Verdict up front: of the seven "duplicates" in the review queue, two were real. Four were twins-in-name-only carrying different content, and two weren't duplicates in any sense — one of them wasn't even about its own title. A dedup pass run on autopilot would have destroyed roughly 40k words of unique material, including one entire flagship essay.

The category error

"Duplicate" got used to mean three unrelated things. Collapsing them into one queue is what made the queue dangerous.

What actually happened on disk

Following the safe-consolidation protocol on the parent page — rescue first, redirect second, archive never:

• The ChatGPT link off the neuroplasticity stub now lives at the bottom of Neuroplasticity. Stub carries a redirect banner and is safe to archive.
• The davidad link off RL (2) now lives at the bottom of RL, with an explicit note that RL notes is not the same page, so the next cleanup pass doesn't re-flag it.
• quotes was retitled Actin treadmilling as a pseudo-organelle (iPALM, Galbraith lab), given a one-line purpose header, and moved under 🧬 Biology, Longevity & Health where it belonged the whole time.

Nothing merged, moved, or archived beyond that. The four complementary pairs were deliberately left alone.

The part that's actually interesting

The dedup queue is an entropy-reducing process being applied to a system whose entire value is its entropy.

This workspace already contains the argument against itself. The RL notes rank methods by effective search entropy and put MAP-Elites, POET, and novelty search in the top tier precisely because they keep many partial solutions alive in separate niches — and put behavior cloning and tight-KL preference fitting at the bottom for collapsing the repertoire. The stated conclusion there: operate in the high-entropy tiers during pretraining, because a monoculture generalizes badly out of distribution.

A deduplication queue is tight-KL preference fitting for a knowledge base. It optimizes a proxy — title collisions — for a target it cannot see, which is conceptual redundancy. And the proxy is not just imperfect, it's anti-correlated with the thing that makes this archive worth having: the same word landing twice in two unrelated domains is not noise to be squeezed out, it's the high-betweenness signal. multiscale meaning both Stefan-problem interface laws and mitochondrial Ca²⁺ nanodomains isn't a filing mistake. It's a bridge. grace meaning both unsupervised-learning robustness and minimum-jerk trajectories is a live analogy someone could get a paper out of.

So the honest recommendation is not "merge less carefully." It's that the archive should be maintained like a MAP-Elites grid, not like a database with a uniqueness constraint. Niches are the point. Merge only when a cell is genuinely empty — a stub with one dead link and no unique content. That describes two of seven pages here, which is about the hit rate you'd expect once you stop trusting the proxy.

Where I'm actually uncertain

Flagging this properly rather than pretending the audit was clean:

• That the two merges were safe — very likely. Both stubs contained a single URL and no prose. Low risk, and the redirect notes make it reversible.
• That the four "keep both" pairs are genuinely complementary — very likely on content, but only plausible on the question of which should be canonical. I judged substance from page bodies alone. I cannot see inbound links or which page you actually navigate to, and those are the two signals the protocol says should decide. If it turns out you always land on the lowercase multiscale and never the capitalized one, that flips the answer.