PeterStrain77/bounded-returns-deep-research

Bounded Returns to Orchestration — Deep Research Evaluation Frames

Judge-level evaluation data from a controlled comparison of automated deep-research architectures over a 90-query manifest. These are the tidy frames the paper's statistics are computed from: the reported numbers are recomputable from the tables here using the analysis scripts in the code repository.

The paper compares eleven architectures on a single shared tool layer: a single-pass baseline, eight orchestration pipelines (P0–P8, GPT-4o), and two local 7B agents (P9, P10), with model and tools held fixed so that orchestration is the variable. Those eleven are the rows of the paper's headline table.

The code repository additionally implements seven later probes and controls (P11–P17), for 18 pattern implementations in total. Two of them appear in these frames: the reference ranking covers 13 arms, but base_p11 and base_p12 were never scored by Claude Opus, so their scores rest on two judges and they are reported as supporting analyses rather than leaderboard entries.

The pattern column carries 73 distinct labels in verdicts, scores, overall_scores and runs (the other configs carry 9–12), because the frames cover far more than the headline arms:

Group	Count	Example
Leaderboard arms	13	base_p0 … base_p12
Variance replicates	32	base_p0_v1 … base_p0_v11
Ablations	8	ablation_p4_no_triangulation
Oracle-retrieval arms	9	oracle_t1_p0 (the e14_oracle_* frames carry 11)
Tavily search arm	6	protocol_a_tavily_p0
Tool-layer disentanglement	2	disentangle_matched_p1
Other probes	3	base_p1_7b, base_p11_16turn

Do not filter on base_p* for a leaderboard view — that prefix matches 48 labels, including every variance replicate. Select the 13 arms explicitly (base_p0 … base_p12), or filter pattern_family == "base" and drop the _v<n>, _7b and _16turn suffixes.

Judgments come from a four-judge panel — gpt52, claude_opus, claude_sonnet, and claude_code — so most frames carry a judge column. Group by it rather than pooling across judges.

What's here

Config	Rows	Description
queries	90	The evaluation manifest: query text, source benchmark, stratum
runs	6,570	One row per (pattern × query) execution, with cost and telemetry
overall_scores	6,509	Per-report aggregate score, per judge
scores	58,552	Per-report × per-rubric-dimension scores
verdicts	248,536	Per-criterion judge verdicts (binary SATISFIED + chain-of-thought), sharded across 6 parquet files
citations	22,903	Extracted citations with URL-resolution status
citations_protocol_a	6,241	Citation subset under the stratified Protocol A sample
c0_verdicts / c0_per_report	3,096 / 269	Claim-level factual verification
e14_oracle_verdicts / e14_oracle_per_report	6,365 / 327	Oracle-retrieval entailment arm

Scoring uses a 9-dimension rubric (information recall, factual accuracy, coverage, analytical depth, citation quality, logical coherence, organization, instruction following, attribution quality), scored as binary per-criterion verdicts with chain-of-thought and aggregated to weighted dimension scores.

Read DATA_DICTIONARY.md (in this dataset, alongside the frames) before using them. It documents every column and, importantly, a set of known upstream data issues that are recorded rather than silently repaired — including a corrupted overall_score for one judge where a _recomputed column must be used instead. Analyses that ignore that section will produce wrong numbers.

Quick start

import pandas as pd

scores = pd.read_parquet("hf://datasets/PeterStrain77/bounded-returns-deep-research/df_overall_scores.parquet")
print(scores.groupby(["judge", "pattern"])["overall_score"].mean().unstack(0))

Or with datasets:

from datasets import load_dataset

ds = load_dataset("PeterStrain77/bounded-returns-deep-research", "overall_scores", split="train")

Reproducing the paper's numbers

The frames alone are not the analysis. The code repository ships the full pipeline implementations, the evaluation harness, and ~115 statistics scripts that turn these tables into the paper's reported values:

If you clone the code repository you already have these frames — they ship in data/analysis/, including an unsharded df_verdicts.parquet. Nothing below is needed; just run the analysis.

