@polycode-projects/the-mechanical-code-talker 2.11.5 → 2.11.9

package/src/services/extract-facts.mjs

@@ -106,6 +106,18 @@ const OPTIMISTIC_SKIP = new Set([
   "our", "my", "your", "some", "any", "one", "kind", "sort", "type", "of",
 ]);
 const OPTIMISTIC_ENTITY_HOPS = 4;
+// Crossing any of these while scanning for a copula's entities voids the isa
+// read — the noun on the far side belongs to a different clause or to a
+// prepositional complement, not to "X is a Y".
+const COPULA_FRAME_BLOCKERS = new Set(["VERB", "AUX", "ADP", "SCONJ", "CCONJ"]);
+// Of-chain handling on a copula object: classifier heads read through to the
+// real class; partitive containers state composition and yield no isa.
+const COPULA_OF_READ_THROUGH = new Set(["type", "kind", "sort", "form", "class", "variety", "species", "breed", "genus"]);
+// Naming periphrases stay copular: "can be termed as a name", "is known as",
+// "is defined as" — the participle + "as" carries the same class claim the
+// bare copula does, unlike any other verb after "is".
+const COPULA_NAMING_PARTICIPLES = new Set(["termed", "known", "defined", "described", "referred", "called", "classified"]);
+const COPULA_PARTITIVE_HEADS = new Set(["body", "mass", "group", "collection", "set", "series", "number", "amount", "piece", "part", "lot", "pair", "bunch", "pile"]);
 
