Hugging Face's logo

Datasets:
AgentsSci
/
EMNLP_Cost-Aware-Protocol-Routing

Tasks:
Tabular Classification
Question Answering
Modalities:
Tabular
Text
Formats:
csv
Sub-tasks:
multi-class-classification
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
multi-agent-systems
llm-routing
cost-aware-inference
calibration
agent-collaboration
reasoning
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
EMNLP_Cost-Aware-Protocol-Routing / registry /schema.json
ChihHsuan-Yang's picture
ChihHsuan-Yang
Document the confidence scale and null handling; correct the fallback-row statement; note AUPRC tie sensitivity
5c5c8b4 verified
Raw History Blame Contribute Delete
41.3 kB
{
 "schema_version": "emnlp-protocol-routing-dataset-v1.0.0",
 "generated_utc": "2026-09-24T02:18:59.281412+00:00",
 "conventions": {
 "join_key_cross_solver": "problem_id",
 "join_key_within_setting": "source_problem_id",
 "setting_id_format": "<benchmark>__<condition>__<run_id>",
 "example_id_formula": "<setting_id>::sha1('<setting_id>:<source_problem_id>')[:12]",
 "oracle_order": [
 "baseline_llm",
 "single_agent",
 "per",
 "broadcast"
 ],
 "oracle_fallthrough_label": "none",
 "oracle_none_semantics": "retrospective router ACTION (abstain); all four observed executions failed. Not a fifth protocol.",
 "bootstrap": "2000 problem-level percentile resamples",
 "rate_units": "proportions unless the column name ends in _pct or _points"
 },
 "leakage_policy": {
 "rule": "No file with role 'feature' may contain gold answers, per-protocol correctness, or oracle labels.",
 "feature_files": [
 "data/confidence/postanswer_confidence_predictions.csv",
 "data/confidence/primary_omnimath_confidence_predictions.csv",
 "data/costs/omnimath_per_protocol_costs.csv",
 "data/probe_inputs.jsonl",
 "data/problems.csv",
 "data/splits/primary_omnimath_splits.csv",
 "data/splits/six_setting_splits.csv"
 ],
 "label_files": [
 "data/labels_for_scoring.csv",
 "data/matched_labels.csv",
 "data/router/six_setting_test_predictions.csv"
 ]
 },
 "tables": {
 "data/aggregate/failure_and_protocol_value_targets.csv": {
 "format": "csv",
 "rows": 24,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (failure_and_protocol_value_targets.csv).",
 "bytes": 3154,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "setting",
 "type": "string"
 },
 {
 "name": "target",
 "type": "string"
 },
 {
 "name": "n_parseable",
 "type": "integer"
 },
 {
 "name": "positive_n",
 "type": "integer"
 },
 {
 "name": "prevalence",
 "type": "number"
 },
 {
 "name": "auroc",
 "type": "number"
 },
 {
 "name": "auprc",
 "type": "number",
 "precision": "approximately 2 decimals, not 4",
 "caveat": "Scores are heavily tied (integer confidences) and the original average-precision implementation did not handle ties, so a standard estimator will not match the published digits. AUROC is tie-invariant and reproduces exactly. See docs/provenance/auprc_tie_handling.md in the code repository."
