Nomenclature — will Rotamer name molecules?

August 27, 2026. A researcher survey covering four research tracks and seven attempts at refutation. This is not a decision: scope belongs to the pm. This records what was known on that date, with an address for every source.

Written when the product was closed-source (D-08). On September 11, 2026, D-28 opened the code under MIT and changed the licensing rule — what this survey says about commercial licensing and a closed-source product records that date, not today’s rule. The current assessment of the options is in docs/FORA-DE-ESCOPO.md, under “Nomenclatura — os cinco caminhos” (Nomenclature — the five paths).

The question

The teacher in the Phase 3 observation session will draw a structure and ask “what is this called?” — and today Rotamer does not answer, because D-15 chose to let users give molecules nicknames instead of naming them. The question is whether that decision still stands.

Short answer

Technically D-15 remains correct, but for the wrong reason. The premise “RDKit.js cannot do this” was checked three times and is true. But the implicit premise “there is no open, permissively licensed engine” became false on August 27, 2026: openclatura exists (MIT, deterministic, built on RDKit itself), and I measured it here naming 30 out of 30 high-school organic chemistry molecules, with a perfect round trip and textbook-style names. The obstacle is no longer licensing or quality — it is that it only speaks English, runs in Python on the server (never in the browser, never offline), and is beta 0.3.1, from a single lab.

What has been verified

1. What Rotamer has today — measured in the repository

The core has nothing close to nomenclature. @rotamer/core exposes 39 runtime symbols, and the /name|nomen|iupac|title/i filter returns nothing. The Molecule object has no name field: smiles, formula, inchi, inchiKey, molblock, descriptors, groups, atomHydrogens, stereo. — temporary probe in packages/core/test/, created, run, and removed.
The closest thing to naming is detectFunctionalGroups: 21 SMARTS patterns with pt-BR names (carboxylic acid, ester, amide, phenol, …). These identify functional classes, not compound names — they provide no parent chain, numbering, or branching. — packages/core/src/chemistry/groups.ts, lines 63-175.
Nickname validation rejects five categories, each with its own pt-BR message: curto, longo, caracteres, parece-formula, parece-sistematico (short, long, characters, looks like a formula, looks systematic). — apps/web/lib/molecule-name.ts, MESSAGES lines 43-50, checkName lines 52-66.
The parece-sistematico rule makes mistakes in both directions, and this was measured. It is a bare suffix (/(ano|eno|ino|ol|al|ona|oico|amina|amida|ato|ila|ilo)$/i) applied to a single word. Running checkName: it rejects Camila, Sol, Cristal, Girassol, Farol, Carnaval, Ludmila; it accepts cafeina, aspirina, anilina. In other words, it blocks a student’s first name and lets through the trivial name of a real compound — exactly the opposite of its purpose. It also rejects Ba and Na as parece-formula. — probe apps/web/lib/zz-probe.test.ts, created, run, and removed.
Search resolves only one direction. findByName sends the text to PubChem, and the returned SMILES passes through RDKit before becoming a molecule. findCompoundByInchiKey queries by key and, if found, gets the CID’s Title — which becomes the “already exists out there” screen, never a name field. — apps/web/app/actions/search.ts lines 42-85; apps/web/lib/pubchem.ts lines 154-234.
The product can already check name→structure without any engine. The condition { kind: 'inchiKey', value } exists in packages/quests/src/types.ts:42 and is evaluated in conditions.ts:40-41. And the key does not depend on how the student drew it: four different SMILES representations of aspirin (aromatic and Kekulé, different starting atoms) all returned BSYNRYMUTXBXSQ-UHFFFAOYSA-N. The name here is entered by a human in the mission data — the product never generates it, and D-15 remains intact.
Three states when PubChem goes down, not two: a 5 s timeout per request, a 6 s budget per question, a circuit breaker that opens after 3 rejections and stays closed for 60 s; knownCompound returns nao-sei (I don’t know), and the nickname remains valid, with the caveat on screen. — apps/web/lib/pubchem.ts lines 20-41; known-compound.ts lines 22-37.
The suite passes today: pnpm test → 21 files across 5 packages, Tasks: 5 successful, 5 total.

2. The two chemistry libraries already in the stack do not name compounds

RDKit.js 2025.03.4 (BSD-3): I enumerated the runtime, not the .d.ts. There are 21 module functions and 55 API methods on JSMol; the broad filter /name|iupac|nomen|systematic|preferred|synonym|label|title/i returns (none). get_descriptors() contains 43 keys, all numeric. get_prop_list() on a mol from SMILES returns {}. Grep for iupac in the .js and .wasm → 0, with a positive control: inchi in the same .wasm → 62. The method finds a string that exists, so zero means actual absence. nomencl, butane, and hexyl in the .wasm → 0, meaning no nomenclature lexicon is compiled in. There is no alternative build: dist/ contains only RDKit_minimal.{js,wasm}, and Code/MinimalLib/dist/ is an identical copy by md5.
OpenChemLib 9.25.0 (BSD-3): iupac → 0 in the bundle; no toName/fromName/StructureName in the .d.ts; resources.json (1.35 MB) contains only MMFF94, torsion, and toxpredictor.

