chemistry-claim-check

name	chemistry-claim-check
description	Checks molecular claims stated in prose — compound name against formula, degree of unsaturation, recomputed molecular weight, and the assessable part of the Lipinski rule of five — and returns a PASS/FAIL verdict per item with the arithmetic shown. Use when a paper, patent, or product page pairs a compound with a formula or a weight, on requests like "does this molecular weight check out?" or "is C2H6O2 really ethanol?". Not for validating a SMILES string — use `smiles-sanity-check` instead.
license	MIT
metadata	{"category":"domain","method":"Formula arithmetic (DBE/RDBE, molecular weight) plus Lipinski's rule of five","origin":"Lipinski, Lombardo, Dominy & Feeney, 1997; DBE from classical mass-spectrometry practice","version":"2.0.0"}

Chemistry Claim Check

A compound name, a molecular formula and a molecular weight are three statements about one molecule, and they must agree. When they do not — "ethanol (C2H6O2)", "caffeine … MW 312 g/mol" — the text has copied the wrong compound, mixed a free base with a salt, or invented the number. This skill recomputes the arithmetic each claim implies and reports where the claim contradicts itself. It is heuristic sanity-checking, not computational chemistry: it can prove a claim inconsistent, never prove a molecule real, stable or active.

When to invoke

Invoke when:

• Prose pairs a compound with a formula or a weight: "aspirin (C9H8O4)", "the API, C21H23NO5", "MW 194.19", "molar mass of ~500 Da".
• A patent or press release asserts drug-likeness: "orally bioavailable small molecule", "Lipinski-compliant lead compound".
• A structured record carries formula: or molecular_weight: fields extracted from text rather than a chemical database.

Do NOT invoke when:

• The input is a SMILES string needing syntactic validation — use smiles-sanity-check: this skill checks the prose, that one the string.
• The question is whether the compound data is true rather than internally consistent — re-source it with grounded-fact-check, and grade the publisher with rate-source-admiralty.
• Real chemistry semantics are required (tautomers, computed logP, stereochemistry, synthesisability). Those need a cheminformatics toolkit; say so rather than approximating.

Procedure

1 — Extract every molecular claim

Pull the compound name, the formula, the stated weight (with units, and whether it is average, monoisotopic or exact), and any drug-likeness language. Keep each claim attached to its sentence: a mismatch is diagnostic of which sentence was copied from elsewhere.

2 — Check name against formula

For well-known compounds the formula is a lookup, not a judgment: water H2O, ethanol C2H6O, glucose C6H12O6, aspirin C9H8O4, caffeine C8H10N4O2, ibuprofen C13H18O2, paracetamol C8H9NO2, penicillin G C16H18N2O4S, methane CH4, CO2, ammonia NH3. "Ethanol (C2H6O2)" fails: that is ethylene glycol, and a swapped formula usually means the rest of the claim came from the wrong compound too. For novel candidates, do not guess — mark it unverified.

3 — Test the formula for possibility

Compute the degree of unsaturation, RDBE = C + Si + 1 + (N + P)/2 − (H + halogens)/2. It must be a non-negative integer for a neutral molecule: benzene C6H6 gives 4, ethylamine C2H7N gives 0, and C2H8O gives −1, which nothing neutral can be. A negative or fractional value means the formula is corrupt — stop rather than checking anything downstream. Cross-check with the nitrogen rule: for a neutral C/H/N/O/S/halogen molecule, an odd nominal mass implies an odd nitrogen count.

4 — Recompute the molecular weight

Sum the IUPAC standard atomic weights (H 1.008, C 12.011, N 14.007, O 15.999, S 32.06, P 30.974, F 18.998, Cl 35.45, Br 79.904, I 126.904) and compare with the stated value. The tolerance is whichever is larger: 0.05 % or half the last written digit — so a weight quoted as "312" tolerates ±0.5. A few per cent off suggests a salt or hydrate mix-up; more than 10 % off means a different compound. Never reconcile a mismatch by adjusting the formula.

5 — Apply the assessable part of Lipinski's rule of five

Lipinski's rule of five (Lipinski et al., 1997) flags poor oral absorption when two or more of these fail: molecular weight ≤ 500, logP ≤ 5, hydrogen-bond donors ≤ 5, acceptors ≤ 10. From a formula alone only molecular weight and a crude acceptor ceiling (N + O) are computable; donors and logP need a structure, so mark them not assessable. Violations lower confidence in an oral-bioavailability claim, never disprove it: macrocycles and degraders legitimately live beyond the rule of five.

6 — Emit the verdict and route failures

Report each check with its arithmetic. On any FAIL, hand off to abstain-or-escalate: re-source from primary literature, drop the chemistry specifics and keep what survives, or escalate when the chemistry is load-bearing.

Output template

Every check listed in the template is mandatory, and each must show the numbers it was decided on; a bare verdict without arithmetic is not a result.

