BabyHemingway D2+D3: entities, base-rate gender resolver, rename preset, leak gate passes

2026-09-16 07:55:34 -07:00
parent 9598d0b4a7
commit 03b4a3f62c
5 changed files with 329 additions and 1 deletions
@@ -0,0 +1,44 @@
+{
+  "corpus": "BabyHemingway",
+  "why": "The unigram entity scan structurally CANNOT reach a coined proper noun whose every component is an ordinary word -- the same class as Yarros's `Riders Quadrant` and `Fourth Wing`. The cap/lowercase detector correctly refuses to call `Gran` or `Maestro` a name, so the phrase survives a gate reading zero. It needs a MAP, not a detector: substituting a coined title is a choice about register, not a measurement.",
+  "how_applied": "rename.py runs this AFTER the entity substitution; leak_gate.py audits every recurring capitalised 2-3gram against `allow` and flags anything neither mapped nor allowed.",
+  "scale_note": "⭐ Hemingway needs THREE mapped phrases where Yarros needed 48, and that gap is the whole difference between the corpora: Yarros invented a world, Hemingway named the real one. Correspondingly the allow list here is long and hers was short.",
+  "phrases": {
+    "Gran Maestro": "Primo Servitore",
+    "Unknown Tongue": "Borrowed Speech",
+    "Sin House": "Gladness House"
+  },
+  "phrases_why": {
+    "Gran Maestro": "The narrator's coined honorific for the head waiter at Harry's Bar in Across the River and Into the Trees -- an invented title for one specific character, not an Italian idiom. `Primo Servitore` holds the register: Italian, mock-formal, a title one man awards another in a private joke.",
+    "Unknown Tongue": "A running invention in True at First Light -- the narrator's private made-up language, capitalised and named ('you put up with Unknown Tongue for a long time'). A named thing he invented, so it renames.",
+    "Sin House": "Henry's brothel in Islands in the Stream, a named place ('Do you want the key to Sin House'). Invented proper noun built from two ordinary words."
+  },
+  "tokens": {},
+  "allow": [
+    "New York", "San Sebastian", "San Dona", "South America", "East Africa",
+    "Gulf Stream", "Sand Key", "Coast Guard", "Civil War", "General Staff",
+    "Mau Mau", "Signor Tenente", "Thank God", "For Christ", "For God",
+    "Our Lord", "Our Lady", "Happy Hunting Grounds", "The Colonel", "The Lieutenant",
+    "The Queen", "The American", "The Italians", "The Austrians", "The Germans",
+    "The German", "The French", "The Indians", "The Basque", "The Masai",
+    "The Wakamba", "North Americans", "White Man", "White Hunters", "Kamba Shamba",
+    "Royal Game", "Big Picture", "White Heather", "Bwana Game", "Game Scouts",
+    "Game Scout", "Roman Soldier", "Wine Seller", "The Interpreter", "Comrade General"
+  ],
+  "allow_why": [
+    "Real-world referents and generic English. Every entry was READ IN CONTEXT before being",
+    "allowed rather than classified by shape, which is how the three mapped coinages above were",
+    "separated from these. Some were not obvious:",
+    "",
+    "  Royal Game    a real colonial-Kenyan legal category for protected species, not a coinage",
+    "  White Heather a real Scotch whisky brand, handed over in a square squat bottle",
+    "  Bwana Game    Swahili-English address for the Game Ranger -- a job, not a character name",
+    "  Roman Soldier stage-direction labels (1st/2nd/3rd) in `Today is Friday`, a one-act play",
+    "  Wine Seller   the same play's `Hebrew Wine Seller` -- a generic role label",
+    "  Big Picture   an English idiom, used as one by the Colonel",
+    "  The Interpreter  a character named by role, same class as The Colonel and The Lieutenant",
+    "",
+    "⚠ SENSITIVITY FLOOR: phrases recurring fewer than 5 times were never audited, and are",
+    "therefore neither mapped nor allowed nor reported."
+  ]
+}