01

drip · methodology

novel peptide candidates, delivered via licensed Aurora technology.

drip operates the unified peptide pipeline. discovery, prediction, structure, simulation, scoring, and calibration ship as one API. the underlying peptide stack is licensed from Aurora and benchmarks publish under the Aurora model card.

aurora-licensed technology · drip-operated pipeline · output-only api

02

benchmarks

interim benchmarks

the following figures reflect the development training run as of early 2026. dataset version, evaluation protocol (k-fold or scaffold split), and negative-set construction ship alongside final benchmarks at launch with reproducibility CSVs and dataset DOIs. these numbers are disclosed now because withholding preliminary data reads as cherry-picking; the standard CADD literature pattern is to disclose with caveats.

what this disclosure does not yet include: holdout set size, train/test split methodology, null-model baseline. those ship with launch alongside the full reproducibility CSV per benchmark and dataset DOIs. one row per sequence, three columns: sequence, model_score, ground_truth_label. researchers download, recompute the metric, verify the claim.

03

methodology

train / test split

every benchmark uses public, published splits exactly as the upstream dataset defines them. no reshuffling, no peeking, no cherry-picking. the harness, the dataset hashes, and the numerical results land in the aurora model card so the work is independently reproducible.

the discipline check is public; the architecture is not. drip publishes the splits, the harness, and the reproducibility csvs. drip does not publish how the underlying peptide stack works internally. that is aurora's trade secret, not drip's to disclose.

04

reproducibility

aurora model card + DOI

the model card is authored and owned by aurora. drip cites it here, in the researcher sdk docs, and in customer onboarding. the card publishes behavior, training-data scope, intended use, known limits, and contact. it does not publish architecture, weights, or mechanism — by license design.

aurora mints the zenodo doi when the model finalizes. the doi becomes the citation key everywhere drip references the underlying peptide stack.

05

not published

what stays with aurora

published: the benchmark harness, the dataset citations, the public splits, the reproducibility csvs, the aurora model card, and the doi. anyone can recompute the headline numbers from first principles.

not published: weights, training corpus, training procedure, inference internals, architecture beyond what the model card discloses. these are aurora-owned trade secrets. drip does not host them, redistribute them, embed them, fine-tune on them, or hand them to any customer at any tier — the license forbids every one of those modes. customers consume output via api; that is the entire interface to the underlying stack.

06

references

cited literature