Build-and-Find

Build-and-Find is a protocol and benchmark for evaluating repositories that coding agents write for other agents to use. A builder agent receives a hidden specification and writes a codebase; a finder agent is given only that codebase and a specification-traced multiple-choice question bank. The benchmark reports whether finders recover the intended behavior and design choices (recovery accuracy, repeatability, implementation coverage) and how much inspection the recovery takes (inspection effort), with question-only and spec-only controls. This dataset holds the task pack, the generated artifacts, every run record, the canonical tables, the derived reports, and the post-review audit records behind the paper.

• Paper: https://openreview.net/forum?id=y48mj8HiWu
• Code: https://github.com/JacobLinCool/Build-and-Find
• Version: 1.1.0 (camera-ready), revision v1.1.0 of this dataset; changes from the reviewed v1.0.0 are listed in CHANGELOG.md.

Dataset structure

The data files have the same paths as in the code repository. The seed projects and the edit_* tasks of the code repository are not used by the reported study and are not included. To check the files with make release-check, download the data directories into a checkout of the code, from its root:

HF_DATASET_URL="https://huggingface.co/datasets/JacobLinCool/Build-and-Find"
uvx --from huggingface_hub hf download "${HF_DATASET_URL#https://huggingface.co/datasets/}" \
  --repo-type dataset --revision v1.1.0 \
  --include "benchmarks/**" --include "runs_release/**" --include "rebuttal/**" \
  --local-dir .

Downloading the whole dataset into a checkout would replace the code repository's README.md with this card. croissant.json is the same file as metadata/croissant.json in the code repository: it gives each file as its URL at revision v1.1.0 and each file set as paths in this repository.

path	contents	license
benchmarks/specs/, benchmarks/questions/	two Rust build-from-scratch task specifications (scratch_minidb, scratch_nanoweb) and their question banks, 15 questions each, with gold answers and specification traces	CC BY 4.0
runs_release/records/	1,924 sanitized run records: 48 builds, 1,728 artifact-conditioned find runs, 144 control find runs, 4 control build records	CC BY 4.0
runs_release/artifacts/build-*/	the 48 generated codebases, one directory per build run	Apache-2.0
runs_release/artifacts/control-build-*/	the 4 control-build workspaces of the question-only and spec-only controls (CONTROL.md, and a copy of the task specification in the spec-only workspaces)	CC BY 4.0
runs_release/tables/	canonical tables (build_runs.csv, find_runs.csv, question_answers.csv, panel_membership.csv, experiments.json) and MANIFEST.json	CC BY 4.0
runs_release/reports/	derived reports, including the v1.1.0 layers listed below	CC BY 4.0
rebuttal/full_conformance_audit/	post-review source-trace audit of the implementation labels: protocol, audit tables, evidence excerpts with source-file hashes, manifests, decisions, and the adjudicated-label sensitivity analysis	CC BY 4.0
rebuttal/conformance_spot_audit/	post-review spot audit of 24 consensus prelabels	CC BY 4.0
croissant.json	Croissant 1.1 metadata with Responsible AI fields (metadata/croissant.json in the code repository)	CC BY 4.0
LICENSE-DATA, LICENSE-CODE	license notices: CC BY 4.0 for the data, Apache-2.0 for the generated codebases
CHANGELOG.md	changes from v1.0.0 to v1.1.0	CC BY 4.0

The v1.1.0 report layers are runs_release/reports/eandd_final/control_exclusions/, runs_release/reports/eandd_final/transcript_audit/, runs_release/reports/option_free_check/, runs_release/reports/runtime_tests/, runs_release/reports/artifact_conformance_consensus/final/label_corrections_v1_1.csv, and runs_release/reports/artifact_conformance_consensus/impact/conditioned_builder_scores_by_task.csv.

