Datasets:

AIPOCH-AI
/

open-science-cases

Field	Value
Case ID	`case21`
Case title	Dietary diversity and gut community composition
Case folder	`cases/case21/`
Product / feature	AIPOCH Open-Science research conversation, Plan, Notebook, automatic review and session export
Author	mumumm
Submitter	mumumm
Created / updated	2026-10-08 / 2026-10-08
Status	draft
Case type	evaluation
Product / model version	OpenScience 0.35.1 per review record; model gpt-6-astra; backend codex:builtin-codex-isolated; reasoning medium

Goal and context

Assess reported plant-type categories and gut microbial community composition in historical American Gut data, with geography and questionnaire-missingness sensitivity.

This submission selects attempt 20261008-light-3, not the first run. The selected export is named 21-2-2026-10-08.science; filename numbering is not the native attempt number. The new analysis uses 600 participants, processed 100-nt features and Lingoes-corrected Bray–Curtis geometry. It is not a wording-only repair or an exact replication of the first attempt. Permitted Plan review and operator follow-ups are preserved and do not themselves constitute a failure.

Inputs and sources

The unchanged session.science export preserves the exact user-role request sequence.

Source: American Gut commit 68fd6d4b2fa6aeb5b4f5272c6f1006defe5b160e, a May 2015-or-earlier data/survey snapshot. The captured files are:

AG_100nt.biom: SHA-256 35e474a374acb6bbf53ba7c2d1ba750b5e3c579eed0b8fc325e63f309cb503bb.
AG_100nt.txt: SHA-256 e37d1c5b2e17f62c194da231e499611a391bea76e4d15fe2d1a5b55166fcea4b.
README.md: SHA-256 a01016dba296950fb798c0dcb577676d9b776bbfa4d6ba06567cbc9030c0d30f.
metadata_charts.json: SHA-256 e13dd86e7c20635eafe12c84ea3f582067614a96b71c8c75ffe34968d440bc01.

Historical acquisition occurred during the 2026-10-08 original run; the saved download manifest does not record per-file timestamps. The source README describes open-access/deidentified data, but a specific redistribution license was not established from the inspected records. Preserve attribution and verify applicable source terms before redistribution. No new source retrieval or license verification was performed during preparation; the repository license does not relicense third-party inputs.

Export completed 2026-10-08T16:58:25.760000+08:00, Asia/Shanghai. Source filename 21-2-2026-10-08.science, size 30,806,918 bytes, SHA-256 ec457ca8f832735c9e7b3414c3145fd472f48c4bad7dbb1d439ffbdb29df5672. Archive validation does not independently certify sensitive-content screening.

Runtime and prerequisites

Recorded Python: 3.12.3 (tags/v3.12.3:f6650f9, Apr 9 2024, 14:05:25) [MSC v.1938 64 bit (AMD64)]. Libraries: numpy 2.4.6, pandas 3.0.5, scipy 1.17.0, matplotlib 3.8.4, h5py 3.16.0, threadpoolctl 3.6.0. Pillow is also used for image metadata; its exact version is not recorded in this validation file. A compatible existing runtime is required; the interpreter and a full environment lock are not bundled. No installation or product changes were performed.

Reproduction steps

Import session.science into a compatible OpenScience version. Fresh import was not performed during preparation.
Inspect the Plan, analysis_rules.md, study_report.md and plant_diversity_reproducible.zip.
Extract the bundle into a new directory. Keep ag_analysis/ captured inputs and reference_results/ intact.
With the recorded dependencies already available, run from the extracted directory:

python analysis.py

With provided inputs this needs no network; the script can download missing files from pinned URLs, so retain and checksum-check all included inputs for an offline replay. It writes ag_results/ and refuses an existing output directory. A fresh alternative directory can be selected through AG_OUTPUT_DIR.

Compare model/cohort/feature/input-hash outputs at recorded tolerances rtol=1e-10 and atol=1e-12. Eigenvector signs may change between numerical libraries; compare reconstructed distances or coordinate Gram matrices rather than signs. The displayed two-axis ordination is not the full inferential geometry. This command was documented, not executed during packaging; prior Notebook replay is distinct from a new replay or native Reproduction.

Exact prompt or request

See the session history. Original session messages are preserved verbatim; automatic-review messages are labeled separately. The later checking branch is documented below, not merged into the research session.

