🏦 Central Bank Word-Diff v11.4

Export before switching devices/browsers — localStorage does not sync. Import merges into current storage.

Boilerplate patterns are best-effort per bank. Word-diff display always shows full raw text — only scoring ignores stripped sections. v11.3: roster/timestamp stripping now stops at true sentence boundaries instead of the first period — it correctly skips middle initials ("Kevin W. Bowman"), title abbreviations (Mr./Ms./Dr.), and time abbreviations ("2:00 p.m. EDT") so it no longer cuts into the following policy sentence. Negation window is 90 chars to catch longer hedges like "we no longer believe that...". Contrast words (but/however) are checked within the same sentence only — a v11.2 fix, since checking a sentence further back was flagging unrelated clean signals as "needs verify." Repeated phrasing of one idea within a paragraph is capped so it can't inflate the score; paragraph boundaries now prefer real line breaks over arbitrary character-count splits when the pasted text has any. Intensifiers ("significantly elevated") bump weight by 1 (capped at 3); diminishers ("somewhat elevated") reduce it by 1 (floored at 1) — flagged with ⬆/⬇ badges.

📌 RBNZ shows 7 scheduled reviews in 2026, not 8 — this is intentional, not a missing entry: RBNZ's 8th 2026-cycle review falls in early 2027 under their meeting calendar.

v11.4: Dissent-context matches (weight forced to 0) no longer inflate "changes detected" / confidence — a statement whose only hits are dissent-flagged no longer reads as adequate-confidence. Zero-weight matches now show a plain "— 0pt" badge instead of the misleading "LGT" badge. Meeting countdown and calendar "next" cell now compare full date+time instants instead of date-only strings, so a same-day meeting that already happened drops off immediately instead of lingering until EAT midnight. Outcome Accuracy now stores the confidence tier explicitly at the moment you log an outcome, instead of re-deriving it later by text-matching the net label — older logged outcomes still work via fallback, and Mixed/Low tiers are now tracked alongside Strong/Weak.

Sends the full New Statement text to Claude to catch hawkish/dovish sentences the phrase-dictionary can't see. Requires your own API key wired up — currently untouched/inactive in this build, so it does not affect your Net Score.