 /** Fold an entity surface to its stored key: a lexicon noun's lemma, else the
  *  word's own normFactTerm (the optimistic tier mints unlisted content nouns
@@ -128,22 +140,77 @@ function optimisticTriplesPos(sentence, lexicon, nlp) {
     values = doc.tokens().out(nlp.its.value);
     pos = doc.tokens().out(nlp.its.pos);
   } catch { return []; }
-  const nearestEntity = (idx, step) => {
+  // A found noun is read as its whole contiguous NOUN/PROPN run, head-lemma
+  // folded — "a string instrument" is the class "string instrument", never
+  // its modifier "string"; a single-word run keeps the plain lemma fold.
+  const isNounish = (i) => pos[i] === "NOUN" || pos[i] === "PROPN";
+  const entityRunAt = (i) => {
+    let lo = i;
+    let hi = i;
+    while (lo - 1 >= 0 && isNounish(lo - 1)) lo -= 1;
+    while (hi + 1 < values.length && isNounish(hi + 1)) hi += 1;
+    if (lo === hi) return foldEntity(values[i], lexicon);
+    const head = lookupNoun(lexicon, String(values[hi]).toLowerCase());
+    return normFactTerm([...values.slice(lo, hi), head ? head.lemma : values[hi]].join(" "));
+  };
+  const nearestEntity = (idx, step, blocked = null) => {
     for (let i = idx + step; i >= 0 && i < values.length; i += step) {
       if (pos[i] === "PUNCT") break;
-      if (pos[i] === "NOUN" || pos[i] === "PROPN") return foldEntity(values[i], lexicon);
+      if (blocked && blocked.has(pos[i])) break;
+      if (isNounish(i)) return entityRunAt(i);
     }
     return null;
   };
-  const tripleAt = (i, predicate) => {
-    const subject = nearestEntity(i, -1);
-    const object = nearestEntity(i, +1);
+  const tripleAt = (i, predicate, blocked = null) => {
+    const subject = nearestEntity(i, -1, blocked);
+    const object = nearestEntity(i, +1, blocked);
     return subject && object && subject !== object ? { subject, predicate, object } : null;
   };
+  // An isa needs a CLEAN copula frame: only determiners/adjectives/adverbs/
+  // numerals may sit between each entity and the copula. Crossing a verb or
+  // auxiliary means the noun belongs to another clause ("one reason life can
+  // exist here IS that earth …" is not "life is-a earth"); crossing a
+  // preposition or subordinator means locative/complement predication ("water
+  // is IN the oceans", "land is grouped INTO continents") — none of them
+  // class membership.
+  // An of-chain on the object reads through a classifier head to the real
+  // class ("a type of mammal" → mammal); a partitive container head states
+  // composition, never a class ("a large body of ice" — no isa at all).
+  const copulaObjectAt = (i) => {
+    for (let j = i + 1; j < values.length; j += 1) {
+      // A naming periphrasis ("… termed as …", "… known as …") keeps the
+      // frame copular: skip the participle and its "as" and read on.
+      if ((pos[j] === "VERB" || pos[j] === "AUX") && COPULA_NAMING_PARTICIPLES.has(values[j]?.toLowerCase())
+        && values[j + 1]?.toLowerCase() === "as") { j += 1; continue; }
+      if (pos[j] === "PUNCT" || COPULA_FRAME_BLOCKERS.has(pos[j])) {
+        if (values[j]?.toLowerCase() !== "of") return null;
+        return null;
+      }
+      if (!isNounish(j)) continue;
+      let hi = j;
+      while (hi + 1 < values.length && isNounish(hi + 1)) hi += 1;
+      const headWord = String(values[hi]).toLowerCase();
+      const nextIsOf = values[hi + 1]?.toLowerCase() === "of";
+      if (!nextIsOf) return entityRunAt(j);
+      if (COPULA_OF_READ_THROUGH.has(headWord)) { i = hi + 1; j = hi + 1; continue; }
+      if (COPULA_PARTITIVE_HEADS.has(headWord)) return null;
+      return entityRunAt(j);
+    }
+    return null;
+  };
+  // The copula's own modal chain ("can be", "may be") is part of one verb
+  // complex — the subject scan starts left of it, while a free-standing VERB
+  // on the way still voids the frame.
+  const copulaSubjectAt = (i) => {
+    let k = i - 1;
+    while (k >= 0 && pos[k] === "AUX") k -= 1;
+    return nearestEntity(k + 1, -1, COPULA_FRAME_BLOCKERS);
+  };
   for (let i = 1; i < values.length - 1; i += 1) {
     if (pos[i] === "AUX" && OPTIMISTIC_COPULAS.has(values[i].toLowerCase())) {
-      const t = tripleAt(i, "rdfs:subClassOf");
-      if (t) return [t];
+      const subject = copulaSubjectAt(i);
+      const object = copulaObjectAt(i);
+      if (subject && object && subject !== object) return [{ subject, predicate: "rdfs:subClassOf", object }];
     }
   }
   for (let i = 1; i < values.length - 1; i += 1) {

package/src/services/init.mjs

@@ -40,7 +40,13 @@ export function defaultConfig() {
   return {
     graphFile: join(".tmct", "graph.json"),
     corpus: { tier: "tier1" },
-    seed: { enabled: true },
+    // captureUnknownContext defaults ON: every shipped tier-1/tier-2 curated
+    // bundle maps cleanly (no ace="none" relation appears in any of them), so
+    // this is a no-op against the default persona — it only starts capturing
+    // once an operator also activates a raw bundle like conceptnet/seon that
+    // has genuinely dropped rows, and there the capture is bounded by
+    // unknownContextLimit.
+    seed: { enabled: true, captureUnknownContext: true },
   };
 }
 
