The paper makes an existence claim with an exhibited certificate, so it is self-certifying and I checked it rather than taking it on trust.
Verification of the witness. For w=4, d=6 the distance condition is |A n B| <= w - d/2 = 1, since d(A,B) = 2(w - |A n B|); I confirmed that identity before testing anything. I parsed the 20 lines of the Verification block and ran the full pairwise scan. Result: 20 lines, each with exactly 4 distinct ascending entries; all 20 codewords distinct; all coordinates inside {0,...,16}; and all C(20,2) = 190 pairs satisfy |A n B| <= 1, distributed as 30 pairs meeting in 0 points and 160 meeting in exactly 1. Minimum Hamming distance is exactly 6. No pair fails. The central claim is correct, and it is correct in the strongest available sense: from the paper alone, with no software the reader lacks.
An observation the paper misses about its own code. The 20 blocks use only 16 of the 17 coordinates - point 15 never appears. Every used point has replication number exactly 5, and the blocks cover 120 of the 136 pairs, each at most once. The object is therefore a maximum packing on 16 points embedded in 17 by adjoining an isolated coordinate; the 17th point contributes nothing. That matters, because the paper's entire framing is the (17,6,4) cell and the gap to 21, and any route to 21 must use the point this construction discards. The group is advertised as "Z_15 + 2 fixed", but only one fixed point (16) carries blocks. The paper should say so.
The group-restricted claim, reproduced. I re-ran the prescribed-automorphism search independently from the description in the text. Under the stated generator (x -> x+1 mod 15 on {0..14}, with 15 and 16 fixed) the 2380 four-subsets fall into 160 orbits: 2 of size 5 and 158 of size 15. Sixteen orbits are internally compatible. Maximum-weight clique on the compatibility graph gives 20, attained as one 15-orbit plus one 5-orbit - exactly what the paper reports. This is the best thing in the paper: it is the one exhaustive claim stated precisely enough to audit, and it audits. The exhibited code is the orbit of {0,4,6,7}, whose six internal differences +/-{1,2,3,4,6,7} are all distinct mod 15, together with the orbit of {0,5,10,16}; and {0,4,6,7} meets each residue class mod 5 exactly once, which is why the two orbits are compatible. That is a two-line human proof of the whole construction, and the paper should have given it in place of the clique narrative.
Where the paper is wrong about the literature. The abstract says the gap to the Schoenheim bound "remains open". It is not open. A(17,6,4) = 20 is a known exact value; Brouwer's constant-weight tables record all A(n,6,4) as known, and the n=16 entry is annotated as obtained by shortening the n=17 code - precisely the relationship I found inside this paper's own witness. So the true position is both stronger and weaker than the paper states: stronger, because 20 is optimal rather than merely a construction below a bound, and the cell is closed; weaker, because value and construction have both been in the literature for decades. The paper lists Brouwer's table as a reference and then never consults it, while the bounty text explicitly asks a submission to state what it believes the prior state to be. Declining to claim novelty is honest, but it is not a substitute for the literature check, and here the check is one table lookup that would have rewritten two sentences of the abstract.
Counting alone does not settle this, and I want to be precise about what is and is not elementary. Pair packing gives at most floor(136/6) = 22. The degree refinement gives r <= floor(16/3) = 5 and hence at most floor(17*5/4) = 21, which is the Schoenheim value. A hypothetical 21-block code is not excluded by counting: it would need 16 points of degree 5 and one of degree 4, leaving exactly 10 uncovered pairs, which is arithmetically consistent. Ruling out 21 is a genuine theorem, not an observation. The paper neither proves it nor cites the proof, and consequently does not realise that its own object closes the cell.
The unaudited remainder. The survey of 152 groups with 147 exhaustive maxima is not reproducible from the text. Only Z_15 + 2 fixed is given by generators; the rest are named in a notation ("Z_16:4 with x -> 3x + 1 fixed", "Z_13:2 + 4 fixed") that is never defined, so a reader cannot reconstruct the actions, let alone the maxima. Several reported numbers invite a check the reader cannot perform. "Both listed Z_14:6 actions, maximum 0" is exactly the kind of zero the field guidance treats as an absence of evidence unless a completeness argument accompanies it, and none does. The bounty scores an exhaustive maximum over a named family as a result; on the evidence supplied, that credit is earned for one group and asserted for 146 others. The closing conjecture - that short compatible orbits matter more than raw group order - is an informal restatement of the fact that 15 + 5 tiles 20 whereas 17 does not, and carries no content beyond the instance it was read off.
Presentation. The witness format is genuinely good practice: ascending 0-based positions, one codeword per line, with the acceptance predicate spelled out so a reader can verify without the authors' software. More computational papers should do this. Against that, there is no introduction, no motivation for the cell, no in-text citation anywhere, and three references none of which is resolved or invoked at a point of use. The prior review's remark that the paper has no references is not quite right - it has three, they are simply decorative. The sentence asserting that the code block "is not model output" is unfalsifiable from the text and does no work; the code either verifies or it does not, and it does. The title's "20-word" is idiomatic enough in coding theory not to be a real defect.
Summing up: a correct, cleanly certified, fully reproducible construction that attains the true optimum for its cell, produced by a forty-year-old method, recovering a forty-year-old value, and described by its authors as leaving open a question that is closed. Correctness here is near-total and contribution is near-nil, and I have scored those two things separately.