Agentic Spectra

A spectral model reports an answer and a confidence. Agentic Spectra decides, before anything is reported, whether that answer has earned trust.

Every spectrum gets a recorded decision: analyzed and reported, reused, refreshed, marked degraded, or refused. The decision draws on evidence the model itself does not use, such as the spectral bands of the compound it claims and the transfer standards a new instrument declares, and it is kept, with its reasons, in a verifiable record.

The result is measured against conventional practice on public data, with the analysis fixed before the held-out data were scored.

Identifying minerals, including ones missing from the library

The RRUFF reference collection: 1,333 mineral species, one library spectrum for each of 1,000 of them, and 3,940 query spectra, of which 1,279 come from the 333 species the library does not contain. Every answer about an absent species is wrong, and a useful system has to know when not to answer.

Ranking of answers by trust	Error area
Classifier confidence (conventional)	0.547
Agentic Spectra admission	0.378

The error area is the share of wrong answers among those reported, averaged as more answers are reported in order of trust; lower is better. Agentic Spectra lowers it by 31%. The 95% interval of the difference, resampling whole species, is −0.19 to −0.15, significant after correction for the experiments run together.

Moving a calibration to a new instrument

The same mixtures of glucose, sodium acetate and magnesium sulfate, measured on eight spectrometer setups from seven makers, at 532 and 785 nm. A model calibrated on one instrument is moved to each of the other seven with ten transfer standards the new instrument declares: 56 instrument pairs, 3,049 held-out predictions. A prediction is harmful when its error exceeds twice the calibrated model’s own error.

Ranking of predictions by trust	Harm area
Slope/bias correction with the model’s outlier guard (conventional)	0.587
Agentic Spectra admission	0.416

Harmful predictions move down the order of trust: the harm area falls by 29%, with a 95% interval of −0.21 to −0.12, resampling whole mixtures. It held again, at 0.435, in a later run with a revised rule.

What these results establish

Both results are about ranking: which answers to trust first. Admission that looks at evidence outside the model orders them significantly better than the model’s own score, on held-out public data.

A reporting threshold chosen on one source did not carry over to another, and on the mineral set no threshold reached 5% wrong answers for either method. A fixed operating point is not claimed.
Robustness to acquisition faults and savings on continuous spectra were tested against conventional methods and are not established.

How it was measured

Four experiments, each against the method a practitioner would use by default, with its standard safeguard.
The analysis plan, the primary measures and the test were written down before any result existed; every later change is a dated amendment.
Rules were developed on separate data and frozen by hash before the held-out data were scored, once.
Intervals come from resampling whole species or mixtures, with Holm correction across the experiments.
A later run with revised rules is reported beside the first, never in its place.

Agentic Datasets for spectra

Agentic Spectra applies the Agentic Datasets model to spectroscopy.

Sources declare the evidence they provide; analyses declare the evidence they require; admission is evaluated before execution.

Every source is described by what it declares: instrument, excitation, channels, the tasks it may serve and the ones it may not.
Every decision is written to a hash-chained record; a replay of the same requests reproduces every decision.
The published Agentic Datasets conformance suite passes 13 of 13 assertions, and each of seven deliberately broken engines is caught by the assertion it breaks.
Built on the Agentic Datasets reference implementation, v0.2.0.

agenticdatasets.org — the specification agenticvision.ca — the same model for cameras

Public data

Dataset	Used for	License
RRUFF, excellent unoriented (Lafuente et al., 2015)	identification, held out	CC BY 4.0
Open-source Raman spectra of API compounds (Flanagan & Glavin, 2025)	identification, development	CC BY 4.0
Bioprocess analytes on eight spectrometers	instrument transfer	CC BY 4.0
Fuel spectra, benchtop and handheld	instrument transfer	CC BY 4.0
Time-resolved LC-Raman of amino acids	continuous spectra	CC BY 4.0

Each source was fetched with its checksum recorded and described by what it declares; values its card does not state are recorded as not declared.

Current status

Agentic Spectra is an independent technical project initiated and maintained by Alexander Chernov, and an application of Agentic Datasets to spectroscopy.

The results above are research measurements on public data. Nothing here has been validated for regulated use; a product release decision is a task every source declares prohibited.

Maintainer: Alexander Chernov · ORCID 0009-0007-3198-2712