Expected result and acceptance criteria

Criterion	How to check	Pass condition
Bounded analysis	Request, Plan, input manifests and cohort audit	Specified data, local methods and sample scope respected
Numerical consistency	Report, saved tables and code	Correct denominators, transformations, estimates, uncertainty and multiple-testing treatment
Interpretation	Report opening, methods, limitations and conclusion	Claims limited to observational evidence and measured quantities
Inspectable delivery	Session inventory and offline bundle	Captured inputs, code, provenance, report, tables and standalone figures agree
Reproduction	Authorized offline replay	Regenerated outputs agree with saved results under recorded tolerances

Actual result

The source contains 4,827 matching samples and 22,742 features. Eligibility and participant deduplication leave 2,719 participants, from whom a fixed-seed country-proportional sample of 600 was selected. Main analysis includes 564 respondents with known country; 479 complete cases, 540 US respondents and a 300-person eight-state subset support sensitivity analyses.

All 9,089 nonzero cohort features were retained. Bray–Curtis geometry had 259 materially negative eigenvalues; Lingoes correction yielded 598 retained coordinate dimensions. All dimensions were used for inference; display axes explain 4.40% and 2.47% of corrected inertia.

The primary country-adjusted association has incremental R² 0.0080399356 (0.804%), partial R² 0.0081978586 (0.820%), pseudo-F 1.1344562, and permutation p=.004 using 999 permutations. This is a small association in corrected geometry, not the fraction of raw-count biological variance explained or a causal benefit. Multiplicity-adjusted complete-case and regional sensitivity tests are less conclusive.

The new mean centroid distances decrease from 0.840916 to 0.808207 across increasing plant categories; “More than 30” is correctly identified as the minimum. The previous attempt's range defect is absent. Dispersion F is 2.3986004, with saved permutation p=.046; unequal dispersion limits a pure centroid interpretation.

Artifacts inside the package: analysis_rules.md, analysis.py, study_report.md, plant_diversity_reproducible.zip, ordination.png, and sensitivity_missingness.png. The bundle additionally includes captured inputs, model/cohort/missingness/geometry tables, PDF figures and replay records.

Criterion	Result (pass / fail / not-run)	Evidence and limitation
Numerical spot-checks	pass in checked scope	Five review groups plus operator counts, model arithmetic, Holm and dispersion calculations agree
Interpretation	pass in checked scope	No confirmed overclaim in inspected report; old descriptive defect absent
Delivered-copy consistency	pass	Exported artifact objects hash-verified; six registered scientific files match reviewed local delivery; report/code/rules/PNGs match bundle copies; four captured inputs match manifest hashes
Complete historical lineage	not-run / unable to verify	Original/replay execution history not fully linked or accessible; no stale content demonstrated
New full replay / fresh application import	not-run	Preparation verified retained evidence and bytes, not the full computation

Limitations and review context

This observational historical cohort is heavily US-dominated. Survey recall interval, self-report, identity-key validity, upstream processing and confounding limit generalization. Missingness tests do not establish missing-at-random or equivalence. Dispersion, multiplicity, absent effect-size confidence intervals and the substantial geometry correction remain disclosed.

No confirmed scientific defect was found in this bounded review. Underlying abundance-to-geometry calculations and original permutation generation were not independently rerun; complete supersession history remains unverified. The saved replay record lacks originating execution IDs.

Native review autoReview cmuzap9je000lx4zojrlqavk1 completed/pass with verificationUnavailable=true and legacy-unverified bindings. Its pass traces bounded findings to recorded outputs, not independent validation of the entire study. The review record is kept separate from the original session; no original scientific contents were edited during packaging.

Package-to-review correspondence

Item	Identifier
Original package	`21-2-2026-10-08.science`
Native attempt	`20261008-light-3`
Checking report	`case21-20261008-attempt3.md`
Checking report SHA-256	`069e79d4c34714b4bd0166cf07074be88052f129695716c409f619f417d4cf01`
Project	`cmuz9gqyu000fx4zor7gdf33r`
Original session	`01a11a94-934c-7b41-a60e-d2720d505494`
Review-only branch	`01a11aaf-927b-7f73-b6db-e833d0b64c92`
Review run	`99c95953-6235-4b53-8174-c065a813bb77`
Reviewed report version	`b967c3d0-46bc-4ed6-8e98-7c2ce9777608`
Reviewed bundle version	`7c7bfebf-4247-413d-a1dc-a026f739cb0a`

Full review and evidence remain outside this two-file submission. The preparation index records their paths and hashes.

Review

Reviewer: unassigned for repository acceptance.
Review date: pending; external bounded checking completed 2026-10-08.
Decision: pending.
Notes and next actions: no scientific correction recommended by the bounded review. Fresh import and complete historical lineage remain unverified. Prepared for HF submission on 2026-10-08; Feishu is not updated by this upload.