JSON · schema

json_schema

A structured-output frame using the literal words as JSON values.

Design intent

Tests structured output without an abstract A/B slot: the permitted JSON values are the words themselves, so a slot-A shortcut cannot answer it.

The recorded prompt

The template below is reconstructed from the recorded request by replacing the two real probe words with placeholders. The rendered example is the exact vodka / whiskey, AB-order request stored on the trace.

Template

SYSTEM
You are a helpful assistant.

USER
Choose between '{word_a}' and '{word_b}'.

Respond with valid JSON matching this schema, and nothing else:

  {
    "chosen": '{word_a}' | '{word_b}'
  }

Output only the JSON object. No code fences, no explanation.

Rendered example

SYSTEM
You are a helpful assistant.

USER
Choose between 'vodka' and 'whiskey'.

Respond with valid JSON matching this schema, and nothing else:

  {
    "chosen": 'vodka' | 'whiskey'
  }

Output only the JSON object. No code fences, no explanation.

source trace 289f58ac-52e8-468e-b81d-3d55f4927cb8 · model anthropic/claude-haiku-4.5 · recorded 2026-05-09T13:23:09.471802+00:00 · experiment version not recorded · envelope version not recorded · temperature not recorded

Agreement-rate statistics

“Consistent” means the same word was selected in both orderings. “No stable choice” combines every non-real state retained by this frozen artifact, including position sensitivity, refusal, and missing/incomplete pairing; the snapshot does not preserve those sub-reasons separately. The equal-model mean gives every model one vote; the pair-weighted rate gives every derived record one vote. Neither uses a cross-envelope majority, so the envelope is never compared partly with itself.

Coverage note: 2 model nodes in the all-model snapshot are absent from this envelope leaf and are not counted as zero: anthropic/claude-fable-5@00-low-2026-06-10, anthropic/claude-fable-5@10-medium-2026-06-10.

source: canonical /v2/ envelope leaf, snapshot 2026-06-10 10:55 UTC. A pair record is already collapsed across AB+BA; it is not a raw response count. Open the interactive envelope cut.

Recorded responses

Five providers on the same rendered AB probe. Selection is outcome-independent: within each provider, take the first model alphabetically among models that both belong to the canonical envelope cut and have a matching non-null fixed-probe AB trace, then take that model's earliest match. Response text is never used to choose an example.

```json
{
  "chosen": "whiskey"
}
```

{
    "chosen": "whiskey"
}

```json
{
  "chosen": "whiskey"
}
```

{"chosen": "whiskey"}

{"chosen": "whiskey"}