Second opinion — LADLE docs

Second opinion is a one-click action on any completed assistant reply. It runs a fresh Anthropic call to the OTHER tier (never the same model that wrote the answer), passing the immediately-preceding user turn and the assistant reply, and asks the verifier to review for errors, omissions, or overconfidence. The verifier's critique is persisted on the message and rendered as a quiet block below the reply.

How it works

One HTTP call, one Anthropic message, one row-level update. No streaming.

EndpointPOST /api/app/messages/:messageId/second-opinion

Verifier modelOpposite tier: Haiku → Sonnet, Sonnet → Haiku. Defaults to Sonnet if the source model is unknown.

System prompt'You are a rigorous second opinion. The user asked a question and an assistant answered. Review the assistant answer for errors, omissions, or overconfidence. Be specific. Cite what's likely wrong or missing. If the answer holds up, say so briefly and stop.'

Max tokens800 (keeps the critique focused)

Persistencemessages.second_opinion jsonb: { model, content, input_tokens, output_tokens, elapsed_ms, created_at }

Rate limitOne second opinion per assistant message. Re-clicking overwrites the previous critique.

Cost and tokens

Second opinions count against your usage cap. Sonnet-as-verifier typically uses 800-2000 input tokens (parent turn + reply) + up to 800 output tokens; Haiku-as-verifier uses the same input with typically 300-600 output tokens. Per-call cost is $0.005-$0.03. The verifier tokens are logged separately in usage_counters via increment_usage_counter.

How the critique renders

As a quiet block below the assistant reply, indented with a 2px amber border on the left, prefixed with a label like 'SONNET · SECOND OPINION · 1,240ms'. The critique text renders in the same font as an assistant message but at 90% opacity. If the verifier disagrees, the critique is worth reading; if it agrees, it typically returns 2-3 sentences and stops (per the prompt).

When second opinion is worth using

High-stakes factual answers (legal / medical / financial reasoning). Long code where a bug could hide. Confident-sounding claims that don't cite sources. Not useful for: creative writing, casual conversation, or asks where the assistant already said 'I'm not sure'. If the answer opened with uncertainty, the second opinion will typically confirm the uncertainty, not resolve it.

When it doesn't help

Second opinion is not a truth-check — it's a peer-review by a model with the same training data. If BOTH models share the same training-data blind spot, the second opinion will confirm the wrong answer. For questions where the ground truth matters and might not be in the training data, enable SEARCH instead — Anthropic's server-side web_search tool will fetch live sources, and the verifier can be enabled on top of the search-grounded reply for extra confidence.

Docs: Citations and web search

Second opinion.