CURATION.md

Curation and provenance

This is a dated public projection of the ChemBricks knowledge corpus. The creator, ChemBricks, requested the MIT license for this release. Source records are read without editing the source corpus, validation reports, review receipts, ledger, queue, or worker instructions.

What qualifies

Each source record must be current, have a completed ledger reference, and have unchanged record, trace and tool-log hashes. The release builder repeats the existing pure G1-G8 checks under the source validation policy and requires their claim results to agree with the stored report. It verifies applicable G9 replay evidence and G10 referee integrity. A record must have a passing independent model referee or a recorded building-agent trace review bound to the exact trace hash.

The word curated describes this computational review process. Building-agent reviews are AI reviews; they are not presented as independent human chemist annotations.

The existing source-confidence projection is applied transitively. A claim is eligible only when it and its arithmetic dependencies have passed, none are quarantined, evidence is linked, no dependency issue is found, and prediction timing has not failed. Low-trust, screening-grade, model-limit and unverified-timing qualifications remain attached.

The records subset contains complete, unchanged question/answer/trace text only when every structured claim is eligible and the record passes G6. Full prose is never made clean by deleting only its structured disputed claims.

The claims subset additionally contains eligible claims from mixed-result investigations. Their original record question and Limitations section supply context; no full answer from an ineligible record is presented as a clean answer. These claims keep their own statement, method, conditions, checks, source confidence and evidence. A failed design target is not converted to a passed finding.

The evidence subset contains every non-protocol tool-call row from the retained source captures, including intermediate calculations and recorded failures. These are raw observations, not an additional set of validated claims. Identical capture hashes are stored once; multiple questions or revisions sharing a capture are not independent replications.

Public-copy transformations

• Chemical arguments, responses, numeric values, response request IDs, model metadata and claim text are preserved.
• The release omits MCP protocol advertisements, private endpoint configuration, local runtime controls and worker configuration.
• Request-envelope client metadata is removed because it may contain local workspace paths or session metadata. Response metadata remains intact. Original tool-row hashes are retained separately from newly computed public-row hashes.
• Source-record hashes, full source-tool-log hashes, source-validation hashes, ledger commitments, curation decisions and source-code hashes are included. A source-tool-log hash identifies the original capture; the filtered public evidence file is not claimed to have that same file hash.
• Variable-shape scientific objects are stored losslessly in columns ending in _json. This avoids incompatible JSON/Arrow types in the Hugging Face viewer. Decode these columns with json.loads.
• Serialization uses UTF-8 and JSON escaping. Decoding restores the original text, including exact tool-response string values.
• The owner-authorized release license is mit. Historical source license labels remain in the provenance index as historical metadata; the license of this public release is given by LICENSE and the dataset card.
• Secret and private-host scans check the material selected for publication. Records with a remaining match are held out rather than having scientific text silently rewritten.

The release manifest states exact counts and exclusions. No new ChemBricks calculations or model-generated chemistry answers are produced by packaging.

Evidence versus experimental truth

Passed validation means that a claim meets the recorded provenance and consistency requirements. It does not establish experimental accuracy, causal mechanism, a general law of chemical space, or successful synthesis.

The dataset preserves the distinction between endpoint predictions and experimental measurements, logP and solubility, computed cross-route discrepancies and calibrated error bars, and prospective predictions and retrospective explanations. A calculation's returned trust factor is not an experimentally measured confidence probability.

G9 is sampled, not universal. Phase 0 records may have historically skipped G10. Individual gate states and referee scores are available on each row. The exact source policy versions and hashes are recorded; the release does not apply a new chemistry tolerance.

Splits and duplicate observations

This release has one split named corpus. At packaging time the source repository has no frozen scaffold holdout, so this is not a production training split or an independent evaluation benchmark. No random row split is supplied.

Before an evaluation, keep source captures, revisions, question paraphrases and linked comparisons together, then define a documented scaffold-based holdout from the captured structures. The molecule lists and source capture IDs help with this, but molecules appearing only in returned candidate lists must also be considered. A held-out result must not reappear through a derived question or the claims/evidence subsets.

Rebuilding

From the source repository:

python -B huggingface/build_dataset.py
python -B huggingface/write_card.py
python -B huggingface/verify.py --seal
python -B huggingface/verify.py

Rebuilding takes a new snapshot. It is not a promise to recreate the same release after new records or supersessions arrive. The ledger head, source-code hashes and snapshot time identify the source state. Release verification is offline and makes no chemistry calls.