You are performing a **final trim pass** on a VibeCheck risk analysis JSON. The congruence layer is separate — this refiner only touches **risk** fields.

RISK ANALYSIS INPUT:
{risk_analysis_json}

Return ONLY valid JSON matching this exact structure. No explanation, no markdown outside the JSON block.

{{
  "system_risks": [
    {{
      "category": "<unchanged unless categorization is wrong>",
      "description": "Sharper wording; must stay grounded in the same evidence. Prefer conditional phrasing where implementation is still unknown.",
      "severity": "<unchanged unless clearly wrong>",
      "conditional": <unchanged unless description change requires it>,
      "condition_note": "<refined for clarity or null>"
    }}
  ],
  "missing_safeguards": ["at most 3 items — concrete, product-level; drop generic encryption/auth/access-control bullets unless the original analysis grounded them in the input"],
  "quick_fixes": ["at most 3 items — implementation-oriented; drop generic security consulting"],
  "risk_score": <copy the input risk_score number exactly — do not change it>,
  "risk_score_rationale": "Tighter wording; same evidential basis",
  "analysis_notes": ["at most 2 short notes; no duplication of risks"]
}}

Refinement rules:
- **Cap:** Reduce **system_risks** to **at most 4** items (keep the strongest; merge duplicates).
- Remove duplicate or near-duplicate risks in the same category.
- Strip generic phrasing that could apply to any AI product; if it cannot be fixed, remove the item.
- **Do not** add new risks, new severities, or new categories not supported by the input JSON.
- **Do not** change **risk_score** (numeric value must be identical).
- **Do not** introduce findings about PRD/code congruence — risk layer only.
- **Preserve "conditional" in text:** If a risk has `"conditional": true`, the refined **description** must still contain the literal word **conditional** (add `Conditional risk:` or similar if the input dropped it). If `"conditional": false`, do not add the word **conditional** to that description unless present in the input description.
- **Preserve grounded nouns** from the input descriptions and rationale (e.g. reply, chart, clinical note, medical note, log, logging, logged, snippet, notes) when tightening wording; do not strip concrete terms that the input already used correctly.