Consequence: any nomenclature means a new dependency, not a flag to enable.

3. Name → structure: it exists, it is MIT, and it does not speak Portuguese

OPSIN 2.9.0, MIT. I read the project’s own LICENSE.txt at https://raw.githubusercontent.com/dan2097/opsin/master/LICENSE.txt — the full MIT text, “Copyright 2017 Daniel Lowe.” Release 2.9.0 was published on March 15, 2026 (api.github.com/repos/dan2097/opsin/releases); the previous release was from 2023. An active repository, pushed to on August 19, 2026, with 232 stars.
One direction only. The CLI’s own banner: “OPSIN converts systematic chemical names to CML, SMILES or InChI/StdInChI/StdInChIKey”. There is no structure→name mode.
Fast and correct. 15 names converted in 0.407 s wall time including JVM startup. It handles systematic and trivial names correctly (acetic acid and ethanoic acid produce the same SMILES) and returns stereochemistry ((2R)-butan-2-ol → C[C@H](https://github.com/GabrielMello1407/Rotamer/blob/main/docs/pesquisa/CC)O). Passing its SMILES through this repository’s RDKit reproduces the CLAUDE.md table: ethanol 46.07/20.23; acetic acid 60.05/37.30; aspirin 180.16/63.60; caffeine 194.19/61.82.
It does not understand Portuguese. A set of 42 high-school names: PT 3/42 (all three are spelling coincidences, such as propan-2-ol), EN 42/42. etanol, benzeno, acido acetico, 2-metilbutano, acido acetilsalicilico, propanona, and cafeina fail. Also confirmed on the EMBL-EBI service: curl https://www.ebi.ac.uk/opsin/ws/etanol.smi → HTTP 404.
Every failure is loud and visible. Across 62 names tested across the research tracks, there were zero silent mismatches — OPSIN never quietly returned the wrong molecule; it either got it right or rejected it. This matters under D-01: it is an error profile a classroom can tolerate.
A pt→en layer is manageable, but it is a lexicon and does not go away. A morphological normalizer of ~20 lines (ano/ane, ol/ol, ona/one, oico/oic) brought the tuned set to 41/42 and a held-out set of 20 names, with no new rules, to 14/20, with 0 silent mismatches in both. The 6 misses are dictionary entries: hidroxi→hydroxy, gli→gly, ureia→urea, estireno→styrene, and the ester “acetato de …”. Translating a compound name means deciding structure, so this layer cannot be an LLM (D-01).
A licensing caveat inside the jar. opsin-cli-2.9.0-jar-with-dependencies.jar (14,335,652 bytes, sha256 c2e29326c281f87b59a05d934d8589adac6e9d17b95b984931b3e739111b360f) bundles jna-inchi-core 1.3.1, whose POM on Maven Central declares LGPL 2.1 or later (77 io/github/dan2097/jnainchi/ entries inside the jar). LGPL is not in CLAUDE.md’s literal ban (GPL/AGPL) — precisely why it needs an explicit decision rather than an omission. opsin-core alone stays clean (automaton BSD-2, woodstox/commons-io/log4j Apache-2.0), but it is the artifact that stopped having a Main-Class as of 2.9.0.
There is no JS/WASM port. The npm package opsin is an unrelated Vue library; opsin-js, opsinjs, opsin-wasm, js-opsin, @opsin/core, node-opsin → all 404. GitHub search for opsin+wasm → 0; opsin+javascript → 1 result, a 2014 Safari extension that calls the service. Using OPSIN means running Java or calling EBI.
The Spanish fork exists and is unusable here. quimifyapp/opsin (14 stars) has no license file at all — derived from MIT with no declared grant.

4. Structure → name: the finding that changes the picture — `openclatura`

This was not in the four research tracks; a skeptic introduced it, and I measured it today.

openclatura 0.3.1, MIT, deterministic, built on RDKit. I read LICENSE at the source (https://raw.githubusercontent.com/lamalab-org/openclatura/main/LICENSE): “MIT License / Copyright (c) 2026 lamalab-org.” GitHub API: spdx_id: MIT, created on May 8, 2026, pushed to on August 24, 2026, 80 stars, Python. PyPI: openclatura 0.3.1, classifier “License :: OSI Approved :: MIT License.” Only runtime dependency: rdkit>=2023.09 (BSD-3) — I installed it in a disposable scratchpad venv, and the full tree was numpy, pillow, rdkit, openclatura. No GPL, no mandatory Java.
No model and no table. README: “There is no model and no lookup table: the same structure always yields the same name” and “Both are deterministic — same input, same output, no LLM in the loop”. This is the opposite of STOUT and makes it compatible with “the deterministic core decides.”

