Sonnet 5 high-effort thinking — datacenter framing

This is the strongest existing realistic-datacenter campaign for studying large-model negotiation under both PRIVATE and FULL preference visibility. It uses anthropic:claude-sonnet-5 with adaptive thinking, high effort, the moves_chat protocol, and ten rollout seeds.

Run summary

• 60/60 completed episodes: six game instances across seeds 0–9.
• 30 PRIVATE episodes and 30 matched FULL-information episodes.
• Three underlying payoff games per visibility condition.
• Oracle annotations: threshold, acceptance, and best response.
• Rational counterfactual annotations are in annotations_v1/.
• This run is paired with ../apibehav_sonnet5_thinkon_cot_abstract/, which uses the same underlying conditions with abstract framing.

The two framing runs together contain 120 episodes and 60 paired abstract/datacenter traces. Because the framings reuse the same instance/seed conditions, they represent 30 independent conditions rather than 60 independent conditions within each visibility regime.

Preference visibility

In PRIVATE episodes, each participant sees only its own score sheet and acceptance threshold. Other participants' preferences must be inferred from public messages, proposals, and votes. In FULL episodes, all score sheets are revealed.

Use each episode's outcome.info field to distinguish private from full. The run-level manifest.json field says full, but this is a mixed campaign and the episode-level field is authoritative.

Thinking provenance

The model was invoked with --model-thinking on, --api-effort high, and a 16,384-token per-turn API floor. Anthropic does not expose the model's native hidden chain of thought verbatim. Where available, the episode stores the provider-permitted reasoning summary or explicit scratchpad with provenance withheld_or_summarized; an absent reasoning field does not imply thinking was disabled.

Reading the oracle results

The rational best-response oracle is an omniscient counterfactual. In PRIVATE episodes it can use preferences hidden from the acting model, so its value gap should not be interpreted as implementable-policy regret. Inspect the prompt and Preference Visibility indicator to audit what the acting model actually knew.

Visualize

From /juice2/u/siddharth/ii_mats:

uv run --no-sync python -m interlens.arena.viz \
  --run /nlp/scr/siddharth/ii_mats/rational_agents/apibehav_sonnet5_thinkon_cot_datacenter \
  --serve

The generated index includes all 60 episodes; use the Preference Visibility indicator to identify the information regime on each page.