The gate we built to fail

What happened when we cashed in the rank check and ran the honest gate. Sequel to BLOG_PHASE1_GSUSTAINED.md; all numbers below are built from committed evidence in data/phase1_gate4_reform.json, full record in ../../PHASE1_GATE4_REFORM_RESULTS.md.

The last post ended on a cautious win. We had a fitness function for ranking candidate blow-up shapes (g_frac, an inviscid growth rate) and although its magnitude was hopelessly on the resolution wall, its rank order was stable: Spearman +0.90 from N=128 to 256. We ran one more expensive check, 256→512, and the rank held again: +0.905. A rank-based fitness looked viable.

So we did the honest thing, and it is worth being explicit about what "honest" meant here, because it is the whole point of the post.

Pre-committing a gate you expect might fail

We had been burned before. The previous currency, ν_crit, printed a clean 6/6 PASS on its pre-committed viability gate, and it was a false pass. The gate had tested the wrong cheat. The real degeneracy (the fitness secretly tracking 1/ω₀²) was caught only afterward, by a substantive correlation the predicate hadn't thought to include.

The lesson we banked: a pre-committed predicate is necessary but not sufficient. So this time we promoted the substantive anti-cheat audits, correlate the winner against the dumbest cheats, from post-hoc diagnostics into first-class gate conditions. If the winning shape is secretly the ω₀→0 corner, or if the fitness is really just a proxy for some trivial scalar, the gate must fail by construction, not pass and let us discover it later during an expensive search.

We wrote the six-property predicate down, had it reviewed, and froze it before running. One condition in particular ("the fitness must carry ω-geometry signal beyond the buoyancy split") we flagged in the frozen doc as likely to fail, because a probe had hinted the relevant partial correlation was only +0.09 against our +0.15 bar. We ran it anyway. A gate that cannot fail is theatre.

It failed. 4 out of 6.

Here is the scorecard on the 40-shape roster (three resolutions, plus an 18-shape controlled split-sweep):

The rank was stable (property 4's Spearman part passed: +0.889, +0.926). It failed for a different, sharper reason, and property 6 failed hard. Both failures say the same thing.

The ghost came back

The de-risk that made us optimistic had fixed the buoyancy split at 0.5: equal vorticity and temperature energy. That pins the scale of ω₀. It was a controlled experiment, and in a controlled experiment g_frac behaved.

But the actual search does not fix the split. It is free to vary it, and freeing it is exactly the knob the old ν_crit cheat used. More temperature energy → more buoyancy forcing → faster vorticity growth, and a smaller ω₀. On the free-split roster, look at who wins (data):

The entire top five is the split → 1 corner. ω₀ → 0. It is the same degeneracy that killed ν_crit, arriving by a different road: ν_crit rewarded it through a small denominator, g_frac rewards it because dumping energy into buoyancy really does make the (vanishing) vorticity grow fastest in rate. The correlation ρ(g_frac, log|ω₀|) = −0.66.

Here is the subtle part, and the reason the gate earned its keep. We also ran the controlled split-sweep (fix a shape's geometry, vary only its split) and it showed a well-behaved interior optimum every time (peaks at split 0.3–0.7, no rail). By that test alone, "the split is not a trivial rail" passes. The de-risk had run essentially that test and concluded the split preference was honest physics.

It was honest physics and a rail. Both. At fixed geometry the best split is interior; but on a free roster where geometry also varies, a high-split shape with the right geometry beats every interior-optimum structured shape. The controlled experiment could not see it. Only interrogating the actual free-search winner could. (Banked lesson, reasserted for the third time: a controlled sub-test passing is not the winner being honest.)

And it's mostly a stopwatch

Two more of property 6's conditions failed, and they matter for anyone thinking of reusing this currency:

• Beyond split, it barely ranks geometry. Partial ρ(g_frac, centroid | split) = +0.11. Control for the split and the fitness stops distinguishing one vorticity shape from another. It is close to a split-meter. (This was the condition we'd flagged as likely-to-fail. It failed.)

• It's largely a formation-time proxy. Partial ρ(g_frac, centroid | t_res) = −0.39. t_res is just when the shape first generates small scales. Control for that, and the structure signal doesn't merely vanish, it reverses. g_frac is, to a large degree, a re-encoding of "how soon does this shape hit the grid limit." So binning the archive on split wouldn't even rescue it; it's a stopwatch as much as it's a split-meter.

The other failure: some "growers" were never growing

Property 4 failed a second axis the rank-check never tested. The rank of g_frac is resolution-stable, but the binary classification "is this shape blowing up" is not. Seven of 37 shapes trip the small-scale guard at N=128, look like growers, and then saturate once the grid resolves them:

advlo_s20:  t_res  2.91 → 3.56 → 4.00   (N = 128, 256, 512)

That is a textbook under-resolution false-positive, the exact artifact the whole project exists to not be fooled by. A rank can be stable while ~19% of what it's ranking is grid noise.

Two currencies, one wall, and what it's actually telling us

Step back. We have now taken two independent fitness currencies, a viscosity-resistance (ν_crit) and an inviscid growth-rate (g_frac), through a pre-committed, cheat-audited viability gate. Both fail, and both fail through the same root degeneracy: on a free genome the "most singular" optimum is the ω₀→0 corner, not blow-up structure. g_frac throws in a bonus problem (it's half a stopwatch).

The deeper reading is a statement about the grid, not the currency. On a uniform mesh, the genuine singular structure forms below grid scale. Everything a scalar fitness can read off the trusted, pre-artifact window is therefore dominated by the degrees of freedom that are resolved (overall amplitude, buoyancy split, formation time) none of which is the thing we're hunting. You can audit away one proxy and the signal just re-expresses itself through the next. Two currencies was enough to see the pattern.

So we are not going to reach for a third uniform-grid currency. That would be pushing harder against a wall, and this project's one firm rule is that a wall is a finding, not a motivation. The honest path to a fitness that tracks real structure (and, not coincidentally, the only path to confirming any candidate singularity) is to stop using a uniform grid: adaptive mesh refinement / self-similar rescaling, so the fitness is measured on resolved structure instead of on the grid's own breakdown.

That's the next decision. As always, the honest framing stands: this is a result about search machinery in a 2D toy model, not about Navier–Stokes, and not a blow-up. But knowing precisely why two reasonable fitness functions both fail (and having a pre-committed gate rigorous enough to prove it rather than guess it) is the kind of negative result the field actually needs. If you're going to evolve initial conditions toward a singularity, here is a wall you'll hit, and here is how to know you've hit it.