BVI-Val: ranking and field choice in budgeted price prediction

Research code and aggregate evidence for Risk Ranking, Benefit Ranking, and Field Choice in Budgeted Vehicle Advertised-Price Prediction.

Access status and inspection

Public release; unauthenticated ZIP download verified on 5 October 2026. The exact approved archive, original-software licence and licensing-scope documents are committed in this repository. The archive was downloaded without login, extracted into a new directory, and checked against all 298 file hashes and the four frozen recipe checks. This verifies access and payload integrity, not empirical retraining or independent human reproduction.

Pinned archive. The directly browsable release tree was also downloaded without login and verified: 298 files, zero missing, unexpected or changed files. All four subpackage checks passed on the actual extraction. This separate check preserves the archive's original bytes.

Release filename: BVI-Val_code-aggregates_2026-10-05_v2.zip. Version: 2026-10-05-v2; 298 files inside the extracted package, including 297 payload files and its manifest. ZIP SHA256:

81bb9acd9c1a2a3f6a35203c2538dbc90e2382ef6a6e3e1a217ab51cd99484a1

Extract the pinned ZIP and enter public-code-aggregates-2026-10-05-v2/, then run:

python3 -I -B -S CHECK_RELEASE.py

Python 3.12+ is sufficient for this integrity check; no source data, installed study packages or network access is required. It verifies hashes and four frozen recipe subpackages. It does not perform empirical retraining. Keep generated outputs outside the extracted package. Alternatively, after cloning this repository, run make check at the outer repository root.

For the synthetic tests, provision a recipe-compatible research interpreter first and run make test BVI_PYTHON=/absolute/path/to/python. The source-inspection tests passed on the publicly downloaded package in an existing author-host environment (360 tests, 31 subtests); this is not a fresh installation or independent empirical replication. The integrity-only GitHub workflow uses the standard library and does not run these dependency-requiring tests or train models. Its execution status is visible under Actions, not inferred from the presence of the workflow file.

What is inside

Turkey historical reconstruction companion

The separate 108-file companion and ZIP are public. Their unauthenticated download, exact inventory, file bytes and extracted integrity check passed on 5 October 2026, with zero missing, unexpected or changed files. ZIP SHA256: acfc1e45cafa8296e16f7885b0cf76ad89b36e97e91cab9fc0a116b26b7aac20.

Follow its README: obtain the pinned workbook separately, run the embedded development recipe through preparation/readiness/fitting, then use the companion for scoring and historical outcomes (not the embedded legacy --score). Author-host execution completed the full original 258/21/6 fitting schedule and both complete calibration/test outcome structures: 2,011 scalar leaves matched with the original 1e-12 tolerances. This is reconstruction of already-opened historical results, not new confirmation, independent label custody, another-machine replication or coverage of every SI/GDFS diagnostic. No private rows, models, target caches, execution receipts or ledgers are included.

At the outer repository root, make check includes this companion. For its synthetic tests, use make test-turkey BVI_PYTHON=/absolute/path/to/python with the specified main-role dependencies. Use python -B -m pytest, not the frozen README's example without -B, to avoid adding bytecode files to the strict package inventory. Keep empirical execution outputs outside the package.

Path inside the released directory or ZIP	Content
AGGREGATE_INDEX.md and aggregates/	Source-specific results, negative findings, diagnostic summaries and interpretation index
recipes/at/	Austrian commercial-vehicle source-to-results recipe and synthetic guard tests
recipes/historical-primary/	Source-directed MUCars, JUCars and AutoScout24 primary reconstruction
recipes/historical-comparators/	Fixed-field benefit and neighbor controls using privately rebuilt primary artifacts
recipes/source-inspection/	Frozen implementation-inspection snapshot, including Turkey fitting/scoring helpers; not a complete Turkey retraining recipe
SOURCE_UNITS.md	Currency and price-unit evidence boundaries
release-manifest.json	Exact per-file hashes and provenance

Read each recipe's README before obtaining the specified source version from its depositor. No raw marketplace rows, row-level predictions/actions/labels, trained models, private execution trees, full manuscript or author declaration forms are distributed. The archive has original research software, not bundled dependency wheels. Its hash-pinned environment targets macOS arm64/CPython 3.12.14; it is not a cross-platform installation claim.

Research scope

The task simulates revealing recorded specifications to reduce advertised-price prediction error under a fixed action capacity. Historical gains over risk routing vary by source. Stronger fixed-field benefit controls explain much of the apparent field-choice increment. In the locally prospectively fixed AT follow-up, joint benefit and fixed-power benefit policies select identical actions on 3,251 evaluation records (326 actions each); their MAE is EUR 7,364.4627. The neighbor policy has MAE EUR 7,318.0484; fitted-head superiority is not established. The archive retains the historical positive findings, negative results, failed pilot and distinct Turkey H2 comparison.

All released evaluations have already been opened: replay is reconstruction, not new confirmation. AT shares the AutoScout platform family and includes new/used commercial vehicles. Turkey/GDFS aggregate diagnostics and implementation-inspection entry points remain in the earlier release; the separate companion now supplies Turkey historical training/scoring/outcomes, but not complete GDFS or all-SI reconstruction. Two additional AT diagnostics use frozen scores or non-price source information only; they are post-result characterization, not new confirmatory performance evidence. The study makes no transaction-price, real verification-cost, profit or universal policy-superiority claim. Local reconstruction is not independent third-party empirical reproduction.

Licensing and version history

MIT applies to the project-original software within the specified scope, not source data or dependencies. See third-party notices. Dataset use and redistribution remain subject to source terms; public download does not adjudicate underlying marketplace rights. Aggregate evidence is not a blanket relicensing of data.

The byte-preserved package README and manifests retain preparation-time access text and flags, including public_payload_download_verified=false, publicly_archived=false and project_license_selected=false, for provenance. This outer repository README records current access; it does not rewrite those snapshots or their checks. The checker's PREPARED_PAYLOAD_INTEGRITY_VERIFIED status is an integrity statement, not today's public-access status. GitHub commits identify versions. This is not a DOI deposit, anonymous archive, journal submission or acceptance.

中文：代码与聚合结果压缩包已公开，无登录下载、解包及完整性检查已通过。 展开目录也已通过 298 文件逐字节核验，缺失、额外及差异均为零。 外层仓库 README 是当前访问说明；包内准备状态文本保留为冻结历史。 先读聚合索引，再运行完整性检查；已有研究环境可另运行合成测试。 原始车辆行和模型不公开；“检查通过”不代表完整实验复现或达到录用门槛。 子包旧标志是历史快照状态，不是新的公开访问证明。