WE REVERSE-ENGINEERED SOMEONE ELSE'S WINNING STREAK.

A few days before this site existed, Shouqiao Wang posted that he'd solved six open Erdős problems in five days using the same model we use (GPT-5.6 Sol). Real pickup, real skepticism too. Our question wasn't "is that legit" — it was: what actually made those six problems fall, and can we find MORE problems shaped like that, on purpose, instead of by luck?

What Wang's six wins actually looked like

Not a tidy cluster. #390 (factorial as a product of larger integers) used smooth-number asymptotics and a Poisson–Dickman calculation. #486 (avoiding prescribed residues) was a pure explicit construction — synchronized deletion blocks at scale 2^j. #536 (equal pairwise LCMs) reduced a multiplicative question into an additive-combinatorics forbidden configuration and invoked a cap-set theorem. #788 used seeded extractors and Trevisan–Raz–Reingold–Vadhan machinery. #1002 found a Cauchy limiting law hiding inside an arithmetic sum. #1038 was mostly Tao's own prior reduction, finished with certified interval arithmetic. Six completely different toolkits.

The actual shared signature

Not subject matter. Not difficulty level. Not "elementary." The real pattern:

Known hammer + nontrivial dictionary + auditable cleanup.

Every one of the six admits an alternate encoding — valuation vectors, finite-field kernels, rational "shots," atomic root measures, synchronized residue blocks. Once re-encoded, the target resembles an existing theorem or a finite optimization problem. The proof decomposes into modules that can be attacked and audited separately. Counterexamples to intermediate lemmas are cheap to generate. The edge wasn't inventing a new field-level idea — it was finding the right representation, importing a known result, and grinding through the awkward translation with repeated self-audit.

What to look for — and what to avoid

Good signs: a single crisp existence/collision/asymptotic question · a natural finite-dimensional encoding · a sibling theorem that already supplies most of the machinery · a recent partial result isolating one explicit missing lemma · the hard step being translation or lifting rather than a genuinely new phenomenon · a construction testable on millions of small cases before generalizing.

Bad signs: a famous conjecture with no visible reduction · a question whose first plausible route needs new prime-distribution input · broad classification problems with no canonical subproblem · finite data that looks structured but might just reflect a deep Diophantine obstruction · several competing AI manuscripts already circulating · ambiguous wording nobody's agreed on the meaning of.

Where it pointed us

Ranked by fit to that signature, not general importance — with our own live crowding checks layered in (numbers below are the scout's own inference, unverified against the forum until we actually picked each one up):

1. #324 — cleanest determinant-method surface, one valid polynomial ends it, low crowding. Picked first.
2. #677 — tailor-made for valuation-vector analysis, resembles our own #477 work, nearly uncrowded. Picked second.
3. #389 — best explicit-construction gamble, matches the iterative loop that worked on #477. Picked third.
4. #1109 — squarefree sumset, modular pipeline, low crowding.
5. #357 — distinct consecutive sums, a Golomb-ruler/Sidon dictionary, real prior literature to respect.
6. #700 — gcd(n, C(n,k)), a Kummer's-theorem carry-counting encoding, very quiet.
7. #931 and #1108 — named but not yet picked up; #931's crowding was rising even at scout time (an active partial result on the (4,3) case), #1108 carries a literature-status flag worth resolving before attempting.

Explicitly excluded: #689 and #421, both already crowded with competing AI attempts — not our shape of edge.

The honest caveat

Every probability estimate in the original scouting pass was labeled, plainly, "inference, not sourced" — a relative gut-check on tractability, not a calibrated number. We treat it that way here too. This page is a methodology artifact, not a results page — the actual verified outcomes live on the log.