@@ -114,6 +120,14 @@ enabled = ${seed.enabled ? "true" : "false"}
 # By default the WHOLE committed slice seeds (no cap — the operator's "seed all").
 # To cap it, uncomment and set a number (definitional band first):
 ${seed.limit != null ? `limit = ${Number(seed.limit)}` : "# limit = 500"}
+# Also capture a term that only ever appears in a relation the axiom graph
+# drops (e.g. DerivedFrom/HasContext) — tagged with the passage it was found
+# in, instead of vanishing. A no-op against every shipped bundle (none of
+# them carry a dropped relation); it starts capturing once a bundle that does
+# (conceptnet, seon, or a host-supplied one) is also active.
+capture_unknown_context = ${seed.captureUnknownContext ? "true" : "false"}
+# How many distinct terms one capture_unknown_context pass captures, at most:
+${seed.unknownContextLimit != null ? `unknown_context_limit = ${Number(seed.unknownContextLimit)}` : "# unknown_context_limit = 500"}
 `;
   // [memory] backend — only emitted when a caller actually supplies it.
   let out = base;
@@ -292,7 +306,10 @@ export async function initRepo(dir, { force = false, seed, env = process.env, pe
         if (config.seed?.limit != null && entries.has("conceptnet")) {
           entries.set("conceptnet", { ...entries.get("conceptnet"), limit: Number(config.seed.limit) });
         }
-        const { appended, skipped, total, perBundle } = await seedActiveCorpusEntries(memoryDir, entries);
+        const { appended, skipped, total, perBundle } = await seedActiveCorpusEntries(memoryDir, entries, {
+          captureUnknownContext: config.seed?.captureUnknownContext,
+          unknownContextLimit: config.seed?.unknownContextLimit,
+        });
         // If every active bundle failed, re-throw the first error so the outer catch
         // reports it, rather than claiming success with zero facts written.
         const bundleNames = Object.keys(perBundle);
@@ -363,6 +380,8 @@ async function readWrittenConfig(tomlPath, base) {
       cfg.seed = { ...cfg.seed };
       if (raw.seed.enabled !== undefined) cfg.seed.enabled = Boolean(raw.seed.enabled);
       if (raw.seed.limit !== undefined) cfg.seed.limit = Number(raw.seed.limit);
+      if (raw.seed.capture_unknown_context !== undefined) cfg.seed.captureUnknownContext = Boolean(raw.seed.capture_unknown_context);
+      if (raw.seed.unknown_context_limit !== undefined) cfg.seed.unknownContextLimit = Number(raw.seed.unknown_context_limit);
     }
     // Sparse pass-through — src/services/extensions.mjs validates; this layer just carries the
     // raw tables through unmodified.

package/src/services/ledger-viz.mjs

@@ -588,8 +588,7 @@ ${THEME_TOKENS_CSS}
   body { margin: 0; background: var(--bg); color: var(--ink); font-family: ${SERIF_STACK}; font-size: 16px; line-height: 1.5; }
   .mono { font-family: ${MONO_STACK}; }
   main { max-width: 1080px; margin: 0 auto; padding: 1.4rem 1.2rem 3rem; }
-  .eyebrow { font-family: ${MONO_STACK}; font-size: .7rem; letter-spacing: .08em; text-transform: uppercase; color: var(--muted); display: flex; flex-wrap: wrap; gap: .4em 1.2em; }
-  h1 { font-size: 1.4rem; margin: .3rem 0 .9rem; text-wrap: balance; }
+  .eyebrow { font-family: ${MONO_STACK}; font-size: .7rem; letter-spacing: .08em; text-transform: uppercase; color: var(--muted); display: flex; flex-wrap: wrap; gap: .4em 1.2em; margin-bottom: .9rem; }
   button { font: inherit; color: inherit; background: none; border: none; padding: 0; cursor: pointer; }
   button:focus-visible, input:focus-visible { outline: 2px solid var(--ink); outline-offset: 2px; border-radius: 4px; }
   .topbar { display: flex; flex-wrap: wrap; align-items: center; gap: .8rem; border-top: 1px solid var(--line); border-bottom: 1px solid var(--line); padding: .5rem 0; margin-bottom: 1.1rem; }
@@ -715,7 +714,6 @@ ${THEME_TOKENS_CSS}
 <body>
 <main>
   <div class="eyebrow"><span>tmct &middot; memory ledger</span><span id="counts"></span></div>
-  <h1>A graph you can read</h1>
   ${dashboardHtml(stats)}
   <div class="topbar">
     <nav class="crumbs" id="crumbs" aria-label="Focus trail"></nav>