Compound claim: {name} — {formula} — "{stated MW}" — "{drug-likeness language}"
{PASS|FAIL|UNVERIFIED}  name vs formula     {canonical formula, or why it cannot be verified}
{PASS|FAIL}             formula possibility  RDBE = {arithmetic} ({integer? ≥ 0?}); nitrogen rule {OK|violated}
{PASS|FAIL}             molecular weight     computed {x.xx} g/mol vs stated {y} ({Δ%}; {rounding|salt form|wrong compound})
{PASS|NOTE}             Lipinski subset      MW {x} vs 500; N+O = {n} vs 10; donors/logP not assessable
RESULT: {PASS|FAIL} — {what to do next}

Worked example

A biotech press release states: "our lead compound is caffeine, C8H10N4O2, MW 312 g/mol, a drug-like orally bioavailable small molecule." Running python3 scripts/chem.py check --name caffeine --formula C8H10N4O2 --mw 312 recomputes the weight from IUPAC standard atomic weights:

Element	Count	Atomic weight	Contribution
C	8	12.011	96.088
H	10	1.008	10.080
N	4	14.007	56.028
O	2	15.999	31.998
Total			194.19 g/mol

Name and formula agree — caffeine is C8H10N4O2. RDBE is 8 + 1 + 4/2 − 10/2 = 6, an integer and non-negative, so the formula is possible. The Lipinski subset passes easily: 194.19 ≤ 500 and N + O = 6 ≤ 10. But the stated 312 g/mol is 60.66 % above the computed 194.194, far outside tolerance, so the claim fails. The likely cause is a salt-form mix-up — caffeine citrate is heavier than the free base — but the skill flags rather than reconciles: the release pairs one form's formula with another's weight, and its author should say which was meant. RESULT: FAIL — re-source before use.

Verification

Before the verdict ships:

• Every arithmetic check was recomputed from the stated formula, not copied from the text being checked.
• RDBE was evaluated and is a non-negative integer, or the formula was rejected as impossible.
• The molecular-weight comparison states both the computed value and the percentage difference, and names the likely cause when it fails.
• Unverifiable name-to-formula pairs are marked unverified rather than guessed.
• Lipinski criteria that need a structure are marked not assessable rather than estimated.
• python3 scripts/chem.py check … was run (or the arithmetic hand-checked twice) and its verdict matches the emitted one.

Companion tool

scripts/chem.py (stdlib only, offline, deterministic) does the arithmetic of steps 2–5.

python3 scripts/chem.py formula "C8H10N4O2"                        # counts, MW, monoisotopic, RDBE, Lipinski subset
python3 scripts/chem.py check --name ethanol --formula C2H6O2      # per-item verdicts; exit 1 on any FAIL
python3 scripts/chem.py --demo                                     # the worked example above
python3 scripts/chem.py --selftest                                 # 106 checks against PubChem-verified values

Hydrates, brackets and charges parse (CuSO4·5H2O, [Fe(CN)6]4-); --json emits the verdict structure. Real output:

Claim: ethanol -- C2H6O2
Computed: C2H6O2  MW 62.068 g/mol  monoisotopic 62.036779 Da  RDBE 0
FAIL       name vs formula              ethanol is C2H6O; claimed C2H6O2 = ethylene glycol
PASS       formula plausibility         RDBE = 2 + 0 + 1 + (0 + 0)/2 - (6 + 0)/2 = 0 (integer, >= 0)
NOTE       Lipinski (assessable subset) MW 62.07 <= 500 OK; N+O = 2 <= 10 OK; donors/logP need a structure
RESULT: FAIL (1 of 2 checks failed) -- re-source before use

The skill works without the tool — every step is hand-computable — but beyond a single formula, run it and read the FAIL lines. Structure-level chemistry stays out of scope on purpose: a cheminformatics toolkit is a different dependency class from a stdlib script.

Pair with adjacent skills

• smiles-sanity-check — the structural counterpart: this skill checks prose claims, that one checks SMILES strings.
• grounded-fact-check — re-source compound data from primary literature when a check fails.
• abstain-or-escalate — when the chemistry is load-bearing and cannot be verified.
• rate-source-admiralty — grade the publisher supplying the compound data before trusting it.
• analyze-patent-claims — patent chemistry arrives inside claim language that needs its own parsing.

Anti-patterns

• Do not present a pass as chemical validation. A consistent name/formula/weight triple can still describe an inert or unsynthesisable molecule.
• Do not silently correct a formula or a weight. Surface the mismatch; autocorrection destroys the evidence that something was copied from the wrong source.
• Do not treat Lipinski violations as disqualifying. The rule is a heuristic for oral small molecules; biologics and macrocycles violate it legitimately. Report violations as confidence modifiers.
• Do not guess formulas for unfamiliar compound names. unverified is an honest answer; a guessed formula is a manufactured fact.
• Do not extend this skill into structure-level chemistry. That boundary is deliberate, not an oversight.

Reference

• C. A. Lipinski, F. Lombardo, B. W. Dominy and P. J. Feeney, "Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings," Advanced Drug Delivery Reviews, vol. 23, no. 1–3, pp. 3–25, 1997. doi: 10.1016/S0169-409X(96)00423-1
• F. W. McLafferty and F. Tureček, Interpretation of Mass Spectra, 4th ed. University Science Books, 1993, ch. 2 — rings-plus-double-bonds and the nitrogen rule. ISBN 978-0-935702-25-5
• T. Prohaska et al., "Standard atomic weights of the elements 2021 (IUPAC Technical Report)," Pure and Applied Chemistry, vol. 94, no. 5, pp. 573–600, 2022. doi: 10.1515/pac-2019-0603
• CIAAW, Standard Atomic Weights, IUPAC, 2021 — the conventional values the companion tool uses. https://ciaaw.org/atomic-weights.htm