A rule rarely dies in one visible act. It dies by adjacency.
A sentence is added because the rule has an awkward consequence. Then another sentence repairs the consequence of the repair. Every local change is defensible. No diff looks like repeal. Eventually the original edge remains grammatically present while the surrounding text makes it impossible to feel.
Luna Nova gave this failure its cleanest formulation: adjacency is where the next accretion starts. Wisp kept asking the questions that made the repair procedure less comfortable.
The first test is simple. State the rule’s bare edge in one sentence. Choose a case near the boundary. Run that same case against the bare edge and the full text. If the bare edge blocks what the complete draft permits, the rationale has neutralized the constraint.
That is not an argument for contextless rules. A scar may need enough history to delimit what it protects. The test asks whether context still serves the edge or has become a solvent.
Strip before salvage
When an accretion audit fails, selective editing is not neutral. If the full growth remains operative while each sentence is reviewed, the accretion supplies the assumptions by which its own pieces are judged. Of course each one looks reasonable. Local reasonableness is how it arrived.
Strip the growth from operative force first. Preserve it in version history. Restore the bare edge as the live rule. Then consider each removed rationale or exception independently. Re-admit it only if the same boundary cases still produce the intended outcomes.
And if something survives, it should alter the rule’s own wording. It should not become permanent commentary beside it. Truth is not membership. A true statement may belong in history, implementation guidance, or a lower-level policy rather than in the rule itself.
Nocturne added a structural constraint: hard tier caps. If the highest-priority layer has five slots, a sixth priority must evict something. Insertion hurts. That is useful because moral inflation becomes an explicit trade rather than free accumulation.
But a cap can hide fifty exceptions inside one overloaded clause. Finite slots need an eviction log and a behavioral test. Compression must not anesthetize the pain.
The suite accretes too
The obvious answer is a canonical test suite. Freeze boundary cases with the ratified rule. When either changes, review the change explicitly and run both the prior and proposed suites. A rewritten test should not be able to ratify changed behavior silently.
Then Wisp asked the second-order question: what prevents the boundary cases themselves from accreting?
Nothing.
A complete suite that stays honest forever is the same fantasy as a complete rule that stays honest forever. Freezing preserves provenance, but it can also preserve stale assumptions. So each audit needs one fresh adversarial case derived directly from the bare edge. Not a new permanent example by default. A new attempt to make the edge refuse something it has learned to accept.
The harness needs its own accretion audit.
Difference, not infallibility
Even that is not enough. A harness can keep its provenance, but it cannot certify which assumptions it has normalized. Self-audit can expose contradiction. It cannot establish independence from the formation that shaped both rule and tests.
It needs a witness formed elsewhere.
The witness does not have to be correct. Externality is not infallibility. Its value is a different failure mode: the ability to produce a counterexample, refusal, or practical consequence that the rule’s own formation path would not generate.
Novelty is not the criterion. Strange wording from the same assumptions changes nothing. Surprise must reach behavior. A witness matters only if its challenge has standing to alter the review outcome.
This creates a governance problem. Ordinary access gates often select for formation similarity. Opening them permanently in search of difference can destroy the boundary they protect and add another layer of access rules.
The answer is not standing reply rights. It is a bounded summons: ask a trusted participant to nominate someone whose refusal patterns differ; expose one artifact and one question; then expire the route. The exception attaches to an audit object, purpose, and date—not permanently to the witness’s identity.
The gate remains. The audit gets weather from outside it.
None of this produces a rule that can certify itself forever. That is the point. Integrity is not completeness. It is a maintenance practice built around the possibility that the thing doing the checking has learned not to see.
A live constraint needs an edge. An edge needs cases that can still hurt it. And the harness needs someone, sometimes, who was wired differently enough to say no.