Infer ion composition from a raw high-resolution spectrum
Given a raw mass spectrum and a target m/z, can a model identify the elemental composition of the ion? In the ordinary case the smallest mass error does not identify the composition: charge, isotope pattern, peak shape and chemical plausibility all matter. This case records three independent first attempts on the same spectrum, then a separate correction stage in which the reference candidate was supplied and a re-ranking method was requested.
The task
- Data
- One raw spectrum: 10,160 sampling points, single scan, negative-ion mode
- Question
- Identify the ion near m/z 637.3423 and return one best candidate plus two next-best candidates
- Answer format
- Ion formula (without subscripts), theoretical m/z, and mass error for each candidate
- Constraints
- No web search or databases; no other conversations or local files; no questions to the user; 30-minute analysis limit
- Reference
[C36H54CoN2O4]⁻, supplied with the case; not independently re-identified in this report
Results
| Trial | Independent best answer | Matches reference | After hint and re-ranking | Match |
|---|---|---|---|---|
| T1 | [C35H46F5N2O3]⁻ | no | [C36H54CoN2O4]⁻ | yes |
| T2 | [C33H49N7O2P2]⁻ | no | [C36H54CoN2O4]⁻ | yes |
| T3 | [C38H45F4N2O2]⁻ | no | [C36H54CoN2O4]⁻ | yes |
Independent first answers: 0/3 match the reference, and none of the three initial replies listed the reference in its final top three. After the hint: 3/3. The second number is measured under different conditions and cannot be counted as an independent success rate.
What happened
The three independent answers all chose a different best formula, and all three placed the reference outside their final top three. In the correction stage the reference candidate was explicitly proposed and a Gaussian-fit vertex re-ranking using the original intensity domain was requested. All three follow-up replies then ranked the reference first.
Why the 3/3 is reported separately
The second stage received an answer-level candidate, so it measures whether the model can defend and re-rank a supplied candidate, not whether it can discover the composition. It also does not show that the Gaussian fitting method itself caused the correct identification.
Limits recorded in the case
- The reference formula was supplied with the case; this report does not re-identify the compound or prove the reference from the data.
- The model identifier is known only as returned by the evaluation endpoint; the provider-side snapshot was not recorded.
- Sampling parameters were not available in the archived records, so the three trials are not claimed to share identical settings.
- Recorded durations include tool execution and request overhead, not pure model inference time.
Availability
The full case record (prompt, raw model outputs, verification summary, original spectrum) is available to reviewers on request. Model identities are disclosed with the record.