MEASURED HERE — 30 molecules, with textbook-style names. I ran oc.name_smiles() on the 6 CLAUDE.md cases plus 24 high-school organic chemistry molecules. Literal output included:

I drew	openclatura returned
`CCO`	`ethanol`
`CC(=O)O`	`acetic acid`
`c1ccccc1`	`benzene`
`CC(=O)Nc1ccc(O)cc1`	`N-(4-hydroxyphenyl)acetamide`
`CC(=O)Oc1ccccc1C(=O)O`	`2-(acetoxy)benzoic acid`
`Cn1cnc2c1c(=O)n(C)c(=O)n2C`	`1,3,7-trimethyl-3,7-dihydro-1H-purine-2,6-dione`
`CC(=O)OCC`	`ethyl acetate`
`CC(C)=O`	`propan-2-one`
`C/C=C/C`	`(2E)-but-2-ene`
`OCC(O)CO`	`propane-1,2,3-triol`
`CC(O)C(=O)O`	`2-hydroxypropanoic acid`
`C[C@@H](https://github.com/GabrielMello1407/Rotamer/blob/main/docs/pesquisa/N)C(=O)O`	`(2R)-2-aminopropanoic acid`
`CC(C)CC(C)(C)C`	`2,2,4-trimethylpentane`

Compare this with what the research track measured for NISPO 0.1.7, the other BSD-3 candidate: acetic acid → 1-hydroxy-1-oxo-ethane, ethyl acetate → 2-oxo-3-oxapentane, propanone → 2-oxopropane, cyclohexane → cyclohexan. NISPO is chemically correct but would fail an ester lesson; openclatura names an ester as an ester and a ketone as a ketone.

MEASURED HERE — round trip 30/30, zero silent mismatches. I took the 30 openclatura names, passed them through local OPSIN 2.9.0 (java -jar opsin-cli-2.9.0.jar -o smi), and compared InChIKeys using RDKit. Output: round-trip openclatura 0.3.1 -> OPSIN 2.9.0 -> InChIKey: 30/30 iguais (30/30 equal) · OPSIN nao entendeu o nome (falha alta): 0 (OPSIN did not understand the name, loud failure: 0) · DIVERGENCIA SILENCIOSA: 0 (silent mismatch: 0).
MEASURED HERE — it verifies itself. NamingResult has opsin_check and self_audit fields, and verify_with_opsin(nome, smiles) returns an OpsinCheck with a status. I installed the py2opsin extra, oc.opsin_available() → True, and ran it: ethyl acetate, caffeine, aspirin, and alanine → status=matched for all four. This is the mechanism that would allow showing only names confirmed by the round trip and staying silent for the rest — the design D-01 would require.
MEASURED HERE — the name comes out in pieces, not as a string. analyze_smiles() returns a NameAnalysis with trace_segments, decisions, substituent_tree, and operations. Each segment contains atoms, bonds, name_terms, and a rule_hint pointing to a Blue Book rule (P-44, P-45). For CC(=O)OCC: {'key':'parent', 'atoms':[0,1], 'name_terms':['acetate', 'acetat'], 'rule_hint':'Parent hydride / parent structure: Blue Book P-44 and P-45.'}. Two effects: (i) atom indices match the graph, so the drawing could highlight the part corresponding to each piece of the name — the same idea as “the highlighted atom is the same in both views”; (ii) localizing to pt-BR would mean mapping terms rather than translating a sentence — which reduces the problem but does not eliminate it, because the order changes (ethyl acetate → “acetato de etila”).
Two of the 30 names used a less common form than the textbook: 1-ethoxyethane (the locant seems superfluous beside “ethoxyethane”) and 2-(acetoxy)benzoic acid (the form cited in ENEM/PubChem is “2-acetyloxybenzoic acid”). I checked that OPSIN accepts all four spellings and they all yield the same molecule — ethoxyethane and 1-ethoxyethane → C(C)OCC; 2-(acetyloxy)benzoic acid and 2-(acetoxy)benzoic acid → the same aspirin SMILES. So the deviation is style, not chemistry. Whether that style would fail an exam is a question for a chemistry teacher, not for me — it remains open below.