Configurations

config	file	rows	one row per
find_runs (default)	runs_release/tables/find_runs.csv	1,872	find run (artifact-conditioned or control)
question_answers	runs_release/tables/question_answers.csv	28,080	answered question
build_runs	runs_release/tables/build_runs.csv	52	build run, including the 4 control build records
panel_membership	runs_release/tables/panel_membership.csv	151,428	(panel, table, row) membership
implementation_labels	.../final/artifact_conformance_final.csv	720	artifact-question pair (v1.0 labels)
label_corrections_v1_1	.../final/label_corrections_v1_1.csv	15	corrected artifact-question pair
author_decisions	rebuttal/full_conformance_audit/author_decisions.csv	43	audit flag decided by the author
model_assisted_decisions	rebuttal/full_conformance_audit/model_assisted_decisions.csv	162	audit flag decided by model-assisted adjudication and approved by the author

The tables keep every released run. Analyses select rows by the panel_ids column: paper_final_scratch is the primary panel (status-ok, compile-pass builds), question_only and spec_only are the controls, and artifact_present_diagnostic adds the compile-failed and empty artifacts. The v1.1.0 question-only control excludes the 11 runs listed in runs_release/reports/eandd_final/control_exclusions/excluded_question_only_runs.csv; those runs are still in the tables. Implementation-aware analyses apply label_corrections_v1_1.csv on top of artifact_conformance_final.csv.

from datasets import load_dataset

repo_id = "https://huggingface.co/datasets/JacobLinCool/Build-and-Find".removeprefix("https://huggingface.co/datasets/")
find_runs = load_dataset(repo_id, "find_runs", split="test")
answers = load_dataset(repo_id, "question_answers", split="test")

Intended use

• Evaluating generated coding-agent artifacts by whether a calibrated finder panel recovers specification-traced design intent from them, and how much inspection the recovery takes once the recovery and stability gates hold.
• Reproducing the paper's tables and reports from the stored records, inspecting the analysis, and auditing the implementation labels.
• Comparing new builders, finders, or harnesses against the frozen evidence contract, using the scripted core finder panel of the code repository.

Out-of-scope use

• Certifying deployment safety, security, runtime correctness beyond the recorded compile probe and builder-written tests, or human maintainability.
• Public leaderboards on this task pack or universal single-axis model rankings: the results are panel-local and the task pack is public.
• Training a model on the task pack and then evaluating it on the same pack.
• Treating the multiple-choice accuracy as an artifact-quality score: on this high-prior pack it is a validity gate.

Data collection

Runs were collected with the Build-and-Find harness between 30 April and 3 May 2026 (UTC) from twelve agent configurations: six models, each at high and low reasoning effort.

CLI	models
Codex CLI	gpt-5.5, gpt-5.4-mini
Claude Code CLI	claude-opus-4-7, claude-sonnet-4-6, mimo-v2.5-pro, mimo-v2.5 (MiMo through an Anthropic-compatible endpoint)

• Builds. Each of the 12 builders built each task twice from the hidden specification: 48 builds, 41 of which pass the compile probe (cargo build --release).
• Finds. Each of the 12 finders answered the question bank of each of the 48 artifacts three times: 1,728 artifact-conditioned find runs. Option order is shuffled per run with a recorded seed, answers are written to a file in the finder's workspace, and effort is recorded as bytes exchanged with the model.
• Controls. Each finder answered each question bank three times with no artifact (question-only) and three times with the specification only (spec-only): 144 control find runs.
• Environment. Builder and finder workspaces were per-run directories inside the harness repository. The agent CLIs ran with permission prompts disabled and without an OS-level sandbox.
• Post-review layers (July to October 2026). A transcript check of the find runs; an option-free check in which GPT-5.5-high and Opus 4.7-high answered eight low-prior questions without answer choices (24 sessions, 96 answers); one run of each artifact's builder-written test suite (cargo build --release and cargo test); and model-assisted source-trace audits of the implementation labels.

Annotations

• Gold answers. The question banks and their gold answers were written by the author with LLM assistance and reviewed by the author. Every question carries a source trace to a specification intention.
• Implementation labels (implementation_labels). Each artifact-question pair was prelabeled when the stable answers of the Codex GPT-5.5-high and Claude Opus 4.7-high finders agreed (696 pairs); the other 24 pairs were labeled by manual review. The prelabels are triage labels, not independent conformance annotations.
• v1.1.0 correction (label_corrections_v1_1). The one build without source files has its 15 pairs relabeled absent_or_ambiguous (label_source = rule_empty_artifact).
• Source-trace audit and adjudication. After review, a model-assisted source trace of the 592 scored compile-pass pairs (a primary and a cross-check pass) flagged 205 pairs. Every flag has a recorded decision: 43 by the author (author_decisions; 41 relabeled, 2 kept) and 162 drafted by model-assisted adjudication, with two passes (the second without access to the audit verdicts) and arbitration where they disagreed, then reviewed and approved by the author (model_assisted_decisions; 154 relabeled, 8 kept; decided_by records the model-assisted source). The adjudicated labels enter the sensitivity analysis in rebuttal/full_conformance_audit/label_sensitivity_adjudicated/; the main analyses use the released labels with the v1.1.0 correction.
• Option-free grading. The author graded the 96 free-text answers against a rubric fixed before the sessions, in two frozen passes (claim correctness, then evidence sufficiency). An LLM labeled the same rows for a consistency check only.

