This repository contains the public research data associated with Governed agentic discovery for auditable sustainability assessment. Release 2.1 adds source-linked reference records and de-identified outcome-level tables to the machine-readable source data for main-text Figures 1–5. The release remains data-only and retains the 11 scientific cases from release 1.0.
| Domain | Cases | Location |
|---|---|---|
| Industrial LCA | 6 | data/industrial/ |
| TDF inventory reconstruction | 1 | data/tdf/ |
| Electronics LCA | 4 | data/electronics/ |
Each case contains the applicable scientific scope, source metadata and public
evidence resources. The TDF directory separates the primary-source metadata
used for de-novo generation from the published-inventory packet used only for
the post-hoc reference audit. data_index.json lists all cases and packets.
Third-party binaries are omitted unless redistribution is permitted; the
manifests retain identifiers and hashes for authorized retrieval.
reference_records/ contains frozen, machine-readable LCI, LCIA, goal-and-scope
and evaluation-reference records. record_set distinguishes the public source
packets from the six blinded industrial configurations used in Figure 3; their
functional units and boundaries are different and the values are not
interchangeable. Held-out evaluation references were unavailable during model
inference and are released here only as static post-campaign data.
The export deliberately omits metric weights, tolerances, scoring logic,
dependency specifications, governance state and assessment histories. See
DATA_DICTIONARY.md and reference_records/README.md.
The sole figure-data freeze is dated 29 August 2026. Machine-readable sources
are in figure_data/; file sizes and SHA-256 hashes are recorded in
figure_data/index.json. Rendered figures and other presentation assets are
intentionally excluded.
Figure 4 received a bookkeeping correction on 30 August 2026: reported confidence for each model-based condition now uses the same 150 corrupted artifacts and 15 clean controls as external correctness. Detection, exact repair, preservation, residual defects, clean acceptance and premature export were unchanged.
Run-level Figure 1 scores and de-identified Figure 4 artifact and defect outcomes are included so that the reported descriptive statistics can be independently reaggregated without access to implementation or validation logic.
This remains a public data release. It does not contain the CogniLCA implementation, validation or scoring logic, metric weights or tolerances, dependency manifests, governance-state records, model credentials, runtime environments or full transcripts.
Author-created tables, metadata and figure-source data are licensed under CC BY
4.0 unless a file states otherwise. Third-party material remains subject to its
original terms; see DATA_LICENSES.md. Citation metadata are provided in
CITATION.cff.