AssayVault

Stability studies (ICH Q1A) produce large volumes of measurements across formulations, storage conditions, timepoints, and methods. The data arrives as messy per-instrument exports - SEC-HPLC, CEX-HPLC, DLS, HIAC, potency - with cryptic sample identifiers and no shared schema.

The question that matters is regulatory: the supported shelf life under ICH Q1E. Answering it takes three separate statistical judgements - per-batch regression, poolability of batches, and admissible extrapolation - that must line up. AssayVault models exactly that chain, end to end.

The awkward part is that none of those three judgements is independent. Whether batches may be pooled changes which regression carries the claim. The claim in turn decides how far beyond the observed data the answer may reach. A spreadsheet holds any one of them comfortably. Holding all three at once, for 24 molecules across 52 projects, is where it stops being the right tool.

The warehouse is 34 dbt models over DuckDB, layered into staging, marts and analysis. The whole graph runs as a Dagster asset job with its data-quality checks wired in as blocking asset checks: nine assets, five checks, one command.

A shelf-life answer per attribute, with its reasoning attached. 24,266 evaluated attribute and condition combinations resolve into 56 claims, each carrying the regression, the poolability verdict and the extrapolation branch that produced it.

A specification history that survives revisions. 363 effective specification versions sit behind the results as a slowly changing dimension, so a result keeps the verdict it had on the day it was measured. Tightening a limit in 2026 does not silently re-flag 2024.

A risk ranking across 19,016 units. Which combinations of formulation, lot, storage condition and attribute deserve attention before they leave specification.

A reference library of what is already on the market. 199 marketed antibody products drawn from FDA, EMA and NMPA labels, queryable by composition, so a new formulation can be compared against approved practice rather than against memory.

The data is synthetic. Every number on this page comes from a generator, not from a laboratory. That is a licensing constraint rather than a shortcut. The deliverable is the pipeline, the guards and the validation, all of which behave the same once real history is plugged in. The ROC-AUC in particular measures how learnable the simulated degradation process is. The model card says so in those words.

Snowflake is prepared, not proven. The same dbt models build against a second profile, which has not been run against a live account.

Nothing here has been audited. The system is built to the habits that make an audit possible: versioned rules, reproducible builds, a traceable path from instrument export to reported number. Whether it would satisfy a particular inspection is a question for a validation exercise that has not been run.

The regulatory logic is an implementation, not an authority. It follows ICH Q1E and the 2025 consolidated draft as a versioned rule base. Every result records which rule set produced it. It does not replace a regulatory assessment.