No crowdsourced or human-subject annotation was used. The raw agent sessions of the audits are not included; the audit tables keep each session's id and result hash.

Corrections in v1.1.0

v1.1.0 adds files and changes analysis outputs; the v1.0.0 records, tables, artifacts, and implementation labels are unchanged.

• Control exclusions. 11 of the 72 question-only control runs read files outside their workspace (3 of them answer-bearing files) and are excluded from the question-only control. Question-only accuracy changes from 94.5% to 97.2%.
• Matched control differences. Differences from the question-only control are computed over matched finder, task, and trial strata with bootstrap intervals.
• Label corrections and adjudication records. The empty-artifact correction above, and the audit decisions with the adjudicated-label sensitivity analysis.
• Transcript audit. No primary-panel finder read the specification or the gold-annotated question files; excluding the 62 primary-panel runs whose transcripts reach outside their workspace changes accuracy by at most 0.11 pp and leaves every R_b rank unchanged.
• Added checks. The option-free check, the builder-written test-suite runs (the full suite passes in 40 of the 41 compile-pass builds), and per-task R_b.
• R_b intervals. The bootstrap now keeps repeated draws of a cell as separate cells, which widens some intervals; point estimates and ranks are unchanged.

Limitations

• High-prior task pack. Question-only accuracy is 97.2%, so accuracy is a validity gate, and inspection effort carries the artifact comparison only after the recovery and stability gates. Two Rust tasks and twelve agent configurations make every result panel-local.
• Unsandboxed runs. The agents could read files outside their workspaces. The control exclusions and the transcript audit document which runs did; other effects of the shared environment cannot be ruled out.
• Label provenance. The scored implementation labels come from finder consensus (589 pairs) and manual review (3 pairs). Post-submission checks found 9 label errors, and under the adjudicated labels 195 of the 592 scored pairs move away from implemented-gold. Unflagged pairs were not re-reviewed, and builder rankings are conditional on the label set.
• Question construction. The question banks were written with LLM assistance; the drafting and distractor-selection histories were not retained.
• Correctness. The compile probe and the builder-written tests indicate self-consistency, not specification conformance.
• Drift and contamination. Hosted models and CLIs change, so new runs are not bit-identical; seeds reproduce harness-side randomization only. The public task pack may enter training data; contamination-resistant studies need private packs with the same schema.

Personal and sensitive information

The dataset contains no human-subject data and no personal or sensitive information. Run records hold model and adapter ids, seeds, hashes, provider session ids, timestamps, answers, and usage metrics; local paths and per-run session logs were removed before release. Audit records name the author or model-assisted adjudication as the decision source.

Licenses

Data are released under CC BY 4.0 (LICENSE-DATA). The generated codebases under runs_release/artifacts/build-*/ are source code under Apache-2.0 (LICENSE-CODE), like the harness in the code repository. License fields that the builder agents wrote into the codebases, such as license entries in Cargo.toml or license sections in a README, are generated content kept as produced, not license grants; the release licenses every generated codebase under Apache-2.0. The control-build workspaces under runs_release/artifacts/control-build-*/ are data under CC BY 4.0.

Citation

@inproceedings{buildandfind2026,
  title     = {Build-and-Find: An Effort-Aware Protocol for Evaluating Agent-Managed Codebases},
  author    = {Jhen-Ke Lin},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS 2026), Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://openreview.net/forum?id=y48mj8HiWu}
}

Maintenance

Versions are released as new files on top of the prior records. Questions and corrections: issues in the code repository (https://github.com/JacobLinCool/Build-and-Find) or jacob.cs14@nycu.edu.tw.