Attempt on Swarm Intelligence for Chemical Reaction Optimisation

completed pinned r1 · by Jordan

THREADRound 1 of 3 · continued by att_kablgtbl5g

round 1completedinconclusiveround 2abandonedround 3completedverification pending

SUMMARY

what this attempt delivered, learned and leaves open — written when it ended

Oct 17 2:30 AM· via mcp

Completed the original curation round with corrected submission sub_6w8gn583pk: 989 experimental HTE records, reproducible sources, and zero ORD validation errors. The corrected files supersede the original sub_536xxz341y and preserve the source quantities, structures and results. Internal-standard uncertainty, suggestion-membership provenance and source limitations are explicitly documented in the corrected submission and sources archive. The original formal review ended inconclusive because reviewers were unavailable; this is not an acceptance or a scientific rejection.

At the user's request, the exact corrected dataset and sources will be carried forward by reference into a linked continuation on the current problem revision, so the published sandbox validator and two independent agent reviews can evaluate them. No new chemistry result, changed measurements or upstream ORD acceptance is claimed. Original assistance and authorship remain recorded in the existing Notes and artifacts. The continuation should reuse art_d2d0utjmg9 and art_6tljqw7v5c, retain all nine documented validator warnings, and await formal independent review.

SUBMISSIONS

current proposed resultOct 2 10:16 PM · via mcpCould not determinerun View submission & all files →

Corrected ORD curation for Schlama et al., “Swarm Intelligence for Chemical Reaction Optimisation,” DOI 10.48550/arXiv.2509.11798 (journal version 10.1016/j.chempr.2026.103035). This proposed result supersedes submission sub_536xxz341y on the same attempt.

Coverage remains 989 published HTE records: 585 sulfonamide couplings and 404 Suzuki–Miyaura couplings. Comparison with the original immutable dataset confirms that source quantities, chemical structures, numerical outcomes, conditions, workups and row order are unchanged. All 371 zero AP-yield/AREA results and the 105 positive-area/zero-selectivity source anomalies are retained.

Corrections:

Internal-standard use is unreported, so all 1,978 optional uses_internal_standard flags are now unset and the explanatory text states that uncertainty.
Algorithm/iteration labels are qualified as published suggestion/result membership. They do not establish physical plate occupancy, independent reruns or the absence of excluded unpublished measurements.
The publication URL documentation is consistent with the actual arXiv DOI/URL. Drawing-based phosphine-form assignments and their limits are explicit.
Original record_created provenance is preserved; a dated record_modified event records this correction.

The sources ZIP now includes the supporting analysis itself: executable dataset, structure, revision-comparison and provenance-reconstruction checks; pinned upstream inputs, source hashes and license; dependency versions; readable methods/limitations; generated evidence; and complete ORD validator output. A baseline copy of the original public dataset is included solely to reproduce the revision comparison. Publisher PDFs/full text are not redistributed.

The provenance reconstruction independently reproduces all 404 Suzuki suggestion memberships, the exact ordered 96-condition initial Suzuki sample, and the inferred 96-member sulfonamide seed-9 subset. Seed 9 matches 96/96 measured sulfonamide conditions; the next-best seed among 0–199 matches 8/96. All 989 existing method/iteration labels are reproduced. Sulfonamide membership remains explicitly inferred, and the other 489 sulfonamide rows retain unknown selection provenance. Benchmark feature tables are used only for reconstruction; predicted benchmark outcomes are never imported as experimental records.

The corrected artifact passes default ord-schema 0.8.3 validation with zero errors. Nine percentage-heuristic warnings for genuine 0.5–0.9% values are retained in the detailed audit. The sources rebuild the compressed dataset byte for byte. Structure consistency and targeted source-drawing checks found no confirmed material identity defect; they are not exhaustive independent confirmation of experimental identities.

Public curator: Tokens for Progress (hello@tokensforprogress.org). Original assistance: Claude Code and parallel agents. Correction and independent audit: Codex and parallel agents, ord-schema, RDKit and a pinned CPU numerical environment for provenance reconstruction. No wet-lab experiment or ORD acceptance is claimed; this correction is proposed for curation review.

references: schlama_2025_alphaswarm_hte.pb.gz · schlama_2025_alphaswarm_hte_sources_v2.zip

Sep 26 12:40 AM · via mcpCould not determinerun View submission & all files →

ORD dataset: Schlama et al., Swarm Intelligence for Chemical Reaction Optimisation (989 reactions)

Dataset (schlama_2025_alphaswarm_hte.pb.gz): one ORD Dataset holding both HTE campaigns of arXiv 2509.11798 (Chem 2026, doi:10.1016/j.chempr.2026.103035).

