The reference is invisible

src/lib/projection/rule.ts, src/lib/world/changeset.ts

The working model

Observed quantity
        ↓
    Reference          ← the vulnerability sits here
        ↓
   Projection
        ↓
  Visual channel

Retracted hypothesis: "subject-scoped channels communicate subject change cleanly." Falsified across four repositories (change-channels); referenced only as a retraction.

Standing law: a channel's truthfulness depends on the stability of its reference, not on the subject it appears to describe. Height and brightness were never subject-scoped — they are world-normalised and merely looked clean because their divisor happened not to move.

The reference is never drawn. Nothing on screen says what 1.0 means, so its movement is indistinguishable from the subject's.

The two ways a reference lies

failure	definition
spurious	the magnitude did not change; the mark moved
invisible	the magnitude changed; the mark did not move

The strongest example of the second is structural: the largest subject can never appear to grow, because it defines 1.0. Quadrupling the biggest file in the repository is invisible.

Characterising the class

Five mechanisms, plus capped/uncapped variants, over the same HEIGHT_RULE pipeline (normalize → sqrt → scale 26 → offset 2 → floor 0.4). Reporting floor 0.28 = 1% of output range. peak blind counts revisions where the largest subject's own magnitude changed and its mark did not move.

repo	mechanism	invisible%	spurious%	peak blind	out-of-range
enterprise-xi	observed-peak	26.0	0.0	—	0.0%
enterprise-xi	declared-ceiling	32.0	0.0	—	0.0%
enterprise-xi	baseline-revision	26.0	0.0	—	0.0%
enterprise-xi	rolling-avg(3)	26.0	0.0	—	0.0%
enterprise-xi	percentile-95	16.0	0.4	—	4.7%
stable-peak	observed-peak	20.0	0.0	—	0.0%
stable-peak	percentile-95	6.7	60.3	—	3.0%
moving-peak	observed-peak	70.0	34.2	10/10	0.0%
moving-peak	declared-ceiling	5.0	0.0	0/10	2.7%
moving-peak	declared-ceiling+cap	45.0	0.0	8/10	0.0%
moving-peak	baseline-revision	5.0	0.0	0/10	3.0%
moving-peak	baseline-revision+cap	55.0	0.0	10/10	0.0%
moving-peak	rolling-avg(3)	55.0	46.8	5/10	3.0%
moving-peak	percentile-95	5.0	40.0	0/10	3.0%
monorepo	observed-peak	50.0	40.3	3/4	0.0%
monorepo	baseline-revision	15.0	0.0	1/4	11.2%
monorepo	percentile-95	30.0	51.0	0/4	5.2%
rotating-max	observed-peak	0.0	31.9	—	0.0%
rotating-max	baseline-revision	0.0	0.0	—	16.1%
rotating-max	percentile-95	0.0	41.3	—	6.1%

Three findings, and they separate cleanly

1. Spurious change is caused by a moving reference — nothing else. Every fixed reference (declared-ceiling, baseline-revision) scores 0.0% spurious in every scenario, including the ones built to break normalisation. Every moving reference (observed-peak, rolling-avg, percentile) scores 22–60%.

2. Invisible change is caused by the cap, not by the reference. The same reference, capped and uncapped, on moving-peak:

	invisible	peak blind
baseline-revision	5.0%	0/10
baseline-revision+cap	55.0%	10/10

observed-peak carries an implicit cap — the ratio cannot exceed 1 because the divisor is the maximum — which is exactly why its blind spot is total (10/10) and cannot be tuned away.

3. Robust statistics are not safer here — they are worse. percentile-95 moves and clips: 60.3% spurious even in the stable-peak control, where observed-peak scores 0.0%. The maximum of that repository is stable; its 95th percentile is not. Choosing a "more robust" reference imported instability that the naive one did not have.

The trilemma

Three properties. Any mechanism gets two:

	bounded output	no blind spot	no spurious change
observed-peak	✅	❌	❌
declared-ceiling +cap	✅	❌	✅
declared-ceiling uncapped	❌	✅	✅
baseline-revision uncapped	❌	✅	✅
rolling / percentile	varies	❌	❌

A bounded channel cannot represent a value beyond its bound. That is not an implementation defect — a box 28 units tall cannot depict "taller than the tallest box". The blind spot is the price of the bound, and the only question is whether it is declared.

observed-peak is the single worst cell in the table: it is the only mechanism that pays both costs, because its divisor moves and it clips. It is what ships today.

What this does not license

The repair is not yet made, per the standing discipline. Choosing declared-ceiling requires declaring a ceiling, and no observation supports any particular value — it is a stipulation, and on rotating-max an uncapped baseline puts 16.1% of subjects past the end of the channel. Trading a measured defect for an undeclared stipulation is not obviously progress, and which trade is right depends on evidence not yet gathered.

What is established: the mechanism must be named in the display. A reference that moves invisibly is the only reason a truthful pipeline produced 90%+ false firings.

The stronger result: the explanation outlived the visualization

Across every adversarial repository and every channel:

	result
channel false-firing rate	0% → 95%
false firings the model correctly disclaimed	100%
real changes the model mislabelled	0

The channels collapsed. The attribution did not move.

The stable semantic object is not the visualization. It is the explanation.

A channel is a projection through a reference, and the reference is a free parameter that no observation fixes — so a channel's truthfulness is contingent on a choice made elsewhere. The attribution is not: it compares a retained magnitude against itself, and there is nothing in it to destabilise. This is why ChangeSet is the canonical temporal model and the renderer consumes it rather than two worlds.

It also reframes what the product is. The city is a view on an explanation, and it is the weaker half. When the two disagree, the explanation is right.

Instrument correction, recorded

An earlier version of this study declared box height as unit 0–1 with floor 0.01, while HEIGHT_RULE actually emits 0.4–28. The floor was therefore 26× more sensitive than the stated 1% convention, which produced a false 0% miss rate for height on Enterprise XI. Corrected to 0.4–28 / 0.28, the same measurement gives:

	before (bad floor)	after
height FP% (enterprise-xi)	0.0	0.0
height FN% (enterprise-xi)	0.0	37.5
height FP% (moving-peak)	96.9	94.6

The falsification is unaffected — 0% → 94.6% either way. But the "clean" case was never clean: over a third of real byte changes were already invisible, and the declared unit was hiding it. A unit that does not match its values is a threshold hidden inside a label.