, and installed as a
**sibling study**. Relay Membrane / Stream **1.0.0 remains frozen, accepted and
mounted** — this document changes nothing about it, which is also what v2's own
charter requires.

## 1 · Technical verification — PASS

Every check was run here, not read from the archive's own audit.

| check | result |
|---|---|
| Acceptance gate (unmodified `validate_stream.py`) | **PASS** — 0.989 / −0.990 / 0.987 / 0.998 |
| All four traces re-emitted from the v2 program | byte-identical to submitted |
| Seed isolation, seeds 0 / 17 / 901 | identical (`aafdcfc4…`) |
| Supplied scheduler | **byte-identical** to the kit (`276093e4…`) |
| `engine.mjs` nondeterminism | zero `Math.random` / `Date` / network |
| Internal manifest, before and after copy | 29/29 OK |
| Instant law | AIRING-BOUND, 33.8% max divergence, 6/6 frames differ |
| Carry | LIVE, digest 0.669034 → 0.705125 |
| Seek equivalence | OK |

The charter's ten technical gates are consistent with what was independently
observed. **v2 is mechanically sound.**

## 2 · What the intervention achieved

Shared-boundary Voronoi territories replace overlapping lobes; the body reads
as one material rather than stacked translucent shapes. Continuous Hermite
evaluation removes the 8-fps stepping. Colour is quantised sRGB with a declared
optical budget. This is a real gain in formal unity and it is visible
immediately.

## 3 · What it cost, measured

### 3.1 The absence notation is half as legible — the serious one

v1 punched a hole through the body to the ground. v2 draws a fused pale
depression. On the same witness (`missing.050`, same key, same seed), measured
on the delivered PNGs:

| | body L\* | absence L\* | ΔL\* | Michelson contrast |
|---|---:|---:|---:|---:|
| **v1** | 149 | 45 | 104 | **0.536** |
| **v2** | 108 | 62 | 46 | **0.271** |

Ground (paper) is L\* 31 in v1 and 27 in v2. So v1's absence sat 14 points from
the ground — it read as *through to nothing*. v2's sits 35 points above it, and
reads as **darker material**.

That is the exact reading the station's honesty rule forbids: *no reading must
not look like a low value*. v2's own artistic gate 4 demands "low focus differs
from missing focus", and the contrast available to carry that distinction has
been halved. This is the one finding I would not promote v2 without addressing.

### 3.2 The population is more uniform, not less

Same instrument, same sealed thresholds, 882 works each. R1 excluded — it was
refuted in the v1 review as sample-size dependent.

| | v1 | v2 | threshold |
|---|---:|---:|---|
| R2 NN duplicate rate | 0.00% | 0.00% | >2% — both clear |
| R3 hue-family entropy | 2.321 | 2.321 | <1.80 — both clear |
| **R4 pressure coverage** | **0.0365** | **0.0445** | <0.10 — **both FIRE** |
| median pairwise, silhouette | 0.701 | **0.596** | descriptive |
| median NN, silhouette | 0.323 | **0.228** | descriptive |
| median pairwise, combined | 1.008 | **0.858** | descriptive |
| within-record silhouette | 0.721 | 0.598 | descriptive |
| median aperture incidence | 0.009 | 0.009 | descriptive |

Silhouette diversity falls **15% pairwise and 29% nearest-neighbour**; combined
diversity falls 15%. This is not a defect — shared-boundary partitions
necessarily produce more similar outlines than freely overlapping lobes, and
that is the intervention working as designed. But it means **v2 buys formal
unity with population variety**, and the trade should be made knowingly.

**R4 is essentially unaddressed.** Pressure coverage moves 0.0365 → 0.0445 and
still fires at less than half the threshold. The one real residual the v1
population review identified is carried forward into v2 intact.

### 3.3 Render cost is 17× higher

| | ms/frame | KB/frame | SVG elements |
|---|---:|---:|---:|
| v1 | 0.16 | 15 | 84 |
| v2 | 2.73 | 148 | 432 |

Fine headless. On the live mount, which re-renders on each of the engine's 8
states per second, v2 rebuilds ~148 KB of SVG eight times a second. It runs,
but it is no longer free, and on a weak client it will be felt.

## 4 · The artistic gates were not run

v2's charter requires **three reviewers**, scoring ten criteria, with a median
of at least 4 and nothing below 3, and it requires blind review for the
anti-pastiche gate. **That did not happen here.** I am one reviewer, not blind,
and I built the measuring instruments — the least independent position
available.

What I can say in that capacity: formal unity (gate 2) and material depth
(gate 6) are clearly improved; motion continuity (gate 5) is improved by
construction; and gate 4 — "low focus differs from missing focus" — is
**measurably weaker than in v1**, per §3.1. Gates 1, 3, 7, 8, 9 and 10 need the
reviewers the charter asks for.

Per the charter's own instruction that technical PASS with artistic FAIL must
be reported as such: **this is technical PASS with the artistic gates
unadjudicated.** v2 stays a study.

## 5 · An instrument failure, recorded

The descriptor extractor initially emitted **882 all-null silhouettes** for v2
and reported success. v2 renamed `geometry.points` to `contour` and changed its
form from pixel `{x,y}` to normalised `{angle,u,v}`; the extractor read absent
fields, produced NaN, and serialised them as JSON null. The comparison in §3.2
was one step from being computed on nothing.

Fixed twice over: the extractor now normalises both contour forms into the same
body space so the two versions are measured with one ruler, and it **throws on
a non-finite silhouette** rather than emitting a null. A descriptor that cannot
be computed is an error, not a value.

## 6 · Disposition

- **v1 1.0.0** — frozen, accepted, unchanged, still the default mount.
- **v2** — installed at `commissions/relay-membrane-v2-study/`, labelled study,
  reachable at `relay.html?v=2` for direct A/B on the same airing.
- **Not published.** The preview carries v1 only.
- **Promotion blocked on:** the three-reviewer artistic gates, and §3.1 — the
  absence contrast — which I would treat as a required fix rather than a
  matter of taste, because it is the station's own epistemic rule and not a
  preference about how a hole should look.