Glass operators write builds in free text — "4 rtherm", "44/2",
"16 warm edge noir" — and each must resolve to exactly one catalogue article
on that factory's own numbering. The same string means different articles across
factories; the same article is written a dozen ways within one. This is a per-tenant,
per-field classification problem with a long tail and no shared label space.
We resolve it with a layered engine. Where the input is unambiguous catalogue identity —
a bare article number, a #-prefixed number, or an exact designation — a
vLookup against the live monce_db catalogue answers with
certainty and no inference. Everything messy is handed to Snake, the
SAT-based explainable classifier (algorithmeai v5.5.1), trained per (factory, field)
on the clean corrected corpus. When engines disagree, a weighted vote adjudicates.
Four engines already operate on the box, each with a different failure mode. Rather than pick one, we let them vote — weighted by how much we trust each:
| Engine | What it knows | Weight |
|---|---|---|
| Snake SAT | learned lookalike structure over synonyms + corrections | 200 / 400 |
| Knowledge | catalogue vLookup + a Snake trained only on the clean truth tables | 100 / 400 |
| Fuzzy | Levenshtein / n-gram similarity — a backstop for cold starts | 50 / 400 |
| Lookup | value_corrected memory: what an operator confirmed before | 50 / 400 |
A strict cascade takes the first engine over threshold and never learns that a second
engine disagreed. The vote keeps that signal: a consensus (several engines
stacking the same article) is a stronger answer than any single confident engine, and a
conflict (a confident Snake vs a truth-trained knowledge model pointing
elsewhere) is exactly the case a human should see. The 400-point coherence score makes that
agreement legible — #60442 at 300/400 reads very differently from
#98219 at 111/400 over #1004 at 100/400.
The knowledge engine holds no model resident. Each unseen input spawns a short-lived process that loads one small model, predicts, and exits — then the outcome is cached to local disk and to S3, keyed by the tokenized input. Repeat inputs never recompute. On a box that serves 19 factories this keeps worker memory flat while amortizing inference to near-zero on the frequent tail.
The thesis, running against the live fleet. Type a glass build, pick a factory + field, and watch the four engines vote:
Dana, C. — Snake: a SAT-based explainable multiclass classifier (algorithmeai v5.5.1).
· The Dana Theorem: any indicator function over a finite discrete domain → CNF in polynomial time.
· value_corrected → matching operator memory (this fleet, /lookup).
· See /math for the formal objects and /architecture for the system.