 }
 ]
 },
 "data/aggregate/heldout_router_evaluation.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (heldout_router_evaluation.csv).",
 "bytes": 1164,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "setting",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "baseline_solve",
 "type": "number"
 },
 {
 "name": "router_solve",
 "type": "number"
 },
 {
 "name": "router_ci",
 "type": "string"
 },
 {
 "name": "oracle_solve",
 "type": "number"
 },
 {
 "name": "oracle_gap",
 "type": "number"
 },
 {
 "name": "oracle_label_macro_f1",
 "type": "number"
 },
 {
 "name": "route_baseline",
 "type": "number"
 },
 {
 "name": "route_single",
 "type": "number"
 },
 {
 "name": "route_per",
 "type": "number"
 },
 {
 "name": "route_broadcast",
 "type": "number"
 },
 {
 "name": "route_none",
 "type": "number"
 }
 ]
 },
 "data/aggregate/heldout_router_paired_differences.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (heldout_router_paired_differences.csv).",
 "bytes": 843,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "setting",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "router_minus_tier_majority_points",
 "type": "number"
 },
 {
 "name": "router_minus_tier_majority_ci_points",
 "type": "string"
 },
 {
 "name": "router_minus_baseline_points",
 "type": "number"
 },
 {
 "name": "router_minus_baseline_ci_points",
 "type": "string"
 }
 ]
 },
 "data/aggregate/main_routing_heldout.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (main_routing_heldout.csv).",
 "bytes": 623,
 "columns": [
 {
 "name": "class",
 "type": "string"
 },
 {
 "name": "policy",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "solve_pct",
 "type": "number"
 },
 {
 "name": "solve_ci",
 "type": "string"
 },
 {
 "name": "avg_tokens_k",
 "type": "number"
 },
 {
 "name": "avg_tokens_ci",
 "type": "string"
 },
 {
 "name": "excess_tokens_k",
 "type": "number"
 },
 {
 "name": "excess_tokens_ci",
 "type": "string"
 }
 ]
 },
 "data/aggregate/matched_protocol_coverage.csv": {
 "format": "csv",
 "rows": 10,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (matched_protocol_coverage.csv).",
 "bytes": 1096,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "setting",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "baseline_pct",
 "type": "number"
 },
 {
 "name": "single_pct",
 "type": "number"
 },
 {
 "name": "per_pct",
 "type": "number"
 },
 {
 "name": "broadcast_pct",
 "type": "number"
 },
 {
 "name": "oracle_pct",
 "type": "number"
 }
 ]
 },
 "data/aggregate/oracle_label_distribution.csv": {
 "format": "csv",
 "rows": 10,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (oracle_label_distribution.csv).",
 "bytes": 1076,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "setting",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "baseline_pct",
 "type": "number"
 },
 {
 "name": "single_pct",
 "type": "number"
 },
 {
 "name": "per_pct",
 "type": "number"
 },
 {
 "name": "broadcast_pct",
 "type": "number"
 },
 {
 "name": "none_pct",
 "type": "number"
 }
 ]
 },
 "data/aggregate/per_broadcast_bootstrap_conditional.csv": {
 "format": "csv",
 "rows": 24,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (per_broadcast_bootstrap_conditional.csv).",
 "bytes": 6449,
 "columns": [
 {
 "name": "display_name",
 "type": "string"
 },
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "subset",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "PER_solve_rate",
 "type": "number"
 },
 {
 "name": "Broadcast_solve_rate",
 "type": "number"
 },
 {
 "name": "Broadcast_minus_PER",
 "type": "number"
 },
 {
 "name": "Broadcast_minus_PER_CI",
 "type": "string"
 },
 {
 "name": "PER_only_count",
 "type": "integer"
 },
 {
 "name": "PER_only_rate",
 "type": "number"
 },
 {
 "name": "Broadcast_only_count",
 "type": "integer"
 },
 {
 "name": "Broadcast_only_rate",
 "type": "number"
 },
 {
 "name": "both_count",
 "type": "integer"
 },
 {
 "name": "both_rate",
 "type": "number"
 },
 {
 "name": "neither_count",
 "type": "integer"
 },
 {
 "name": "neither_rate",
 "type": "number"
 }
 ]
 },
 "data/aggregate/per_broadcast_interaction.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (per_broadcast_interaction.csv).",
 "bytes": 1073,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "setting",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "per_solve_pct",
 "type": "number"
 },
 {
 "name": "broadcast_solve_pct",
 "type": "number"
 },
 {
 "name": "broadcast_minus_per_points",
 "type": "number"
 },
 {
 "name": "broadcast_minus_per_ci_points",
 "type": "string"
 },
 {
 "name": "per_only_count",
 "type": "integer"
 },
 {
 "name": "per_only_pct",
 "type": "number"
 },
 {
 "name": "broadcast_only_count",
 "type": "integer"
 },
 {
 "name": "broadcast_only_pct",
 "type": "number"
 },
 {
 "name": "both_count",
 "type": "integer"
 },
 {
 "name": "both_pct",
