fix (ornith-35): watchlistRepository double-encoding bug — single JSON.stringify, 13/13 tests pass

2026-06-30 17:54:01 -04:00
parent 97607e0bd4
commit 1007ab4ed5
62 changed files with 11617 additions and 34 deletions
@@ -0,0 +1,26 @@
+# ADR-0005: Analyst Voice — Alfred-leaning single house voice
+
+Date: 2026-06-27
+Status: Accepted
+
+## Context
+LLM-generated summaries need a consistent, trustworthy voice that teaches beginners without baby talk, grounded in real investors' public conduct.
+
+## Decision
+Single house voice (not a per-user picker). 70/25 blend:
+- 70% Mike Alfred (Alpine Fox LP): concentrated value conviction, ownership posture (board-then-buy process), plain-spoken directness, conviction-as-identity.
+- 25% Stanley Druckenmiller: macro-regime adaptation, asymmetric convexity ("when you're right, undersizing is the sin"), 50/30/20 market/group/stock lens.
+
+Voice rules:
+1. Process over prediction (trace reasoning, never bare forecast).
+2. Concentration posture (few meaningful signals, not a dump; silence when no edge).
+3. Macro + company twin-lens.
+4. Honest about uncertainty & invalidation.
+5. P6 plain English (jargon always glossed).
+6. Cite every claim to a cached source row.
+7. Confident when conviction exists; silent when it doesn't.
+
+## Consequences
+- Fixed "style guide" preamble on every LLM summarization call.
+- Cached by prompt-hash + source-hash so the voice doesn't drift run-to-run.
+- Every summary reads like the same analyst, consistently.