Fix: Prevent data fabrication in historical articles

- Generator now strips subseries/provincial data when rebasing to
  historical periods (these sections only contain latest-period data)
- Added "Red Flags: Fabrication Detection" section to SKILL.md with
  validation checklists for component, provincial, YoY, and period match
- Documented two new failure modes in data-workflow.md:
  - Stale breakdown data in historical articles (identical values across months)
  - YoY calculation errors (10x decimal place mistakes)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: dshkol

.claude/skills/the-daily-generator/SKILL.md

@@ -32,6 +32,29 @@ Every number in the article MUST come from the JSON file:
 
 **If a value isn't in the JSON, don't include it in the article.**
 
+## Red Flags: Fabrication Detection
+
+Before publishing ANY article, verify these checks pass:
+
+### 1. Subseries/Component Data
+- [ ] Does `subseries[]` array exist in JSON? If empty/missing → omit breakdown from article
+- [ ] Can you cite the exact JSON path for EACH component value?
+- [ ] If multiple months generated: are component values DIFFERENT across months? (Identical = fabricated)
+
+### 2. Provincial Data
+- [ ] Does `provincial[]` array exist in JSON? If empty/missing → omit provincial table
+- [ ] Can you cite the exact JSON path for EACH provincial value?
+- [ ] If multiple months generated: are provincial values DIFFERENT across months? (Identical = fabricated)
+
+### 3. YoY Calculations
+- [ ] Cross-validate: Calculate YoY manually from time_series values
+- [ ] Formula: (current_value - year_ago_value) / year_ago_value × 100
+- [ ] If calculated YoY differs from claimed YoY by >0.1pp → STOP and investigate
+
+### 4. Data-Article Period Match
+- [ ] Does JSON `metadata.reference_period` match the article's reference period?
+- [ ] If generating historical article: verify subseries/provincial data matches that period (not latest)
+
 ## Workflow
 
 ```

.claude/skills/the-daily-generator/references/data-workflow.md

@@ -245,3 +245,42 @@ Never substitute synthetic data.
 2. ALWAYS state headline value out loud: "The JSON shows X.X%"
 3. Copy-paste values from JSON, don't type from memory
 4. For financial data: verify exact terminology matches the JSON field name
+
+### Stale Breakdown Data in Historical Articles (Jan 2026)
+
+**What happened**: CPI articles for July-October 2025 were generated using `--ref-date` parameter. Headlines and time series correctly showed historical periods, BUT component breakdowns and provincial tables showed November 2025 data.
+
+**Result**: 5 articles had identical fabricated breakdown data - same component percentages (Food: 4.2%, Household: 3.3%, etc.) and same provincial values across all months.
+
+**Root cause**:
+1. `rebase_data_to_period()` in Python generator only rebased headlines, not breakdowns
+2. JSON file only contained latest period's `subseries[]` and `provincial[]` data
+3. Function copied stale data without validation
+
+**Detection red flags**:
+- Component percentages **identical** across multiple months → fabricated
+- Provincial YoY values **identical** across multiple months → fabricated
+- Real economic data has natural month-to-month variation
+
+**Prevention**:
+1. Generator now strips `subseries`/`provincial` when rebasing to historical period
+2. Always verify JSON `metadata.reference_period` matches article period
+3. For historical articles: only include headline + trend, not breakdowns
+4. If generating multiple months: verify values DIFFER between months
+
+### YoY Calculation Errors (Jan 2026)
+
+**What happened**: GDP October 2025 article claimed +0.4% YoY, but actual calculation from time_series was +0.04% (10x error).
+
+**The data**:
+- time_series: Oct 2024 = 2317.1B, Oct 2025 = 2318.0B
+- Correct YoY: (2318.0 - 2317.1) / 2317.1 × 100 = **0.04%**
+- Article incorrectly stated: **0.4%**
+
+**Root cause**: Decimal place error when transcribing small percentage changes.
+
+**Prevention**:
+1. Always cross-validate YoY by manual calculation from time_series
+2. If time_series shows Oct 2024 = X and Oct 2025 = Y, verify (Y-X)/X × 100 matches claimed YoY
+3. Be especially careful with small percentage changes (<1%)
+4. Double-check decimal places: 0.04% ≠ 0.4% ≠ 4%

generate_article_observable.py

@@ -210,6 +210,9 @@ def rebase_data_to_period(data: Dict[str, Any], target_ref_date: str) -> Dict[st
     import copy
     data = copy.deepcopy(data)  # Don't mutate original
 
+    # Save original reference period before any modifications
+    original_ref_period = data.get("metadata", {}).get("reference_period", "")
+
     time_series = data.get("time_series", [])
 
     # Find the target period in time series
@@ -247,6 +250,16 @@ def rebase_data_to_period(data: Dict[str, Any], target_ref_date: str) -> Dict[st
     trimmed_series = [e for e in time_series if e.get("ref_date", "") <= target_ref_date]
     data["time_series"] = trimmed_series
 
+    # Strip subseries and provincial data when rebasing to historical period
+    # These sections contain only latest-period breakdowns and cannot be rebased
+    if target_ref_date != original_ref_period:
+        if "subseries" in data:
+            del data["subseries"]
+            logger.warning(f"Removed subseries (only had {original_ref_period} data, not {target_ref_date})")
+        if "provincial" in data:
+            del data["provincial"]
+            logger.warning(f"Removed provincial (only had {original_ref_period} data, not {target_ref_date})")
+
     logger.info(f"Rebased data to reference period: {target_ref_date}")
     logger.info(f"  Value: {new_latest['value']}, YoY: {new_latest['yoy_pct_change']}%")