Artmetria
betaResearch moduleSleepers
A structured analytical surface to detect potential discrepancies between catalogue attribution, visual evidence, and taxonomic coherence — and to identify possible undervaluation as a hypothesis, not a verdict.
What it does
Market discrepancy detection, expressed as probabilities and signals.
Sleepers compares catalogue claims with corpus evidence and explicit negative controls. Outputs are calibrated for interpretation — not for standalone decision-making.
- Detect potential discrepancies across attribution, image regions, and taxonomy.
- Analyze visual and taxonomic signals with preserved explainability.
- Identify possible undervaluation only as a structured hypothesis with confidence bounds.
How it works
Signals are normalized, weighted, and calibrated — then read with care.
The pipeline produces a mismatch probability, a confidence estimate, and an expected-upside index. Each step is designed to remain inspectable.
Visual, regional, taxonomic, and undervaluation-related inputs are normalized to a common scale.
A weighted composition yields a latent score, then a bounded probability via a smooth transform.
Confidence reflects match dispersion, top-match agreement, and penalty strength from negative examples.
Example analysis
A real lot the engine surfaced
An actual weakly-attributed lot, shown in retrospect: the engine read a visual affinity the catalogue missed, and the lot later hammered well above its estimate. A closed past sale, shown to illustrate the method — not advice.
.jpg?mode=max)
- Catalogued as "SHENG MAO (ATTRIBUTED TO, ACTIVE 1313-1362)" — a weak attribution the engine re-read.
- Closest visual affinity with Yuan Yao in the museum corpus.
- Later sold for 62,500 — +213% above the house's high estimate.
Interpretation
Responsible reading
Artmetria provides structured analytical context. Sleepers does not authenticate works, issue valuations, predict prices, or recommend transactions.
- Outputs are probabilistic and may be wrong; they require expert review when material decisions are at stake.
- Corpus coverage and catalogue quality affect conclusions; gaps in data appear as uncertainty, not silence.
- Expected upside is a relative index within the model, not a forecast of financial return.
Use cases
Where teams use Sleepers
Research triage
Rank candidates where attribution, corpus similarity, and taxonomic paths diverge meaningfully.
Catalogue review
Support specialists who must articulate why a line is coherent — or why it deserves a second look.
Institutional workflow
Attach structured rationale to internal memos without replacing connoisseurship or provenance research.
Beta status
Active development
Interfaces, weights, and coverage will evolve. The module is in active beta — read every output as a hypothesis to verify, not a verdict.