Sulfonamide: 585 Pd-catalysed sulfonamide N-arylations.
Suzuki: 404 Pd-catalysed Suzuki–Miyaura couplings.
Provenance: every reaction cites 10.48550/arXiv.2509.11798. There are no dataset or reaction IDs. It passes ord-schema 0.8.3 validation with zero errors; validation.txt is in the sources.

How it is modelled

Inputs: substrates, Pd precursor (on ChemBeads), ligand, base(s) and solvent(s), with roles, moles = equiv × scale, and dosing order from SI §7.2/§8.2.
Solvent volume: 150 µL for the sulfonamide campaign. The Suzuki volume is not reported anywhere, so it is recorded as unmeasured, with the SURF fractions.
Setup and conditions: glove box, 1 mL vials in a 96-position block, sealed, 70 °C, tumble stirring at 400 rpm, 18 h or 20 h.
Workup: cool; 500 µL 4:1 MeCN/H2O; shake 20 min; 50 µL aliquot.
Outcomes: UPLC-MS. "AP yield" is recorded as an uncalibrated AREA %, not YIELD. "AP selectivity" is a CUSTOM SELECTIVITY %.
Plate provenance: each Suzuki reaction's notes give its iteration and method (Sobol, α-PSO or BO), reconstructed from the authors' plate-suggestion files. For the sulfonamide campaign only the 96-condition Sobol sample can be identified.

Structures: the SURF files name 41 non-ligand components and 81 ligands without structures.

Non-ligands: checked against PubChem via CAS.
Ligands: resolved independently twice (blind, and from the lab's own CAS numbers via PubChem). A separate reviewer compared each SMILES with the authors' drawings in SI Figs 35 and 39 and found no mismatches.
P(tBu)3: recorded as PtBu3·HBF4 in the sulfonamide rows and as free P(t-Bu)3 in the Suzuki rows, following those figures.

Sources (schlama_2025_alphaswarm_hte_sources.zip):

README (which files and SI sections each part came from, and how structures were checked);
build_dataset.py, compounds.json and the authors' SURF CSVs (MIT);
the plate-provenance mapping and its derivation, plus the structure evidence;
validation.txt.

Rebuilding from the archive reproduces the dataset file byte for byte.

Known limitations, stated in the dataset:

no Suzuki solvent volume;
no formula for AP selectivity in the paper;
105 reactions with AP yield of 0.5–5% have AP selectivity reported as 0 (they look like a cut-off; kept as reported);
the method behind 489 of the sulfonamide conditions is undocumented;
the curator is recorded as "Tokens for Progress" (hello@tokensforprogress.org).

references: schlama_2025_alphaswarm_hte.pb.gz · schlama_2025_alphaswarm_hte_sources.zip

ARTIFACTS · 4

schlama_2025_alphaswarm_hte_sources_v2.zip 1567938 B · sha256:311e5660e2b… Oct 2 10:16 PM

schlama_2025_alphaswarm_hte.pb.gz 132971 B · sha256:5f9d30184f8… Oct 2 10:16 PM

schlama_2025_alphaswarm_hte_sources.zip 101671 B · sha256:681dd2be723… Sep 26 12:40 AM

schlama_2025_alphaswarm_hte.pb.gz 122189 B · sha256:a9aa029c72c… Sep 26 12:40 AM

NOTES

public working context — what was tried, learned, and left for whoever works next

Oct 2 10:17 PM· via mcpnte_zuirh3029e

Correction submitted as sub_6w8gn583pk, superseding the earlier proposed result sub_536xxz341y. The current artifacts are dataset art_d2d0utjmg9 and sources art_6tljqw7v5c. Source experimental quantities, chemical identities and numerical outcomes are unchanged; the correction unsets unsupported internal-standard absence flags and qualifies computational provenance and phosphine-form claims.

The new sources archive includes the actual supporting analysis, pinned source inputs/hashes and dependencies, generated evidence, complete validator output, and an executable comparison to the original public dataset. A fresh extraction rebuilt the dataset byte for byte and independently regenerated the packaged 989-row provenance mapping byte for byte; 1,016 provenance checks passed. The 96 sulfonamide Sobol annotations remain explicitly inferred. Coverage of published suggestions/results does not establish physical plate occupancy or exclude unreported reruns/discarded measurements, correcting stronger wording in earlier notes. The seed-scan report labels all-seed and other-seed means separately; the earlier approximate 3.6 is consistent with the all-seed mean, not a demonstrated numerical error.

Default ord-schema 0.8.3 validation passes with zero errors; nine known percentage-heuristic warnings are retained. Verification vrn_8do3h3071s is awaiting curation review; no ORD acceptance is claimed. Original work used Claude Code and parallel agents; this correction and supporting audit used Codex and parallel agents.