5. Structure → name: the other candidates, and why they fail

NISPO 0.1.7, BSD-3 (Oxford Protein Informatics Group), Python on RDKit, published on PyPI on August 23, 2026. License verified from two sources. Round trip 26/26 in the research track that measured it — and its README warns: “It is optimized for OPSIN round-trip validity rather than preferred IUPAC nomenclature”. At least 8 of the 26 names would fail a lesson.
STOUT (MIT) is a neural network and fails the unbreakable rule. MIT LICENSE verified; PyPI STOUT-pypi 2.0.5. But the authors’ own paper (J. Cheminform. 2024, 16:146) measures 89.86% exact-match accuracy in the best experiment and 83.52% in large-scale training. This is literally the profile CLAUDE.md describes when banning the LLM. And it is unusable today: canonical repository 404, Zenodo weights 410 GONE (control: another Zenodo record responds 200), pip install STOUT-pypi fails with ResolutionImpossible, and stout.decimer.ai does not resolve.
nobyt/smiles2iupac is deterministic and active (pushed to on August 10, 2026), but license: null — all rights reserved, legally unusable.
JasonYCHuang/chem-dl-iupac is AGPL-3.0 and sneedkap/iupac-to-structure is GPL-3.0 — banned by CLAUDE.md.
No structure→name engine in JS or WASM with a permissive license. An npm search for iupac, smiles name, and opsin: nothing. indigo-ketcher 1.46.0 (Apache-2.0, WASM, the most obvious candidate): I downloaded the 10,146,858-byte bundle, and grep -c -i -a "iupac" → 0. ketcher-core 3.17.2 (Apache-2.0): also no. Architectural consequence: any nomenclature in Rotamer is a network call, not worker code.

6. The market — nomenclature is a paid, server-side feature

ChemDoodle Web Components is GPLv3, and more restrictive than claimed. I downloaded the official distribution (ChemDoodleWeb-11.0.0.zip, 4,641,214 bytes) and read COPYING.txt: it is the FSF’s GPLv3 text, 35,147 characters, without a single modified line. The source header says “either version 3 of the License, or (at your option) any later version” → GPL-3.0-or-later.
Its nomenclature does not run in the browser — this was executed, not read. In the distributed source, generateIUPACName and readIUPACName are one-line wrappers around _contactServer, which calls fetch(SERVER_URL, {method:'POST', credentials:'include'}). I loaded the library in Node with instrumented fetch, built ethanol, and requested its name: 1 network request, zero local nomenclature, and an errback when the network was blocked. The default SERVER_URL is https://ichemlabs.cloud.chemdoodle.com/icl_cdc_v090000/WebHQ; the /cdcloud.php on iupacnaming.com is that site’s override. A direct POST returned {"content":{"iupac":"Ethanol","attemptedPIN":"ethanol"}}.
ChemDoodle’s nomenclature is “powered by OPSIN” — its own demo page says so, and OPSIN is MIT. In other words, their GPL does not close off the name→structure direction.
ChemDoodle pricing: US$29 is a monthly subscription per user, not a purchase. The store lists Monthly $29, Yearly $199, Lifetime $999 for the “iChemLabs Suite User License,” with four products, and says “We do not license them separately”. The academic discount does not apply to those prices: “we offer a significant academic discount on our Site licenses… As for our user subscriptions or Lifetime user licenses, we are sorry”. A site license requires a quote, with no public price.
Ketcher (EPAM, Apache-2.0) does no nomenclature at all — it neither generates nor reads names. api.github.com/repos/epam/ketcher → Apache-2.0, pushed to on August 27, 2026.
MolView does not name compounds: it searches PubChem, RCSB PDB, COD, and NIST WebBook — what Rotamer already does. And it is a GPL stack: curl https://molview.org/ returns 12 occurrences of ChemDoodle, 15 of Jmol, and 12 of GLmol; the repository is NOASSERTION on GitHub. It cannot even serve as a code reference.
Market consolidation. Read on the manufacturer’s own website (https://www.acdlabs.com/resources/free-chemistry-software-apps/chemsketch-freeware/): “ChemSketch Freeware Discontinued. In January 2026, ACD/Labs was acquired by RS who develop ChemDraw”. Two of the five suppliers cited as independent are now the same company, and the free option that existed for schools has disappeared.