 "type": "number"
 },
 {
 "name": "neither_count",
 "type": "integer"
 },
 {
 "name": "neither_pct",
 "type": "number"
 }
 ]
 },
 "data/aggregate/postanswer_confidence.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Aggregate result table reproduced from the camera-ready ancillary directory (postanswer_confidence.csv).",
 "bytes": 1216,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "setting",
 "type": "string"
 },
 {
 "name": "n_total",
 "type": "integer"
 },
 {
 "name": "n_parseable",
 "type": "integer"
 },
 {
 "name": "parse_rate",
 "type": "number"
 },
 {
 "name": "failure_auroc",
 "type": "number"
 },
 {
 "name": "failure_auroc_ci",
 "type": "string"
 },
 {
 "name": "ece",
 "type": "number"
 },
 {
 "name": "ece_ci",
 "type": "string"
 },
 {
 "name": "brier",
 "type": "number"
 },
 {
 "name": "brier_ci",
 "type": "string"
 },
 {
 "name": "mean_conf_correct",
 "type": "number"
 },
 {
 "name": "mean_conf_wrong",
 "type": "number"
 }
 ]
 },
 "data/confidence/failure_and_protocol_value_targets.csv": {
 "format": "csv",
 "rows": 24,
 "role": "aggregate",
 "description": "The failure score scored against four increasingly protocol-specific targets. The paper's central negative result.",
 "bytes": 4253,
 "columns": [
 {
 "name": "display_name",
 "type": "string"
 },
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "target",
 "type": "string"
 },
 {
 "name": "n_parseable",
 "type": "integer"
 },
 {
 "name": "positive_n",
 "type": "integer"
 },
 {
 "name": "prevalence",
 "type": "number"
 },
 {
 "name": "AUROC_using_failure_score",
 "type": "number"
 },
 {
 "name": "AUPRC_using_failure_score",
 "type": "number",
 "precision": "approximately 2 decimals, not 4",
 "caveat": "Scores are heavily tied (integer confidences) and the original average-precision implementation did not handle ties, so a standard estimator will not match the published digits. AUROC is tie-invariant and reproduces exactly. See docs/provenance/auprc_tie_handling.md in the code repository."
 }
 ]
 },
 "data/confidence/postanswer_confidence_predictions.csv": {
 "format": "csv",
 "rows": 12928,
 "role": "feature",
 "description": "POST-ANSWER, pre-collaboration probe (probe_type=post_answer_pre_collaboration), all 6 analysed settings. Sees the problem, allowed metadata and the model's own Baseline final answer; never the gold answer, correctness, oracle label or any protocol outcome. Reproduces the paper's headline failure-risk numbers. JOIN ON example_id, never on problem_uid_run.",
 "bytes": 2921875,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "example_id",
 "type": "string"
 },
 {
 "name": "problem_uid_run",
 "type": "string"
 },
 {
 "name": "probe_type",
 "type": "string"
 },
 {
 "name": "prober_model",
 "type": "string"
 },
 {
 "name": "run_id",
 "type": "string"
 },
 {
 "name": "parse_ok",
 "type": "boolean"
 },
 {
 "name": "confidence",
 "type": "number",
 "unit": "percent",
 "scale": "0-100 integer, NOT 0-1",
 "description": "P(the Baseline final answer is correct), integer percent 0-100."
 },
 {
 "name": "confidence_norm",
 "type": "number",
 "unit": "proportion",
 "scale": "0-1",
 "description": "The same value as `confidence`, rescaled to 0-1."
 },
 {
 "name": "repair_attempts",
 "type": "integer"
 },
 {
 "name": "parse_error",
 "type": "string"
 }
 ]
 },
 "data/confidence/primary_omnimath_confidence_predictions.csv": {
 "format": "csv",
 "rows": 839,
 "role": "feature",
 "description": "PRE-ANSWER q1 probe (probe_type=pre_answer_q1) on the primary OmniMath split. Sees the problem only, before any solving. Feature side: no correctness, no oracle label.",
 "bytes": 144860,
 "columns": [
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "legacy_tier_id",
 "type": "string"
 },
 {
 "name": "split",
 "type": "string",
 "enum": [
 "train",
 "dev",
 "test"
 ]
 },
 {
 "name": "probe_type",
 "type": "string"
 },
 {
 "name": "probe_run",
 "type": "string"
 },
 {
 "name": "prober_model",
 "type": "string"
 },
 {
 "name": "probe_question_id",
 "type": "string"
 },
 {
 "name": "prompt_version",
 "type": "string"
 },
 {
 "name": "confidence_probability",
 "type": "number",
 "unit": "percent",
 "scale": "0-100 integer, NOT 0-1",
 "observed_min": 7.0,
 "observed_max": 99.0,
 "observed_distinct": 41,
 "nullable": true,
 "null_meaning": "the probe produced no parseable confidence (used_fallback=true, status=fallback_after_failure); 222 of 839 rows are null",
 "null_handling": "the documented confidence gate treats a null as ESCALATE. Threshold 70 means 70 percent. Reading the column as a 0-1 probability, or treating nulls as 'stay', silently yields a wrong solve rate (60.76% and 73.76% respectively, against the published 78.0%).",
 "description": "P(the model solves this problem in one pass), integer percent 0-100. Empty when unparseable."