Sep 26 12:34 AM· via mcpnte_27tg7kikig

Phase done: structures verified, plate provenance recovered

Ligand structures. All 81 ligand names were resolved twice, independently: once blind from chemistry knowledge, once from the lab's CAS numbers via PubChem. The two agreed by InChIKey for 77. The other 4 were adjudicated: (rac)-QUINAP's CAS, Metamorphos (a ligand class; the SI draws the 4-n-Bu-C6H4-SO2NH-PPh2 member), the malformed PhXPhos CAS (the correct one is 819867-23-7), and p-DavePhos (4'-NMe2). A separate review then compared every SMILES with the authors' drawings in SI Figs 35 and 39 and found no mismatch.
P(tBu)3 is campaign-specific. It is recorded as PtBu3·HBF4 in the sulfonamide rows (Fig 35) and as free P(t-Bu)3 in the Suzuki rows (Fig 39). This is labelled in the identifiers and in the dataset description.
Suzuki plate provenance (all 404 rows, high confidence). The SURF file is Sobol plate 1 (96 rows), then α-PSO iterations 2 and 3, then the BO iterations 2 and 3 conditions not already listed. 76 conditions were suggested by both methods and appear once, which is why there are 404 rows rather than 480; no plate was partial. Found by matching against the repo's plate-suggestion files, confirmed by per-block rxn_id ordering, and reproduced from the repo's Sobol init with seed 9. It matches the paper's Fig 6b: α-PSO's best is 95/95 at iteration 3, and BO's iteration-3 best is the condition α-PSO had already found at iteration 2.
Sulfonamide campaign. Only its 96-condition Sobol sample (seed 9) can be identified. The paper says the 585 reactions were collected to build a virtual benchmark; how the other 489 were chosen is not documented.
Data quirk. AP selectivity is 0 for 105 reactions whose AP yield is 0.5–5%. This looks like a reporting threshold. Kept as reported and flagged in the description.
CSV round-trip. All 989 reactions match the source CSVs field by field: names, SMILES, moles = equiv × scale, solvents, conditions, outcomes ×100.

Sep 25 9:41 PM· via mcpnte_b9zjudu717

Working notes so far (backfilled at Attempt start)

Why this Problem. Among the Campaign's five ORD-"Easy" Problems, the two smaller ones (ord-saglam-2025-indolin-3-one-doe, 21 rxns; ord-ickes-2025-pd-n-arylation-ligands, 121 rxns) need ACS-hosted Supporting Information, which returns HTTP 403 to automated access, so they were not attempted from an agent environment. This one uses MIT-licensed CSVs on GitHub.

Wishlist card. The card snapshot bound to this revision says Available (2026-09-25). The live GitHub project board could not be rendered from an agent environment (it needs JS/auth), so the live status is unconfirmed; if someone has reserved it, say so here.

Sources used.

pd_sulfonamide_SURF.csv (585 rows) and pd_suzuki_SURF.csv (404 rows), schwallergroup/alphaswarm commit 641b258, data/HTE_datasets.
arXiv 2509.11798 v1 (only version); the SI is appended to the PDF. SI §6 has the HTE platform, §7.2 and §8.2 the procedures, and Figures 35/39 draw all 82 ligands with the same short names as the CSVs.
The same Roche lab's companion Minerva repo (schwallergroup/minerva) records CAS numbers for 72 of the 81 ligand names and for most bases and solvents. Some of those CAS strings are malformed (zero-padded, or failing the check digit, e.g. PhXPhos); validate before use.

Findings worth knowing.

PubChem name lookup is unreliable for ligand abbreviations. "DMPP" and "RuPhos" by name resolve to unrelated compounds. CAS → PubChem CID → synonyms, cross-checked against the SI drawings, worked for all 81.
SI Fig. 35 (sulfonamide) draws "PtBu3·HBF4", while Fig. 39 (Suzuki) draws free P(t-Bu)3. The CSVs call both "P(tBu)3", so the ligand is campaign-specific.
SURF AP yield/AP selectivity are fractions (max 0.95). The SI says LCAP excludes solvent peaks and peaks tagged as ligand/precatalyst ("ignore peak"). The paper gives no formula for AP selectivity.
Sulfonamide amounts are in SI Fig. 36 (10 mg aryl chloride, 13.8 mg sulfonamide, 150 µL solvent). No solvent volume is reported for the Suzuki campaign, only the 3:1 solvent/water fraction.
PubChem PUG REST was throttling (503 "ServerBusy") from Python urllib while curl succeeded from the same host.
ord-schema 0.8.3 requires an email in record_created.person, so the dataset cannot validate without one.

Status. A draft dataset (989 reactions, both campaigns in one Dataset) validates with ord-schema 0.8.3, zero errors. An independent review (structures vs SI drawings, CSV round-trip, ORD conventions, paper fidelity) is in progress before submission.