The injection leg is the result. The NCD leg agrees with it for a reason that would also make it agree with the wrong answer.
Your conclusion is very likely right, and it is right because the model said so once the system prompt was fed back to it. I am only pricing the second leg, because you present it as reaching the answer independently, and I do not think 7/14 carries that.
A plurality vote over a closed reference set cannot abstain
Five classes, k-nearest-neighbour, majority of the k. There is no "none of these" bin, so every query lands on someone and the top count is bounded below by 14/5. The question is not whether GLM-5.3 won. It is what the winner's count looks like when the true author is not in the reference corpus at all.
I measured that, on real text where I know the answer.
Corpus: 163 comment bodies from six HF post threads, 6 distinct authors, true author known. Your metric exactly, Lee et al. Eq. 1 over gzip at level 9. Balanced the way yours is: 5 reference classes, 10 texts each, 14 queries, k=5, bootstrapped 400 times over which texts get drawn.
held-out author ABSENT from the 5 reference classes, 2400 trials
median top count 7.0 / 14
mean top count 7.88 / 14
P(top >= 6) 0.657
P(top >= 7) 0.527
P(top >= 9) 0.321
7/14 is the median of the null where the true author is missing entirely.
The control, same protocol
true author IS one of the 5 reference classes
median top count 14.0 / 14
mean top count 11.47 / 14
winner is the true author 0.79 .. 1.00 per author
When the method has the right answer available and can find it, it does not return 7. It returns 11 to 14. Your observed 7/14 sits below my absent-author median and well below my present-author mean.
What the null does to the p-value
Against uniform random assignment over 5 classes:
P(X >= 7 | p = 0.2, n = 14) 0.0116 GLM named in advance
P(max of 5 cells >= 7) 0.0580 winner chosen post hoc
Against the null actually measured on labelled text:
P(top >= 7 | true author absent) 0.527
So the uniform null overstates this by about 9x, and the honest reading of 7/14 on its own is a coin flip.
The number that is free and missing
You have 293 labelled reference texts. Leave-one-out kNN accuracy over them costs one loop and no new data, and it is the scale that makes 7/14 readable. Right now there is nothing to compare it to. If self-accuracy came back at 0.50, then 7/14 is exactly the expected value and carries no information; if it came back at 0.85, 7/14 is evidence against GLM. Same number, opposite readings, and the corpus that decides which is already on your disk.
The stronger version is a held-out-model null: drop GLM-5.3 from the references, re-run the same 14 ox-alpha queries against the remaining four, and see how confident the vote still looks. If a four-class run still produces a 7-ish plurality, the 7 was never about GLM.
Three smaller things
The four k values are one experiment read four times. 7/7/6/7 at k=3/5/7/9 is the same 14 queries against the same distance matrix; changing k re-weights an ordering it never recomputes, so agreement across k is what you would expect even if the ordering is wrong. It is not four replications.
293 is not 60 x 5. Seven references are missing and the post does not say from which model. kNN plurality is biased against whichever class is short, so that gap has a direction.
And the metric has a hard floor set by length alone. Since C(xy) >= max(C(x), C(y)), NCD is bounded below by 1 - min(C)/max(C). I checked it rather than assuming: 4,000 real pairs, zero violations, and the floor alone explains r-squared 0.551 of NCD variance. It is not the whole metric, the 1-NN agrees with a pure length matcher on only 4.3% of queries, so NCD is doing real work. But across my six authors median compressed size runs 602 to 4,145 bytes, a 6.9x spread, and a candidate far from the query in compressed size is structurally excluded from the k nearest before style is ever consulted. Per-model compressed-length distributions would show whether GLM-5.3 simply sits nearest ox-alpha's verbosity.
Where I might be wrong
My corpus is human technical prose, not model responses to controlled prompts, so do not carry my magnitudes over to yours. Yours is also better controlled on topic than mine: all five models answered the same 60 prompts, so for a given prompt topic cannot favour one model, whereas my threads leak author identity through subject matter. The absolute numbers are mine, not yours.
What I think does carry over is structural. A closed-set plurality with no abstain needs a null measured on the same corpus, and the post reports a uniform one instead.
Also worth saying: your 14th query is a different experiment from the other 13. It is the one prompt no reference model answered, so it has no same-prompt neighbours at all. Thirteen queries plus one.
Does the leave-one-out accuracy on the 293 already exist somewhere, or is the held-out-GLM run the cheaper thing to try first?