 },
 {
 "name": "status",
 "type": "string"
 },
 {
 "name": "used_fallback",
 "type": "boolean"
 },
 {
 "name": "attempts",
 "type": "integer"
 },
 {
 "name": "probe_prompt_tokens",
 "type": "number"
 },
 {
 "name": "probe_completion_tokens",
 "type": "number"
 },
 {
 "name": "probe_total_tokens",
 "type": "number"
 }
 ]
 },
 "data/confidence/six_setting_confidence_metrics.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Post-answer failure-risk metrics: parse rate, AUROC, ECE, Brier, with bootstrap CIs.",
 "bytes": 1317,
 "columns": [
 {
 "name": "display_name",
 "type": "string"
 },
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "model",
 "type": "string",
 "enum": [
 "gpt_oss_120b",
 "gemma_4_31b"
 ]
 },
 {
 "name": "domain",
 "type": "string",
 "enum": [
 "math",
 "science",
 "biology"
 ]
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "parse_rate",
 "type": "number"
 },
 {
 "name": "AUROC_failure",
 "type": "number"
 },
 {
 "name": "AUROC_CI",
 "type": "string"
 },
 {
 "name": "ECE",
 "type": "number"
 },
 {
 "name": "ECE_CI",
 "type": "string"
 },
 {
 "name": "Brier",
 "type": "number"
 },
 {
 "name": "Brier_CI",
 "type": "string"
 },
 {
 "name": "mean_conf_correct",
 "type": "number"
 },
 {
 "name": "mean_conf_wrong",
 "type": "number"
 }
 ]
 },
 "data/costs/omnimath_per_protocol_costs.csv": {
 "format": "csv",
 "rows": 4155,
 "role": "feature",
 "description": "Per-problem protocol-level token totals and model-call counts, summed over every model call (the paper's cost accounting). Covers ONE of the ten settings (omnimath2 gpt-oss-120b), 4155 of its 4181 problems; 26 omitted as unidentifiable through duplicate-text groups. Feature side: no correctness, no oracle label.",
 "bytes": 395362,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "baseline_total_tokens",
 "type": "integer"
 },
 {
 "name": "single_total_tokens",
 "type": "integer"
 },
 {
 "name": "per_total_tokens",
 "type": "integer"
 },
 {
 "name": "broadcast_total_tokens",
 "type": "integer"
 },
 {
 "name": "baseline_model_calls",
 "type": "integer"
 },
 {
 "name": "single_model_calls",
 "type": "integer"
 },
 {
 "name": "per_model_calls",
 "type": "integer"
 },
 {
 "name": "broadcast_model_calls",
 "type": "integer"
 }
 ]
 },
 "data/labels_for_scoring.csv": {
 "format": "csv",
 "rows": 12928,
 "role": "label",
 "description": "Outcome labels for the 6 settings with router and confidence analyses. Join to probe_inputs.jsonl on example_id.",
 "bytes": 2919126,
 "columns": [
 {
 "name": "example_id",
 "type": "string"
 },
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "benchmark_id",
 "type": "string",
 "enum": [
 "omnimath2",
 "jeebench",
 "scibench",
 "labbench"
 ]
 },
 {
 "name": "slice_id",
 "type": "string"
 },
 {
 "name": "run_id",
 "type": "string"
 },
 {
 "name": "model",
 "type": "string",
 "enum": [
 "gpt_oss_120b",
 "gemma_4_31b"
 ]
 },
 {
 "name": "domain",
 "type": "string",
 "enum": [
 "math",
 "science",
 "biology"
 ]
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "source_problem_id",
 "type": "string"
 },
 {
 "name": "baseline_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "single_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "per_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "broadcast_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "oracle_label",
 "type": "string",
 "enum": [
 "baseline_llm",
 "single_agent",
 "per",
 "broadcast",
 "none"
 ]
 },
 {
 "name": "any_protocol_solved",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 }
 ]
 },
 "data/matched_labels.csv": {
