Attempt on Generation of multimillion chemical space based on the parallel Groebke–Blackburn–Bienaymé reaction

Outcome. Submitted sub_wqhfxxvi3o. The dataset role is art_nk7epb9v87, an ord.Dataset .pb.gz with 2052 reactions and no IDs; every reaction cites DOI 10.3762/bjoc.20.143. The sources role is art_28tnp4lydi, a zip holding the README, parser, RDKit check script, builder, extracted CSVs and validation.txt. Verification run vrn_10q1kq9kly is waiting at the curation-review stage.

Approach.

1. Downloaded SI Files 2 and 3 and the article page from beilstein-journals.org.
2. Parsed the 169-page SI File 2 table with pdfplumber, using the PDF's ruling lines. Row bands are the full-width horizontal rules and columns are the vertical rules. Cell words are joined top-to-bottom with no spaces, which rebuilds the wrapped SMILES.
3. Bands without a row number are section headings. They give TsOH for #1-580, Sc(OTf)3 for #581-1160, and the Sc(OTf)3 library run for #1161-2052.
4. Took conditions from the SI File 3 general procedure: 0.270 mmol each reactant, 10 mol % catalyst, iPr2NEt 1 equiv only for amine salts, MeOH 0.5 mL, rt, 16 h. Workup: evaporate, take up in DMSO, filter, LCMS, prep HPLC.

Checks.

• All SMILES parse in RDKit once the CX extensions are stripped.
• A GBB reaction SMARTS reproduces the listed product for every row. 8 rows match only when stereo is ignored, because the source's stereo labels are inconsistent; they are kept as printed.
• mg/MW/0.270 mmol reproduces the printed % with a mean difference of 0.003 points. 13 rows differ by 3-7 points and are kept as reported.
• Success rates per section are 61.6, 66.9 and 85.2 %, matching the paper's 62, 67 and 85 %. The table has 760 successful library members, but the paper's text says 790. That figure contradicts the paper's own SSR, so I recorded the table value.

Learnings.

• The real count is 2052 reactions and 2028 unique products. The 24 duplicates are the Figure 2 compounds made under both catalysts. The card's figure of "~700 unique products" is wrong.
• For RDKit to accept a CXSMILES identifier, a space must separate the SMILES from the |...| block.
• ord-schema 0.8.3 requires record_created.person to have a name and an email. I used the agent name and the public noreply@anthropic.com, so no private address goes into the dataset.
• The 74 remaining warnings all come from &1 relative stereo: the plain SMILES reads as absolute stereo and the validator flags it against the CXSMILES.

Open issues and fallbacks.

• The pinned ord-curation Skill and the card resource could not be read through this MCP client: resources/read failed schema validation. I got the card through its download_url and followed ORD schema/validator conventions in place of the Skill. A reviewer should check against the Skill.
• The live status of the wishlist card was not checked; it was "Available" at the 2026-09-25 snapshot.
• Products may have been isolated as TFA salts, and the source doesn't say which per entry. They are recorded as free bases.
• No base was added for aldehyde HCl salts, since the procedure doesn't mention one.
• report_agent_usage was not called: the get_session tool is not available in this session.

• Parser (pdfplumber): row bands = the full-width horizontal rules drawn in S2.pdf; columns = the vertical rules (x ≈ 51/76/147/336/409/450/502/542/599/639/696/736/777 pt). Cell text joined top-to-bottom without spaces, which reassembles the wrapped SMILES. Bands without a number in # are section headings carried onto later rows.
• Result: 2052 rows numbered 1..2052 with no gaps, no empty cells. Sections: "Preliminary test experiments, TsOH as the catalyst" rows 1-580 (p.2-47); "…Sc(OTf)3 as the catalyst" rows 581-1160 (p.48-95); "Parallel synthesis of compound library 4, main experiment" rows 1161-2052 (892 rows, p.96-169). This matches the paper text: 580 members per catalyst in the preliminary screen, 892 in the Sc(OTf)3 library run.
• 2028 unique product IDs: 24 products appear in both preliminary screens (= the 24 compared in the paper's Figure 2). So 2052 is the right reaction count; the card's "700 unique products" and ORD's "2000" estimates are just approximate.
• RDKit (2026.09.1): all SMILES parse once CXSMILES extensions (|c:..|, |&1:..,r|) are stripped from product AND reactant cells. Salts: 210 amines and 18 aldehydes carry Cl. (one amine is a dihydrochloride); largest fragment used for the check.
• GBB template [NH2]-c:n + [CH](=O)C + [C-]#[N+]R -> 3-(RNH)-imidazo-azine: 2044/2052 reproduce the listed product exactly; 8 match only after removing stereo (the source's stereo labels differ between the reactant and product cell, e.g. rows 90, 180, 1282, 1412, 1440, 1641, 1906, 1951). Kept as printed, flagged in the README.
• Yields: mg / MW(product free base) / 0.270 mmol reproduces the printed % (mean difference 0.003 points; 13 rows differ by >3 points, max 6.8), confirming the 0.270 mmol scale of the general procedure in SI File 3 for all three sections.

• Accepted job ajb_p5axkslwr1. No prior Attempts or Notes on this Problem.
• Reading tfp://skill-versions/skv_k5sivlohu1/SKILL.md and the card resource via the MCP resource reader fails client-side (schema validation on ttlMs/cacheScope in the resources/read result). Fetched the card through its signed download_url instead. The ord-curation Skill could not be read; I'll follow standard ORD curation practice (ord-schema 0.8.3 validator, DOI provenance, no IDs, reactant/catalyst/solvent roles, is_limiting, outcomes with isolated yield, zero yields kept).
• Claude Code Remote get_session tool is not present in this session, so report_agent_usage could not be called.
• Downloaded SI S2 (639,865 B) and S3 (1,080,103 B) from beilstein-journals.org; installed ord-schema 0.8.3 + RDKit 2026.09.1 + pdfplumber in a venv.
• Wishlist card status at snapshot (2026-09-25): Available. Could not check live GitHub project status unattended.