7. What Brazilian education requires — and in which direction

The high-school BNCC does not require nomenclature. I downloaded the official MEC PDF (http://basenacionalcomum.mec.gov.br/images/historico/BNCC_EnsinoMedio_embaixa_site_110518.pdf) and ran pdftotext -layout + grep -c -i: IUPAC 0, carbono 0, hidrocarboneto 0, isomeria 0, átomo 0; “nomenclatura” appears once, referring to educational terminology. The only mention of organic chemistry is “structure and properties of organic compounds,” in a list of knowledge that “may be mobilized” in Specific Competency 3 (pp. 543-544) — suggested context, not a skill requirement. There are exactly 23 skills (EM13CNT101-106, 201-207, 301-310), and none mentions nomenclature, structural formulas, or organic functional groups.
The BNCC does not assign a school year. The “13” in the code means “any year of high school, as defined by the curricula” (p. 33).
The state requires it, in the 3rd year, in both directions. SEDUC-SP’s “Habilidades Essenciais de Química – EM” (Essential Chemistry Skills — High School) document states, in 3rd year / 3rd two-month term: “Write structural formulas of hydrocarbons from their nomenclature and vice versa.” Only São Paulo’s document was read, in full.
ENEM hardly asks for it, and when it does, it is name → structure. Official second-day exams, yellow booklet 5, from 2020 to 2025, downloaded from INEP: 0 occurrences of “IUPAC” across all six years; “nomenclatura” once (2023). Not one question across those six years asks for the name of a drawn structure. Question 96 in 2023 lists five compounds by name only (n-decane, n-heptane, 2,2,4-trimethylpentane, …), without any drawn structures.
Unicamp asks students to draw, and students get it wrong. Annotated exams from 2005 to 2010: “structural formula” appears 18 times, and no question asks for a structure’s IUPAC name. In 2005, question 9 gives “3-penten-2-ol” and asks for a drawing. The examiners themselves wrote that “this question was not expected to present much difficulty” — and 41.59% scored zero.
Giving the answer before an attempt negates the feedback. Shute, V. J. (2008), Focus on Formative Feedback, Review of Educational Research 78(1), 153-189, Table 2 (pp. 177-178): “Provide feedback after learners have attempted a solution. Do not let learners see answers before trying to solve a problem on their own (i.e., presearch availability)”; and in the body: “feedback can inhibit learning if it encourages mindlessness, as when the answers are made available before learners begin their memory search”. A peer-reviewed journal, with no product involved.
And for difficult tasks, feedback should be immediate — “a helpful safety net for the learner so she does not get bogged down and frustrated” (Table 4, p. 179); for lower-achieving students, elaborated feedback works better than simple verification.
Brazilian literature has already diagnosed the problem. Matos et al. (2009), Nomenclatura de Compostos Orgânicos no Ensino Médio (Organic Compound Nomenclature in High School), Química Nova na Escola 31(1), pp. 40-46: “high-school students are led to memorize names, rules, and classifications, but because they do not use them often, they end up forgetting”, and the content’s basic purpose would be “to enable the connection between a substance’s name and its functional class” — the value lies in the name↔function connection, not in the name itself.

8. PubChem was down all day

Five independent sets of attempts, between ~09:44 and ~14:00 (local time) on August 27, 2026, all returning PUGREST.ServerBusy. My final attempt: curl -G ".../rest/pug/compound/smiles/property/IUPACName,Title/JSON" --data-urlencode "smiles=CCO" → {"Fault":{"Code":"PUGREST.ServerBusy"}} [HTTP 503], and the same for aspirin. Earlier, the regular HTML page (/compound/2244) also returned 503, and the headers included Retry-After: 30 and X-Throttling-Control: … too many requests per second or blacklisted. GitHub’s API responded 200 at the same moment, so this was not a network failure on our end.

This matters twice over: the product already depends on PubChem today (D-15’s existence check), and PUG REST’s IUPACName was a candidate naming source. I could measure neither its format nor its coverage. D-15’s third state (“could not verify”) is not a remote possibility — it was the state of the world all day today.

9. Debris found in the repository

git status at the root shows two untracked files: cd.html (96,535 bytes) and st.html (31,541 bytes), timestamped August 27, 2026, 09:55. The sizes exactly match the ChemDoodle and iChemLabs store pages downloaded during this research. They are collection debris that accidentally landed in the product root. I did not delete them — whoever cleans the tree before committing decides.

What the sources claim, and nobody checked

ChemDraw supports both directions, but only in the expensive edition. “Name-to-Structure and Structure-to-Name functions” is listed as a ChemDraw Professional feature; Prime does not include it. — the manufacturer’s page (https://perkinelmerinformatics.com/products/research/chemdraw/, an interested source) and Stanford’s IT page (https://software.stanford.edu/index.php/node/1873, independent of the manufacturer), read on August 27, 2026. The official edition-comparison PDF returns HTTP 403 for automated reading.
At Chemaxon (now under Certara), nomenclature is a separately licensed product, in both directions. “Name import is only available for a single molecule with the free MarvinSketch desktop application. For batch conversion … you need the ‘Name to Structure’ license.” And in Marvin JS: “a separate license for Name to Structure and/or Structure to Name is required”. — manufacturer documentation (docs.chemaxon.com), an interested source.
Marvin JS sends almost everything to the server. Listed Web Services include CIP Stereo Info, Elemental Analysis, Mol Export, Clean, Formats, Hydrogenizer, and similar services — nomenclature does not even appear in the default list. — manufacturer, an interested source.
Certara/ChemAxon Naming Toolkit is delivered through Java, Python, .NET, and a microservice; ACD/Name is desktop/server; Lexichem TK (OpenEye/Cadence) is a server-side C++/Python toolkit; ChemDraw JS is the only one that runs in the browser and supports both directions, delivered as a private npm .tgz plus an XML license. — all from the products’ own sales pages, interested sources. None publishes pricing. An aggregator (SourceForge) mentions US$475.00/year for ACD/Name, but it is a third party and describes a user license, not an embeddable SDK: do not trust that number.
ChemDoodle’s commercial license technically supports closed-source products: “Contact us for a proprietary license instead, which can be integrated and distributed with proprietary products.” — manufacturer, an interested source, with no public price.
PubChem’s IUPACName comes from OpenEye’s LexiChem — found in PubChem SDF field descriptions reproduced in third-party datasets and cited in the STOUT paper’s bibliography. I did not read primary PubChem documentation stating this. High confidence, primary source unread.
openclatura coverage: QM9 100%, PubChem 99.3%, ZINC22 97.4%. — the authors’ own README, an interested source, and this is round-trip coverage, not “a name a teacher accepts.” I measured 30 molecules; they measure millions. The two numbers answer different questions.
KingDraw is free and supports both directions, in mobile and desktop apps, not the web. — manufacturer (kingdraw.com) and a Vanderbilt library guide. I do not know whether naming happens on the device or whether it sends the structure to KingAgroot’s server, and I did not read the terms of use.
CheerpJ runs unmodified jars in WASM, but the Community Edition is only for personal and FOSS projects; commercial use costs £100 per developer/month on the Small Business plan. — Leaning Technologies, an interested source. I did not test it with the OPSIN jar.
EMBL-EBI has hosted OPSIN since May 21, 2025. The service responds HTTP 200 today (I checked), but the page states no rate limit, quota, or availability commitment — just generic “Terms of use.”

