Formula Discovery — candidate formulae the bank does not know

Formula Discovery套语发现

A closed formula bank can only find what it already names — the ontology lock-in problem. This finding runs the search the other way: over the decree-register population of I.PHI (4,434 texts matching a bank seed), it collects recurrent 3–6-token phrasings the bank does not contain, clusters their variants, and asks of each: is it specific to the decree register, or just ordinary Greek? Register lift answers that — how many times more often the phrase appears in decrees than in the background corpus. High lift + high frequency = a likely real formula the ontology is missing. 封闭的套语库只能找到它已命名之物 —— 即本体锁定问题。本分析反其道而行：在 I.PHI 的政令语域群体（4,434 件命中套语库种子的文本）之上，收集套语库所无的 3–6 词复现措辞，聚合其变体，并对每一条追问：它是政令语域所特有，抑或只是寻常希腊语？语域提升度回答此问 —— 该措辞在政令中出现的频率是背景语料的多少倍。高提升度 + 高频率 = 本体所遗漏的、大概率真实的套语。

deterministic · review-only ⚠ A curator queue, not new knowledge. No model is involved — pure counting over folded text (lowercased, diacritics stripped). Every stored attestation offset is re-sliced against the source text at build time (the build fails otherwise), so every sample link below shows a real occurrence. Nothing is auto-promoted into the 58-formula bank; acceptance is a human decision logged in schema/ontology-revision-log.json. Re-run deltas: discovery log. The corpus is Athenian/decree-weighted — lift values inherit that bias. ⚠ 此为编纂者审核队列，而非新知识。 不涉及任何模型 —— 纯粹是对折叠文本（小写、去变音符）的计数。每处存储的见证偏移在构建时均回切原文验证（否则构建失败），故下方每个样本链接都是真实出现之处。任何候选都不会自动收入58 条套语库；是否采纳由人决定，并记录于 schema/ontology-revision-log.json。语料以雅典/政令为主 —— 提升度数值继承此偏向。

The candidate queue候选队列

Sorted by salience (frequency × log lift). Click a column head to re-sort; click a sample id to read the text at PHI.按显著度（频率 × 对数提升度）排序。点击列头可重排；点击样本编号可往 PHI 读取原文。