Does the fly-brain generator match a standard search, and is anything it finds new?

E136 — The fly has not been used, and nothing found so far is new

Two failures, both the operator's to point out and both correct.

First: the fly generator has not been run once in this session. stage_b.py:295 registers a fly arm — the mushroom-body circuit on the measured MaleCNS connectome, which is the thing this project exists to test. Every run recorded in E114, E115, E123, E124, E125, E133 and E134 used uniform, cem or archive. The connectome generator has been sitting unused while a Gaussian hill-climber did the searching.

Second: nothing reported so far is a discovery. There was no novelty check anywhere in the codebase, so no find has ever been compared against the literature. Adding one (scripts/search/novelty.py, against the equiatomic refractory family from Senkov 2010/2011 plus the long-known binaries) and scoring what has been reported:

Mo0.50 Ta0.50            KNOWN   MoTa,        d = 0.000
MoNbTaTiW  (best 5-el)   KNOWN   MoNbTaTiW,   d = 0.000
MoNbTaVW                 KNOWN   MoNbTaVW,    d = 0.000
Mo0.55 Ta0.41 Cu0.04     KNOWN   MoTa,        d = 0.110

The last line is the one that stings: E133 reported a Mo-Ta-Cu family as something 1705 enumerated points could not reach, and presented it as the payoff of searching continuously. It is Mo-Ta with four per cent copper, inside the same tolerance the generator's own distinct uses to decide two compositions are the same alloy. It was never a new family. E133's breadth claim is withdrawn on that point; the W-Nb and Mo-Nb finds are likewise long-known binaries.

The novelty module reports unchecked rather than novel for anything drawing on elements its reference list does not cover, because an incomplete list makes a find look newer than it is.

Predicted, running the fly arm constrained to five elements or more, with novelty scored:

1. The fly's best in-class find is worse than CrossEntropy's, because E124 and E125 measured it losing on every budget tested and the memory records the generator line as killed on pre-registered criteria.
2. Most in-class finds score known — the equiatomic refractory quinaries are a small, well-explored set, and any search that works will land on them.
3. Any novel verdicts are six-element or Cr/Hf-containing compositions, since those are where the reference list thins out, and they should be treated as unverified rather than as discoveries until checked against a real database.

Falsified if the fly produces an in-class composition that is both novel and passes every correction. That is the outcome the project was built for and it has never once been tested.

E136, the fly's first run in this session:

arm      AUC_Q      distinct   best meV/atom
cem      13.57         34.50            -78
fly       1.00          5.00            -74

Prediction 1 confirmed, but far more narrowly than E124/E125 implied. The fly loses decisively on rate — AUC_Q 1.00 against 13.57, and 5 qualifiers against 34.5 — but its best find is 4 meV/atom off CrossEntropy's, and its best in-class composition, W0.412 Mo0.238 Ta0.210 Nb0.077 Ti0.063 at -79.2, is one CrossEntropy also reached independently. It is a poor rate, not a poor ceiling.

Predictions 2 and 3 on novelty are falsified. I predicted most in-class finds would score known and that novel verdicts would be confined to six-element or Cr/Hf compositions. Most score novel, and they are ordinary five-element refractory alloys.

And that is a limitation of the check I built, not a discovery. The reference list holds only equiatomic alloys, so any non-equiatomic composition is at least some distance from every entry and reads novel by construction. W0.412 Mo0.238 Ta0.210 Nb0.077 Ti0.063 is a non-equiatomic MoNbTaW-plus-Ti; whether the literature calls that a new alloy is a question the module cannot answer, because its tolerance of 0.15 is the generator's own distinct threshold — a search-space measure, not a claim about publication. These should be read as "not in my reference list", never as "new to the field".

So the honest position. The class-constrained search finds non-equiatomic refractory quinaries at -79 meV/atom that the enumeration could not reach and that are not in the reference list. Whether any is genuinely unpublished needs a real database query, which is task 1 below and has not been done. What is established is narrower and still worth having: the high-entropy class was never excluded by physics, only by how I sampled it.

232 passing.