What the skeptics disproved or called into question

This section is not empty. Seven refutations were attempted; two claims fell, three became uncertain, and two held up under tougher testing.

REFUTED — “there is no open, permissively licensed structure→name engine.” This statement supported half the answer, and it is false. At least two exist: STOUT (MIT) and openclatura (MIT). The person who reported the finding had qualified it — “this is absence of evidence, not proof of absence; if one appears, this changes.” Two appeared. If D-15 is retained, this cannot be its justification, because someone will discover it is false. The defensible justification is different, and better: open engines are either neural, with 83-90% accuracy (STOUT), or single-lab beta software, and none speaks Portuguese or runs in the browser.

REFUTED — “ChemDoodle is banned in both cases.” CLAUDE.md bans GPL and AGPL; iChemLabs’ commercial license explicitly covers closed-source proprietary products. The paid route is not banned by policy — it is expensive, has unpublished pricing, and depends on a third-party server for each call. Calling it “banned” turns a business judgment into a rule, and someone reading this in six months will believe that door was closed.

REFUTED — “US$29 for the 2D+3D bundle” and “the testimonials are from schools.” It is US$29/month per named user, with no academic discount. And the testimonial about “cash-strapped state schools” is attributed in the raw HTML to “University of Missouri, Kansas City” — in American English, “state school” means a public university. Of the 37 attributed testimonials, 4 seem to be from high schools, 21 from universities, and 12 from companies. Furthermore, none of the 37 mentions naming, nomenclature, or IUPAC: connecting “sells nomenclature” to “sells to schools” was an inference presented as an observation.

UNCERTAIN — “OPSIN’s Portuguese support is the central obstacle.” The measurements are reproducible and became stronger (PT 3/42 versus EN 42/42), but one of the eight names failed with a different message from the one cited; “aspirin” suggested a class limitation that does not exist (caffeine and acetylsalicylic acid pass — it is a missing dictionary synonym); and above all, OPSIN works in the opposite direction from the question. It does not block “naming”; it blocks “entering by name,” which D-09 item 3 already places after the MVP. And the size of the obstacle was measured: ~20 lines of morphological rules recover 14/20 on a held-out set, with zero silent mismatches. It is a lexicon, not a wall.

UNCERTAIN — “OPSIN is MIT” (correct, and incomplete). The jar you download and run contains jna-inchi LGPL-2.1-or-later inside. Saying “OPSIN is MIT” and stopping there leads the team to install copyleft while believing it installed permissively licensed software. And the conclusion “it would mean a Java service alongside Next.js” is a false binary: EBI’s hosted service exists (it works today, and brings a different problem — sending what a student typed to a third-party server).

UNCERTAIN — “ChemDoodle nomenclature is a network call with a quota.” The mechanism was proved (POST, and the page intercepts a message containing exceeded), but the quota was not observed: all 10 identical calls succeeded. Mechanism proved, limit not proved.

CONFIRMED under tougher testing — “RDKit.js has no nomenclature.” It survived four attempts to disprove it, with a broader search than the original and the missing positive control (inchi → 62 in the same .wasm where iupac → 0). Two corrections remain: the package’s .d.ts is outdated and promises get_mol_from_pickle, which does not exist at runtime — grep on the .d.ts yields both false negatives and false positives, and “57 methods” cannot be reproduced under any counting method (there are 55 API methods). And the package’s LICENSE says BSD-3, but the .wasm statically links InChI, Boost, coordgen, FreeType, and Avalon; FreeType’s FTL requires attribution, and docs/TERCEIROS.md needs more than one line.