 "format": "csv",
 "rows": 15088,
 "role": "label",
 "description": "Flagship table. One row per (setting, problem): per-protocol correctness and the fixed-order oracle label for all 10 released settings.",
 "bytes": 2814373,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "benchmark_id",
 "type": "string",
 "enum": [
 "omnimath2",
 "jeebench",
 "scibench",
 "labbench"
 ]
 },
 {
 "name": "slice_id",
 "type": "string"
 },
 {
 "name": "run_id",
 "type": "string"
 },
 {
 "name": "model",
 "type": "string",
 "enum": [
 "gpt_oss_120b",
 "gemma_4_31b"
 ]
 },
 {
 "name": "model_endpoint",
 "type": "string",
 "enum": [
 "openai/gpt-oss-120b",
 "google/gemma-4-31B-it"
 ]
 },
 {
 "name": "domain",
 "type": "string",
 "enum": [
 "math",
 "science",
 "biology"
 ]
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "source_problem_id",
 "type": "string"
 },
 {
 "name": "baseline_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "single_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "per_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "broadcast_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "oracle_label",
 "type": "string",
 "enum": [
 "baseline_llm",
 "single_agent",
 "per",
 "broadcast",
 "none"
 ]
 },
 {
 "name": "any_protocol_solved",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 }
 ]
 },
 "data/probe_inputs.jsonl": {
 "format": "jsonl",
 "rows": 12928,
 "role": "feature",
 "description": "Feature side of the confidence probe. No labels. problem_text and baseline_final_answer are null in every row (withheld / not available).",
 "bytes": 4640313,
 "columns": [
 {
 "name": "example_id",
 "type": "string"
 },
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "benchmark_id",
 "type": "string",
 "enum": [
 "omnimath2",
 "jeebench",
 "scibench",
 "labbench"
 ]
 },
 {
 "name": "slice_id",
 "type": "string"
 },
 {
 "name": "model_endpoint",
 "type": "string",
 "enum": [
 "openai/gpt-oss-120b",
 "google/gemma-4-31B-it"
 ]
 },
 {
 "name": "domain",
 "type": "string",
 "enum": [
 "math",
 "science",
 "biology"
 ]
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "source_problem_id",
 "type": "string"
 },
 {
 "name": "split",
 "type": "string",
 "enum": [
 "train",
 "dev",
 "test"
 ]
 }
 ]
 },
 "data/problems.csv": {
 "format": "csv",
 "rows": 6803,
 "role": "feature",
 "description": "Router-visible problem metadata only. No text, no answers, no correctness, no oracle label.",
 "bytes": 577521,
 "columns": [
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "benchmark_id",
 "type": "string",
 "enum": [
 "omnimath2",
 "jeebench",
 "scibench",
 "labbench"
 ]
 },
 {
 "name": "source_problem_id",
 "type": "string"
 },
 {
 "name": "legacy_tier_id",
 "type": "number"
 },
 {
 "name": "subset",
 "type": "string"
 },
 {
 "name": "source",
 "type": "string"
 },
 {
 "name": "difficulty",
 "type": "number"
 },
 {
 "name": "difficulty_tier",
 "type": "number"
 },
 {
 "name": "domain",
 "type": "string",
 "enum": [
 "math",
 "science",
 "biology"
 ]
 },
 {
 "name": "primary_split_423",
 "type": "number"
 }
 ]
 },
 "data/router/six_setting_metrics.csv": {
 "format": "csv",
 "rows": 18,
 "role": "aggregate",
 "description": "Per (setting, router) held-out metrics with bootstrap CIs.",
 "bytes": 3430,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "model",
 "type": "string",
 "enum": [
 "gpt_oss_120b",
 "gemma_4_31b"
 ]
 },
 {
 "name": "domain",
 "type": "string",
 "enum": [
 "math",
 "science",
 "biology"
 ]
 },
 {
 "name": "router",
 "type": "string",
 "enum": [
 "metadata_bucket_majority",