The steps below are for using the Hub copy on its own. The analysis scripts read from data/analysis/ and expect df_verdicts.parquet as a single file, so the shards must be reassembled:

git clone https://github.com/Peter-McCann-Strain/bounded-returns-deep-research
cd bounded-returns-deep-research
pip install -e ".[paper]" huggingface_hub   # hub client is only needed to fetch

# 1. Pull these frames into data/analysis/ (where the scripts look for them).
python - <<'EOF'
from huggingface_hub import snapshot_download
import glob, shutil, pathlib, pandas as pd

src = snapshot_download("PeterStrain77/bounded-returns-deep-research",
                        repo_type="dataset")
dst = pathlib.Path("data/analysis"); dst.mkdir(parents=True, exist_ok=True)
for f in glob.glob(f"{src}/*.parquet"):          # parquet only — do not overwrite
    shutil.copy(f, dst)                          # the repo's own docs with Hub copies

# 2. Reassemble the sharded verdict frame into the single file scripts expect.
shards = sorted(dst.glob("df_verdicts-*-of-*.parquet"))
pd.concat([pd.read_parquet(s) for s in shards], ignore_index=True) \
  .to_parquet(dst / "df_verdicts.parquet", index=False)
print(f"reassembled {len(shards)} shards -> data/analysis/df_verdicts.parquet")
EOF

# 3. Recompute the paper's headline statistics.
python papers/paper_a_bounded_returns/analysis/build_numbers.py

build_numbers.py recomputes every headline statistic from the parquet files and writes canonical_numbers.json, which is the paper's single source of truth. Fourteen of the analysis scripts read df_verdicts.parquet directly, so if you took the Hub-only route, skipping the reassembly step will make them fail with a missing-file error.

Run the analysis scripts on a throwaway checkout: several read results/**, which is not redistributed, and overwrite shipped files with empty results rather than stopping. See the code repository's README for the full warning.

Revision note (August 2026)

This is a corrected revision. Three changes matter to anyone who downloaded an earlier copy:

• One benchmark query's real-world identity is now removed everywhere. The query names a real small business and a named individual. Earlier redaction reached the query prompt only, so the identity survived in the citations the agents actually retrieved and in the judge text quoting them: business domains and URL slugs in citations.cited_url/domain/ cited_title, the trading city, two community names, and 16 cells of verdicts reasoning. The rubric criteria in eval_queries_v2.json leaked it too, having been written against the unredacted prompt. All are now [REDACTED].

• A third party's email address was removed. A scraped academic byline left a researcher's address in citations.cited_title (2 rows) and in one verdicts.evidence quote. The quote's analytic content is preserved -- the judge was citing it as an example of a malformed reference -- with the address replaced by [email redacted].

  Both redactions are scoped to the affected query_id, so nothing else moved: row counts are unchanged (22,903 citations, 248,536 verdicts, 359,458 across all frames), every pattern mean in the paper still recomputes to the same value, and unrelated text that merely contains a matching substring -- Yellowstone's "Mesa Falls", "nevertheless", "cleverthai.com" -- is untouched.

• The analysis scripts that consume these frames were corrected. The cluster bootstrap resampled clusters without preserving draw multiplicity, which made every confidence interval and p-value derived from it too narrow, and 32 build scripts overwrote sibling keys in the results store instead of merging into it. Numbers recomputed from these frames with the current scripts will differ from a run of the older ones.

Scope and limitations

• Judge scores are model judgments, not ground truth. The repository ships the human-calibration protocol and judge-vs-human agreement analysis; consult those before treating a score as an absolute quality measure.
• report_path in runs points into a results/ tree that is not redistributed. The generated reports are large and carry mixed third-party content. Paths resolve only in a full working checkout.
• Rows are not independent. They are clustered by query and by pattern; naive pooled statistics will understate uncertainty. The analysis scripts use clustered and stratified estimators for this reason.
• Mixed licensing on data-derived rows. Query text and cited content derive from multiple upstream benchmarks under different terms. Apache-2.0 covers the code; see DATA_LICENSES.md and NOTICE in the code repository before redistributing data-derived material.

Citation

@software{mccannstrain_deep_research,
  author = {McCann Strain, Peter},
  title  = {Deep Research: Bounded Returns to Orchestration},
  year   = {2026},
  doi    = {10.5281/zenodo.21118281},
  url    = {https://github.com/Peter-McCann-Strain/bounded-returns-deep-research}
}