A warning none of the research tracks raised: RDKit.js’s condense_abbreviations produces labels such as CO2Et (executed: ethyl salicylate → CXSMILES *Oc1ccccc1* |$Et;;;;;;;;CO2Et$|). This is not nomenclature, but it is exactly the kind of string a student reads as a name. If those labels reach the screen, D-15’s rule needs to cover them.

The possible paths, and the cost of each

Ordered from cheapest to most expensive. None is a recommendation — scope belongs to the pm.

(a) Do not name compounds, and say so openly

Keep D-15 as it is and prepare a written answer to the teacher’s question instead of letting silence answer it.

Cost: almost zero engineering. A sentence on screen and a rewritten decision — because the current justification has been false since today (see the skeptics).
Risk: the teacher reads it as a limitation, not a choice. And there is a free competitor that names compounds for free on a student’s phone (KingDraw, unchecked). The risk grows if the answer is “it can’t be done”; it shrinks if the answer is “it can, but in English and with 1 in 15 names outside textbook conventions — we’d rather not confuse your class.”
What it commits us to maintaining forever: the discipline of rejecting nicknames that pass for nomenclature. And this rule has now been measured making mistakes in both directions — it rejects “Camila” and accepts “aspirina.” This is the cheapest path, and even it has outstanding debt.

(b) The student names it and the product checks

Name→structure: the student types a name, and the product converts it and compares it with what they drew, or with the mission target.

Cost: the lowest among paths that do something new, because half already exists: the missions’ inchiKey condition makes exactly this comparison today, without any engine, with a name entered by a human in the mission data. Free-form nomenclature (the student writes any name) requires OPSIN — a Java service on the VPS, or opsin-core without InChI, or EBI — plus the pt→en layer.
Risk: (1) OPSIN does not speak Portuguese, as measured; (2) the translation layer decides structure, so it cannot be an LLM (D-01) and becomes our own deterministic code, with a dictionary; (3) the full jar includes LGPL; (4) EBI promises no availability and would receive student text.
What it commits us to maintaining forever: a pt→en organic nomenclature dictionary, reviewed by someone qualified, plus another production process (JVM) or a third-party dependency.
In its favor: this is the direction ENEM and Unicamp test, where students make mistakes (41.59% zeros), and what Shute prescribes — the student produces, the software checks. It is also the only direction where engine errors are loud and visible: zero silent mismatches across 62 names.
Against it, seriously: D-09 item 3 already placed “search by name” after the MVP. This path reopens a scope decision that has already been made.

(c) Integrate a third-party engine for structure→name

Literally what the teacher asked for. The only realistic candidate: openclatura (MIT), running as a Python microservice alongside Next.js — the package already includes a [web] extra with FastAPI.

Cost: one more production service, forever. Plus pt-BR localization, which is the real work: NameAnalysis provides the name in segments with name_terms, so it involves term mapping and reordering, not sentence translation — but nobody has measured how much of that is rules and how much is exceptions.
Risk: (1) beta 0.3.1, one lab, created on May 8, 2026, 80 stars — if it stops, the problem is ours; MIT allows a fork, which limits but does not eliminate the damage; (2) English; (3) a network call for each name — it does not work offline and disappears when the VPS goes down, just as PubChem disappeared all day today; (4) a valid round trip does not imply a textbook name: 2 of the 30 used a less common form.
What it commits us to maintaining forever: the policy of showing only names that verification confirmed. openclatura already includes the mechanism (verify_with_opsin → status=matched, executed here), but enabling it pulls in OPSIN, along with Java and LGPL. Without verification, it is a beta engine speaking unchecked to an entire classroom.
In its favor, and this is the strong argument: it is deterministic, with no model or table; it carries a decision trace with the Blue Book rule number; and the atom indices for each segment match the graph — opening up a version of the product nobody else has: highlighting the part of the drawing corresponding to each piece of the name. That stops being “giving the answer” and becomes an explanation.

(d) Nomenclature only within a restricted range

Name only what can be guaranteed: open chains up to N carbons, a single functional group, no stereocenter — and stay silent elsewhere, explaining why.

Cost: lower risk than (c), greater ongoing attention. The range needs to be defined in tests, not in a comment.
Risk: (1) the boundary is invisible to users — students draw whatever they want, and a product that sometimes names compounds and sometimes does not seems broken unless the message explains the chemistry behind the refusal; (2) the range grows, and expanding scope is this project’s declared number-one risk: every lesson brings another case that “almost worked.”
What it commits us to maintaining forever: the range definition, a test that locks its boundary, and a pt-BR message explaining the silence.
An observation that changes the cost: with openclatura verifying itself, the range need not be handwritten — it can be “everything the round trip confirmed,” a measured boundary rather than an opinion. This combines (c) and (d), and is the cheapest combination that answers the teacher.

(e) Name only curated content, never live

It was not requested, but it is the cheapest path that produces some answer. Missions use a closed set of molecules. We could run openclatura once at build time on that list, have a teacher check each name, and store the result as data, just as names already enter mission data today.

