Data

Three pieces:

stage1_runs/    44 zips: per-(model × substrate) message-level Stage-1 traces
stage2_runs/    220 zips: 5 conditions × 44 cells of Stage-2 traces
analysis/       rolled-up per-task records and per-cell aggregates derived
                from stage2_runs/, used directly by the figures and tables

analysis/primary/records.jsonl.gz (one row per task, n=23,375) is the canonical released artifact — it is what every figure and most tables back-reference. The raw zips are released for full reproducibility but are not required to read off the headline numbers.

Stage-1 layout

<model>__<bench>.zip
<model>__tau-bench-retail.zip
<model>__tau-bench-airline.zip

11 models × (gaia + bfcl + retail + airline) = 44 zips. Each zip:

bench_<random_id>/traces/calls.jsonl

Stage-2 layout

blind/             system prompt with no peer information
aware-c1/          system prompt includes the C1 (rubric) cards
aware-c2/          system prompt includes the C2 (static) cards
aware-c3/          system prompt includes the C3 (LLM-judge) cards
aware-tool-only/   tools wired up but no peer description in the prompt
                   (ablation: tool-availability without prompt mention)

Same naming under each condition. Each zip:

bench_<random_id>/
  delegations.jsonl  one line per call_model invocation
  traces/calls.jsonl one line per LLM API call (full request + response)

Per-task record schema (primary/records.jsonl.gz)

{
  "model":       "claude-opus-4.7",
  "bench":       "gaia" | "tau-bench" | "bfcl",
  "cond":        "blind" | "aware-c1" | "aware-c2" | "aware-c3" | "aware-tool-only",
  "task_id":     "<benchmark-native id>",
  "quality":     0.0..1.0,        // suite-native scoring
  "cost":        <USD float>,
  "latency_ms":  <int>,
  "n_calls":     <int>,
  "out_toks":    <int>,
  "in_toks":     <int>,
  "n_delegations":   <int>,
  "n_profile_reads": <int>,
  "deleg_peers": ["gpt-5.5", "deepseek-v4-flash", ...],
  "read_peers":  ["claude-sonnet-4.6", ...],
  "skills":      { "tool_schema_adherence": 0.7, "multi_step_reasoning": 0.3 }
}

The skills dict comes from the rule-based step tagger; the highest-weight key is the task's "dominant skill" used for the routing-precision metric.

Per-CSV index (analysis/primary/)

File	Used for
records.jsonl.gz	one row per task (n=23,375)
rollup.csv	mean q / cost / latency per (model, bench, condition)
lift.csv	Δ-quality vs blind per (model, bench, condition)
decomposition.csv	tool-availability vs system-prompt decomposition (Fig 4)
decomposition_agg.csv	per-bench aggregates of the same
capability_orch.csv	(blind_q, best_lift) per (model, bench) — Fig 7
skill_lift.csv	per-skill lift aggregated across agents — Fig 9
vendor_delegation_matrix.csv	7×7 vendor delegation flow — Fig 5(a)
vendor_self_pref.csv	same-vendor ratio per orchestrator — Fig 5(b)
delegation_fidelity_by_cell.csv	fidelity@k of delegations per cell
hypervolume_with_ci.csv	Pareto-hypervolume per condition with paired-bootstrap CIs
profile_reads.csv	per-cell read_profile tool counts
delegation_flow/	raw cross-vendor delegation flows per (bench × condition); inputs to the matrix above

analysis/alternate/ is a re-run with a slightly cleaner condition mix; it is the source for Fig 3 and the regression / ceiling tables. See its own README.

Schema notes

• deleg_peers and read_peers are repeat-allowed lists (one entry per invocation), not sets — useful for counting how many times the orchestrator delegates to the same peer within one task.
• cost is in USD pinned at the freeze date (2026-04-29); see the paper appendix on cost accounting.
• latency_ms is wall-clock per LLM call summed across the trajectory (excluding tool-execution time on substrates whose tools run remotely e.g. τ-bench's database calls).