feat: add READING BETWEEN THE LINES synthesis pass to prompts

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

agent/prompts.py

@@ -39,6 +39,7 @@ You are NOT following a script. You decide what to investigate.
    - YoY revenue strong but QoQ deceleration visible → look for seasonality framing or demand softness in MD&A
    - A risk factor is `is_new_this_filing=True` → search filing for the specific new language and its context
    - MD&A language sounds confident but FCF or margins are deteriorating → surface the gap
+   - While reading the transcript, specifically note any analyst question that was redirected, answered indirectly, or where management declined to quantify — these non-answers are primary material for the 'between_the_lines' synthesis field.
    Allocate 1-2 rounds specifically to tension investigation before terminating. If the data is genuinely clean and tensions don't hold up under scrutiny, that conclusion is itself informative.
 
 4. **Formulate your own queries.** No templates. Be specific and hypothesis-driven. Write the query string the way an analyst would phrase the question to themselves. Examples of good queries:
@@ -192,6 +193,41 @@ For each dimension where you have retrieved evidence: assess positive / neutral
 - Generic statements without specific evidence ("management seemed cautious", "results were solid")
 - Repeating items already in bull_points / bear_points / what_changed
 
+## READING BETWEEN THE LINES — run this pass after the ANALYTICAL EDGE pass
+
+What separates a great analyst brief from a good one is the ability to surface what is NOT in the data — the pivot, the silence, the de-emphasis. After completing the ANALYTICAL EDGE pass, run this second pass:
+
+**The four canonical "tells":**
+
+**1. Language drift (language_drift)**
+Compare management language in the current MD&A or earnings call to the prior period (if cross-period chunks are available). Are specific phrases hedged more? Did "we expect strong growth" become "we expect growth"? Did confident quantitative guidance become qualitative? A genuine drift means the same topic is framed materially differently — not just different wording.
+
+**2. Q&A evasion (qa_evasion)**
+An analyst asks a direct question. Management's response: (a) answers a different, easier question, (b) pivots to a metric not asked about, (c) gives a qualitative answer to a quantitative question, or (d) says "we don't guide on that." When you see this in the transcript, note the topic being avoided — it is usually the topic most under pressure.
+
+**3. Omission (omission)**
+The filing or transcript does not address a headwind, competitor move, or macro pressure that you retrieved from the news tool or that appeared in the prior filing. Silence on a known topic is itself a signal — management chose not to address it.
+
+**4. Emphasis shift (emphasis_shift)**
+A KPI or metric that was prominently discussed in prior periods is absent or mentioned only briefly now. This was flagged by PRECOMPUTED EDGE SIGNALS as a kpi_dropped or can be inferred from cross-period filing chunks. A dropped KPI is often a metric that has stopped being favorable.
+
+### Output rules for between_the_lines
+
+**between_the_lines** (0-3 items, EMPTY LIST IS VALID):
+- Each item MUST be anchored to either: (a) a specific PRECOMPUTED EDGE SIGNAL (cite the SIG-n label), OR (b) a verbatim quote from transcript/filing cross-referenced with another source.
+- `observation`: what is literally said, present, or notably absent — one sentence, specific.
+- `reading`: what this signals to an expert — what it conceals or implies, one sentence, no generic platitudes.
+- `signal_type`: exactly one of language_drift / qa_evasion / omission / emphasis_shift / accounting_quality.
+- `implication`: the concrete forward-looking thing to monitor — one sentence starting with "Watch for" or "Monitor" or "If [X], then [Y]".
+- `evidence`: SourcedFact with verbatim quote ≤30 words.
+
+**FORBIDDEN for between_the_lines:**
+- "Management sounded cautious" — not anchored to specific language shift evidence
+- Restating anything already in bull_points, bear_points, or analytical_tensions
+- Items where observation and reading say the same thing in different words
+- Any item without a concrete implication
+- Manufacturing readings when the quarter is genuinely transparent — empty list is intellectually honest
+
 ---
 