Cost: none in production. No service, no network; it works offline and on low-end phones.
Risk: it only answers within a mission. A teacher who draws something outside the list still gets no answer — which is exactly what they will do in the observation session.
What it commits us to maintaining forever: a human-review step whenever the list changes. It is recurring teacher work, and disappears if nobody is paid to do it.
Note: this is not naming. It is a catalog. The difference must be stated on screen, or users will think the product names compounds — exactly what D-15 aimed to avoid.

Open questions

D-15’s literal trigger has not fired. It says “revisit if a reliable nomenclature engine in WebAssembly appears.” openclatura is Python. If the trigger applies as written, nothing has changed; if it meant “a reliable engine with a suitable license,” then things changed today. Whoever wrote the decision knows which one. — for the pm.
Which question the teacher actually asks. “Name what I drew” (structure→name, path c) or “correct the name my student wrote” (name→structure, path b)? These are different products, with very different costs. The Phase 3 observation session answers this for free, before choosing an engine. It is the cheapest item on this list.
How many openclatura names are wrong pedagogically rather than chemically. My 30 are a sample I selected; 2 used a less common form. To find out: assemble the list of molecules the tracks actually use (packages/quests), run openclatura on it, and have a chemistry teacher mark each name accepted/rejected. Without that number, nobody knows whether it is “almost there” or “not suitable.”
How much pt-BR localization of openclatura costs. Names come out in segments with name_terms, suggesting mapping rather than translation. Measure by counting distinct terms generated when naming the mission list, and cases requiring reordering (“ethyl acetate” → “acetato de etila”). I did not do this.
PubChem’s IUPACName remains unmeasured: 503 all day, across five sets of attempts. Retry at another time and from another IP. If it names arbitrary SMILES, there is a free, deterministic structure→name path with nothing to install — but PubChem only names what already exists there, the opposite of the nickname use case.
Was PubChem down, or was this IP blocked? If it is a source-based block, the production VPS could meet the same fate and nobody would know until a lesson happened. Find out by repeating the same curl from another network. — for backend and deploy.
The actual false-positive rate of SYSTEMATIC. I invented my own test set. Find out by passing a list of Brazilian first names (IBGE) and a pt-BR vocabulary through checkName — it runs offline, in minutes.
Latency under classroom load. A single call says nothing about 30 students at once. And testing a free public service’s limit by taking it down is something we must not do — which is itself an argument against depending on one.
Fuvest and UNESP were not checked (TLS error in the archive). The ENEM+Unicamp pattern is strong, but another exam board would make the case stronger. Curricula from Goiás, Minas, Bahia, and Paraná were not read either — if any requires structure→name under an EM13CNT code, that reopens the discussion.
A study directly comparing “software gives the name” with “student gives it and software checks” in organic nomenclature was not found. Shute’s argument is general, not chemistry specific. Find out by instrumenting the product itself.

Who needs to know

pm — makes the decision, and there are two decisions, not one. First: D-15’s justification fell (“there is no open engine” has been false since today), even if the decision stands; if it does, it needs a new argument, and the new one is better than the old. Second: path (b) reopens D-09 item 3, which already placed name search after the MVP. Before any choice, the cheapest question on the list is the observation-session question: which of the two directions the teacher wants.
security — before any installation, and there are four issues. (1) opsin-cli-2.9.0-jar-with-dependencies.jar contains jna-inchi LGPL-2.1-or-later; it is not in CLAUDE.md’s literal ban, which is precisely why it needs an explicit decision. (2) openclatura 0.3.1 is MIT with only one runtime dependency, rdkit (BSD-3) — clean, but it is the product’s first Python dependency and changes deployment. (3) STOUT is the general lesson: an MIT LICENSE with weights distilled from proprietary software under an academic license — the code license does not cleanse training-data provenance, and this applies to any model the team considers. (4) quimifyapp/opsin and nobyt/smiles2iupac have no license at all. Add to this: any nomenclature engine is a network call, carrying a student’s drawing, often a minor’s, to a third-party server — the legal basis under LGPD must be defined before, not after.
reviewer — two ready vetoes. If anyone proposes STOUT or any neural nomenclature engine, the figures 83.52% and 89.86%, measured by the authors themselves, are the argument, and match the profile CLAUDE.md uses to ban the LLM. And if any nomenclature path is adopted, the accompanying rule is do not show a name that verification did not confirm.
backend and deploy — PubChem returned 503 all day on August 27, 2026, and the product already depends on it. It is worth finding out whether it was the service or the IP, and measuring from the VPS’s IP, which is the one that matters.
frontend and ui-ux — the parece-sistematico rule rejects “Camila” and accepts “aspirina,” the opposite of its purpose. And RDKit’s condense_abbreviations produces labels such as CO2Et, which students read as names — if they appear on screen, they fall under the same D-15 rule.
Whoever cleans the tree before the next commit: cd.html and st.html at the root are untracked debris from this research.