The Borgmann Search

Logology · method note

The Borgmann Search

An exhaustive search stops when no word in the list fits. A Borgmann search asks a language model to break that specific wall — and keeps the model out of everything else.

Named for Dmitri Borgmann, who founded logology, the method exists because of a stubborn fact about record word puzzles: the hard ones are not blocked by computation. They are blocked by vocabulary. A nine-letter word square needs eighteen entries that all cross correctly, and somewhere in the grid there is a slot whose letter pattern no entry in your list satisfies. The search has not failed to look hard enough. It has run out of words.

That wall is where a language model earns its place. It knows words the lists do not — real-but-unlisted compounds, names, phrases — and when those genuinely run out it can propose a clueable coinage, something a solver would accept given a fair clue. Grant's asymmetric 8×8 was solved this way. Some records are solvable no other way.

The discipline is in the staging. A model is slow, costly, and confidently wrong in ways local data is not, so the method treats it as the scarce resource: everything that can be done in C++ over local data is done in C++ over local data, and the model is handed one small, pointed question at the end.

search runs exhaustive, in C++ hits a dead slot local candidates corpus match, ranked minus the list we have LLM judges accept / marginal / reject + a clue, batched cache verdicts nothing judged twice, ever separate tier marked in the result for expert review pattern shortlist rulings accepted re-run the search with the judged entries in place last resort: nothing survived judging → invent a clueable coinage already ruled on → answer without touching a model
The whole method. Four of the five stages are local computation; only the oxblood box costs a model call, and the dotted return path above it means most requests never reach that box at all. The dashed red path is the one the name is famous for, and it fires last.

The five stages

Each stage exists to keep work away from the model, or to make the model's answer reusable and reviewable.

StageWhat happensWhy it is here
1 · bottleneck Find the slots that have no fitting entry after constraint propagation, with the exact per-cell letter sets the crossings still allow — g[r][e]de. A wipe-out inside a search that later succeeds is ordinary backtracking, not a bottleneck. And if the structure forces the wall, change the layout rather than buy words.
2 · local first Match those patterns against local corpora, ranked by curated status, number of sources, English probability and frequency. A 21-million-entry corpus holds a large middle tier of real-but-unlisted words. Asking a model to invent what you already have on disk is pure waste.
3 · judge The model sees a short ranked list and returns a verdict per entry — accept, marginal, reject — with a one-line clue. It invents only where local data had nothing. Judge, don't generate. A model asked to produce words will produce plausible ones; a model asked to rule on candidates can be checked. The clue is the test of fairness.
4 · cache Every verdict is stored, keyed by the word, so no string is ever judged twice. Later rounds get cheaper and faster. The cache is also where a human expert's ruling lives, outranking anything a model said.
5 · tier Re-run the search with the judged entries as a separate tier, and mark every one of them in the output: real-but-unlisted, or coinage. A record square resting on a borrowed word must say so. Marking them as a class is what makes expert review possible instead of archaeology.

Two searches, two kinds of wall

The stages do not change between problems; what changes is what counts as a bottleneck, and when the model is allowed to be consulted.

9×9 asymmetric word square

bottleneck = a prefix

The solver explores prefix-first. When its curated list offers no continuation for a partial square, that request is the bottleneck report — no separate pass needed.

Judging happens between runs, not inside them. The solver logs every prefix it could not satisfy; a later batch pass works that log worst-first and feeds the verdicts back. During the search itself the oracle answers only from cache.

15×15 symmetric stack

bottleneck = a slot pattern

A stack is a band of grid-spanning 15-letter entries. Depth 5 is the published record; depth 6 is impossible from the lexicon up, so the tier decides whether a given layout fills at all.

The fill engine reports where it got stuck and distinguishes structurally dead slots from vocabulary-dead ones. Dead patterns are aggregated across many runs, so the commonest walls get judged first.

What the method guards against

Every rule below was added after something got through that should not have.

A corpus is not a dictionary

Web and Usenet text is thick with source code, so the most seductive candidates are programming identifiers that look like compounds: scriptdir, scripturl, indexhtml. Being attested proves nothing. The judging prompt has to name that trap outright, or a model will accept all three with a straight face.

A circular clue is a rejection

The single most effective rule in the whole method: write the clue first and let it decide the verdict. If the clue merely restates the word's own parts — "a directory containing scripts" for SCRIPTDIR — the entry is out. This catches transparent compounds no blocklist would, and once it was in place the models began rejecting their own proposals and citing the rule by name.

Real is not the same as recognisable

A gazetteer is full of genuine places. Asked for names, a model will happily hand back hamlets in Yorkshire and localities in Victoria, each with an honest clue, and a record square built on them is indefensible. The usable test is blunt: if the honest clue would have to be "a village in X", that is itself the signal to reject.

Fluency is not evidence

Asked to rule on BHPERFORM — a string it had coined moments earlier from bhp + perform — a frontier model returned accept with the gloss "the maximum power output of an engine under optimal conditions." That reads like a motoring magazine. It is entirely invented. The expert logologist scored it 0.

The model did not merely over-accept. It fabricated a plausible definition for a word of its own making, and the fluency of the definition was what made it convincing.

Which is why stage 5 is not optional, and why an expert's number outranks any amount of model agreement.

The model belongs in batches, never in the inner loop

Consulted inline, a judging call blocks the search for 25 to 60 seconds. Measured on the 9×9 run: two lanes made 132 such calls in three hours and neither finished its first start word. The same judging done in batches, between runs, costs the same money and nothing in search time. "The LLM is the scarce step" is not a slogan about cost — it is a statement about where it may sit in the control flow.

What it is not

Two things get mistaken for a Borgmann search, and both are weaker.

Padding the word list is the opposite of the method. Junk entries produce junk crossings that no real word can complete, so the pool grows while the solution space does not. In the 15-letter case this is provable: adding 441,000 scraped phrases to the row pool still exhausts to zero at depth 6, because the binding resource was never the count of strings — it was the count of genuinely real ones.

Searching a pre-widened lexicon is closer, and easy to mistake for the real thing. If the judging happened over a band of candidates chosen in advance, rather than over the patterns the search actually got stuck on, then stages 1 and 3 have been quietly disconnected. The search may be running on better vocabulary, but nothing is breaking its walls. The test is simple: when the search hits a slot no entry fits, does anything consult the model about that slot?