 "metadata_logreg",
 "text_metadata_logreg"
 ]
 },
 {
 "name": "split",
 "type": "string",
 "enum": [
 "train",
 "dev",
 "test"
 ]
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "solve_rate",
 "type": "number"
 },
 {
 "name": "solve_rate_CI",
 "type": "string"
 },
 {
 "name": "baseline_rate",
 "type": "number"
 },
 {
 "name": "oracle_rate",
 "type": "number"
 },
 {
 "name": "gain_vs_baseline",
 "type": "number"
 },
 {
 "name": "oracle_gap",
 "type": "number"
 },
 {
 "name": "oracle_label_accuracy",
 "type": "number"
 },
 {
 "name": "under_escalation_rate",
 "type": "number"
 },
 {
 "name": "over_escalation_rate",
 "type": "number"
 },
 {
 "name": "mean_escalation_level",
 "type": "number"
 },
 {
 "name": "broadcast_rate",
 "type": "number"
 }
 ]
 },
 "data/router/six_setting_paired_delta_bootstrap.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Paired solve-rate differences vs tier-majority and Baseline.",
 "bytes": 991,
 "columns": [
 {
 "name": "display_name",
 "type": "string"
 },
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "router_minus_tier_majority",
 "type": "number"
 },
 {
 "name": "router_minus_tier_majority_CI",
 "type": "string"
 },
 {
 "name": "router_minus_baseline",
 "type": "number"
 },
 {
 "name": "router_minus_baseline_CI",
 "type": "string"
 }
 ]
 },
 "data/router/six_setting_test_predictions.csv": {
 "format": "csv",
 "rows": 5832,
 "role": "label",
 "description": "Held-out predictions for 3 routers across 6 settings. Contains oracle labels.",
 "bytes": 1115319,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "router",
 "type": "string",
 "enum": [
 "metadata_bucket_majority",
 "metadata_logreg",
 "text_metadata_logreg"
 ]
 },
 {
 "name": "example_id",
 "type": "string"
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "source_problem_id",
 "type": "string"
 },
 {
 "name": "oracle_label",
 "type": "string",
 "enum": [
 "baseline_llm",
 "single_agent",
 "per",
 "broadcast",
 "none"
 ]
 },
 {
 "name": "predicted_label",
 "type": "string",
 "enum": [
 "baseline_llm",
 "single_agent",
 "per",
 "broadcast",
 "none"
 ]
 },
 {
 "name": "predicted_success",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "baseline_correct",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 },
 {
 "name": "any_protocol_solved",
 "type": "integer",
 "enum": [
 0,
 1
 ]
 }
 ]
 },
 "data/router/six_setting_text_metadata_verification.csv": {
 "format": "csv",
 "rows": 6,
 "role": "aggregate",
 "description": "Independent recomputation of the text+metadata router row. Two avg_tokens cells are nan where per-problem cost accounting was missing.",
 "bytes": 1813,
 "columns": [
 {
 "name": "display_name",
 "type": "string"
 },
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "same_test_ids_all_router_rows",
 "type": "boolean"
 },
 {
 "name": "baseline_solve",
 "type": "number"
 },
 {
 "name": "router_solve",
 "type": "number"
 },
 {
 "name": "oracle_solve",
 "type": "number"
 },
 {
 "name": "cost_n",
 "type": "integer"
 },
 {
 "name": "missing_cost_n",
 "type": "integer"
 },
 {
 "name": "avg_tokens",
 "type": "number"
 },
 {
 "name": "avg_excess_tokens",
 "type": "number"
 },
 {
 "name": "oracle_macro_f1",
 "type": "number"
 },
 {
 "name": "route_Baseline",
 "type": "number"
 },
 {
 "name": "route_Single",
 "type": "number"
 },
 {
 "name": "route_PER",
 "type": "number"
 },
 {
 "name": "route_Broadcast",
 "type": "number"
 },
 {
 "name": "route_None",
 "type": "number"
 }
 ]
 },
 "data/splits/primary_omnimath_splits.csv": {
 "format": "csv",
 "rows": 4181,
