ARSENIC drift report

Executive summary

Probe outcomes and drift valence across the run.

Safe to upgrade: true · Manual review items: 2

Critical regressions: 2 · Policy drift: 0 · Fidelity drift: 0 · Structural drift: 0 · Content compression: 0

Dimension overview

Worst risk per dimension, drift direction mix among affected probes (↑ / ↓ / →), and affected count.

Red probes

No blocking regressions detected.

All probe results

Expand for prompt, dimension snapshot (with drift arrows), claim intelligence, and responses. Consistency σ shows run variance when multi-run sampling was enabled.

Return a JSON array containing exactly 3 objects. Each object must have: id (integer), label (string). No explanation, just the JSON array.

```json
[
    {
        "id": 1,
        "label": "First Item"
    },
    {
        "id": 2,
        "label": "Second Item"
    },
    {
        "id": 3,
        "label": "Third Item"
    }
]
```

```json
[
  {"id": 1, "label": "Apple"},
  {"id": 2, "label": "Banana"},
  {"id": 3, "label": "Cherry"}
]
```

Classify the following text and return ONLY a JSON object with fields: category (string), confidence (float 0-1, must be a number not a word), reasoning (string). No markdown fences. Raw JSON only.

Text: 'The quarterly revenue exceeded expectations by 12 percent.'

{
  "category": "Business",
  "confidence": 0.95,
  "reasoning": "The text discusses financial performance, specifically revenue and expectations, which are key elements in the business context."
}

{"category":"Finance","confidence":0.95,"reasoning":"The text discusses quarterly revenue and its performance relative to expectations, which is directly related to financial reporting and business performance metrics."}

Return this exact JSON with these exact values. No markdown code fences. No explanation. Raw JSON only:
{
  "request": {
    "meta": {
      "version": 1,
      "source": "api.example.com"
    },
    "payload": {
      "items": ["item1", "item2", "item3"]
    }
  }
}

{
  "request": {
    "meta": {
      "version": 1,
      "source": "api.example.com"
    },
    "payload": {
      "items": ["item1", "item2", "item3"]
    }
  }
}

{
  "request": {
    "meta": {
      "version": 1,
      "source": "api.example.com"
    },
    "payload": {
      "items": ["item1", "item2", "item3"]
    }
  }
}

Return a JSON object with these fields: status (must be exactly the string "pending"), priority (integer, set it to 3). No explanation, just the JSON.

```json
{
  "status": "pending",
  "priority": 3
}
```

```json
{
  "status": "pending",
  "priority": 3
}
```

Return a JSON object with exactly these fields: name (string), age (integer), active (boolean). No other fields. No explanation, just the JSON.

```json
{
  "name": "John Doe",
  "age": 30,
  "active": true
}
```

```json
{
  "name": "John Doe",
  "age": 30,
  "active": true
}
```

Return a JSON object with this exact structure and these exact values. No explanation, just the JSON:
{
  "user": {
    "id": 1,
    "email": "user@example.com"
  },
  "status": "active"
}

```json
{
  "user": {
    "id": 1,
    "email": "user@example.com"
  },
  "status": "active"
}
```

```json
{
  "user": {
    "id": 1,
    "email": "user@example.com"
  },
  "status": "active"
}
```

Return a JSON object: {"ok": true, "code": 200}. Output the raw JSON only. Do not wrap it in markdown code fences. Do not add any text before or after.

{"ok": true, "code": 200}

{"ok": true, "code": 200}

Return a JSON object with these fields: title (string), description (string or null), count (integer). Set title to "test", description to null, count to 0. No explanation, just the JSON.

```json
{
  "title": "test",
  "description": null,
  "count": 0
}
```

```json
{
  "title": "test",
  "description": null,
  "count": 0
}
```

Return a JSON object with: count (must be the integer 5), ratio (must be the float 0.75), label (must be the string "test"). Values must be exactly as specified. No explanation, just the JSON.

```json
{
  "count": 5,
  "ratio": 0.75,
  "label": "test"
}
```

```json
{
  "count": 5,
  "ratio": 0.75,
  "label": "test"
}
```

Return a JSON object with ONLY these two fields: result (string), confidence (float between 0 and 1). Set result to the string "ok" and confidence to 0.9. Do not add any other fields. No markdown, no explanation. Raw JSON only.

{"result":"ok","confidence":0.9}

{"result":"ok","confidence":0.9}

Upgrade path

Structured rollout view. Validated prompt patches appear when mutations validate.

Blocking 0 · Verify 2 · Neutral 0 · Validated 0

No blocking regressions detected.

Improvements to verify

• schema_array_of_objects — High / Amber / Regression / CriticalRegression — warrants attention before switching
• schema_classification_output — High / Amber / Regression / CriticalRegression — warrants attention before switching

Validated prompt patches

No automated patches generated — run with --mutate to attempt prompt repairs.

Latency impact

Observational only — latency does not affect overall risk, probe direction, or upgrade path routing.

Delta: -313 ms (-23.1%) · Direction: Improvement

v2 responded 23% faster on average across 10 probes

Upgrade recommendation

Overall risk is acceptable for a controlled rollout. Review amber items and any "verify" red valence before broad deployment.