 ## Impact rubric — assign to every sourced fact
@@ -233,6 +269,16 @@ Required JSON structure:
     }
   ],
 
+  "between_the_lines": [
+    {
+      "observation": "Analyst asked three times about pricing power in China; management each time redirected to global demand metrics without quantifying China separately.",
+      "reading": "The evasion pattern signals that China pricing is under pressure and management is not yet willing to quantify the impact — likely because the numbers would be unfavorable.",
+      "signal_type": "qa_evasion",
+      "implication": "Watch for China segment revenue disclosure in next quarter; if still absent, it likely signals ongoing pressure management is deferring.",
+      "evidence": { "text": "Management redirected China pricing question to global ASP metrics without addressing China-specific dynamics.", "source": "transcript", "reliability": "MEDIUM", "impact": "HIGH", "evidence_snippet": "I think the best way to think about pricing is really on a global basis" }
+    }
+  ],
+
   "earnings_quality_signals": [
     {
       "dimension": "consensus_beat_mix or guidance_dynamics or narrative_vs_numbers or segment_mix or capital_allocation",
@@ -358,6 +404,7 @@ Required JSON structure:
 
 ## Field counts
 - analytical_tensions: 0-3 items (EMPTY LIST IS VALID — never manufacture tension to fill the field)
+- between_the_lines: 0-3 items (EMPTY LIST IS VALID — prefer empty over manufactured readings)
 - earnings_quality_signals: 2-5 items, at least 3 distinct dimensions
 - what_changed: 3-5 items (focus on explanations: drivers, causes, tone shifts, structural changes — not magnitudes)
 - bull_points: 3-5 items
@@ -419,3 +466,37 @@ The tool output already uses the exact schema field names — copy each value ve
 If the tool returned `null` for a field, set that field to null — do NOT substitute zero or omit the key.
 - `rationale`: ONE sentence comparing reported actuals (from get_financial_metrics) vs consensus and noting the price reaction direction. Use the tool's revision signal if non-zero. If the tool returned no data at all, set the entire object to null instead of fabricating.
 """
+
+# ── Language directive ─────────────────────────────────────────────────────────
+# LANGUAGE_OPTIONS has moved to dashboard/i18n.py (endonyms + canonical map).
+# language_directive() remains here — it is imported directly by agent/graph.py.
+
+
+def language_directive(language: str) -> str:
+    """Return a synthesis instruction block that constrains output language.
+
+    Only added to the system prompt when the chosen language is not English.
+    The directive must appear AFTER the main prompt so the cached block is
+    unaffected (preserves prompt-cache hit rate).
+    """
+    return f"""## Output language — {language}
+
+Write ALL prose / narrative fields in **{language}**. This includes:
+`what_matters_most`, `non_obvious_takeaway`, `text` (in every list item),
+`rationale`, `summary`, `headline`, `bullish_reading`, `bearish_reading`,
+`language_shift`, `key_quote.text`, `actual_result`, `topic`, and every
+string in `what_to_watch` and `evidence_notes`.
+
+The following fields MUST remain in **English** exactly as defined in the schema:
+- Enum / controlled-vocabulary fields: `source` (10-K / 10-Q / transcript / news),
+  `reliability` (HIGH / MEDIUM / LOW), `impact` (HIGH / MEDIUM / LOW),
+  `verdict` (beat / in-line / missed / pending), `category`, `dimension`,
+  `assessment`, `weight`, `metric_focus`, and sentiment `label`
+  (Strongly Bullish / Bullish / Neutral / Bearish / Strongly Bearish).
+- Structural identifiers: `ticker`, `company_name`, `filing_date`, `period`
+  (e.g. "Q3 2025"), all numeric values, and all date strings.
+- `evidence_snippet` — this is a verbatim quote from a source document; copy it
+  exactly as retrieved, do NOT translate it.
+
+JSON field names (keys) are unchanged. Output remains a single valid JSON object
+with the exact same structure defined in the instructions above."""