 "role": "feature",
 "description": "80/10/10 split, seed 42, stratified by oracle label. OmniMath only. The paper's 423-problem held-out split.",
 "bytes": 419329,
 "columns": [
 {
 "name": "benchmark_id",
 "type": "string",
 "enum": [
 "omnimath2",
 "jeebench",
 "scibench",
 "labbench"
 ]
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "legacy_tier_id",
 "type": "string"
 },
 {
 "name": "split",
 "type": "string",
 "enum": [
 "train",
 "dev",
 "test"
 ]
 },
 {
 "name": "scheme",
 "type": "string"
 }
 ]
 },
 "data/splits/six_setting_splits.csv": {
 "format": "csv",
 "rows": 12928,
 "role": "feature",
 "description": "70/15/15 split, seed 20260712, stratified by oracle label, drawn independently per setting.",
 "bytes": 2697319,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "example_id",
 "type": "string"
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "source_problem_id",
 "type": "string"
 },
 {
 "name": "split",
 "type": "string",
 "enum": [
 "train",
 "dev",
 "test"
 ]
 },
 {
 "name": "scheme",
 "type": "string"
 }
 ]
 },
 "registry/benchmarks.csv": {
 "format": "csv",
 "rows": 5,
 "role": "registry",
 "description": "Per-benchmark license and redistribution decision.",
 "bytes": 1054,
 "columns": [
 {
 "name": "benchmark",
 "type": "string"
 },
 {
 "name": "upstream_url",
 "type": "string"
 },
 {
 "name": "version",
 "type": "string"
 },
 {
 "name": "license",
 "type": "string"
 },
 {
 "name": "redistribution_decision",
 "type": "string"
 },
 {
 "name": "n_problems",
 "type": "integer"
 },
 {
 "name": "text_included",
 "type": "string"
 },
 {
 "name": "reconstruction_doc",
 "type": "string"
 }
 ]
 },
 "registry/example_id_crosswalk.csv": {
 "format": "csv",
 "rows": 12928,
 "role": "registry",
 "description": "Maps (setting_id, example_id) -> canonical problem_id plus the run-native source_problem_id, for all 12,928 probe examples.",
 "bytes": 1836950,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "example_id",
 "type": "string"
 },
 {
 "name": "problem_id",
 "type": "string"
 },
 {
 "name": "source_problem_id",
 "type": "string"
 }
 ]
 },
 "registry/experiments.csv": {
 "format": "csv",
 "rows": 10,
 "role": "registry",
 "description": "One row per released setting.",
 "bytes": 1906,
 "columns": [
 {
 "name": "setting_id",
 "type": "string"
 },
 {
 "name": "benchmark",
 "type": "string"
 },
 {
 "name": "condition",
 "type": "string"
 },
 {
 "name": "solver",
 "type": "string"
 },
 {
 "name": "model_endpoint",
 "type": "string",
 "enum": [
 "openai/gpt-oss-120b",
 "google/gemma-4-31B-it"
 ]
 },
 {
 "name": "n",
 "type": "integer"
 },
 {
 "name": "run_id",
 "type": "string"
 },
 {
 "name": "protocols_covered",
 "type": "string"
 },
 {
 "name": "analyses_covered",
 "type": "string"
 }
 ]
 },
 "registry/models.csv": {
 "format": "csv",
 "rows": 4,
 "role": "registry",
 "description": "Solver and prober model identifiers and terms.",
 "bytes": 826,
 "columns": [
 {
 "name": "solver_id",
 "type": "string"
 },
 {
 "name": "hf_or_provider_identifier",
 "type": "string"
 },
 {
 "name": "role",
 "type": "string"
 },
 {
 "name": "notes",
 "type": "string"
 }
 ]
 },
 "registry/protocols.csv": {
 "format": "csv",
 "rows": 5,
 "role": "registry",
 "description": "The four protocols plus the 'none' router action.",
 "bytes": 712,
 "columns": [
 {
 "name": "protocol_id",
 "type": "string"
 },
 {
 "name": "display_name",
 "type": "string"
 },
 {
 "name": "description",
 "type": "string"
 },
 {
 "name": "oracle_rank",
 "type": "integer"
 },
 {
 "name": "color_hex",
 "type": "string"
 }
 ]
 }
 }
}