Closed-set classifiers recognize known fakes.
PELLUCID reads unknown chemistry.
This walkthrough simulates the full PELLUCID pipeline on a sealed pharmaceutical shipment: acquire a packaging-corrupted spectrum, invert the physics with THE LENS, then hand the recovered chemistry to a live AI — THE LEXICON — which describes what is in the box in open vocabulary and issues a calibrated, auditable VERDICT.
Neural-operator inversion of light transport. Strips foil, bottle, blister, humidity — recovers the chemistry underneath, quantitatively.
Maps recovered spectra into a shared embedding with the written corpus of chemistry. Names substances it was never trained on.
Generative counterfeit digital twin. Manufactures synthetic threats to train and red-team the system before the market does.
Calibrated risk, plain-language narrative, spectral evidence trail. Built for FDA auditability and port-throughput tempo.
Scan Bay — select a sealed shipment
Each scenario is a sealed unit on the inspection line. The raw acquisition is what any TA1 sensor actually sees: attenuated peaks buried under packaging fluorescence, container signatures, and noise.
THE LENS — physics inversion
A physics-informed neural operator inverts the radiative transfer through the packaging. Watch the fluorescence baseline collapse, container peaks subtract, and the attenuated chemistry sharpen. Packaging is treated as a causal intervention — not noise.
Run the inversion to extract the recovered peak set.
THE LEXICON — open-vocabulary chemical read
This step is driven by a live Claude model acting as the chemistry-language layer. It receives only the recovered peak table and the declared product — never the answer — and must describe what is physically in the box, including substances outside any training catalogue.
THE VERDICT — calibrated, auditable decision
The regulator's view: one calibrated risk score, a plain-language narrative, and a spectral evidence trail. Rendered as the certificate an FDA reviewer or port operator would actually sign off on.
Run THE LEXICON first — THE VERDICT is generated from its structured output.
THE FORGE — generate a counterfeit that doesn't exist yet
The FORGE is the standing red team. Here, a live Claude model invents a novel, physically plausible adulteration scenario — a threat no database has catalogued — and injects it into the Scan Bay so you can run the full pipeline against it. If PELLUCID can read the FORGE's inventions, it can read the market's.
Architecture — sensor-agnostic by design
Live AI: THE LEXICON and THE VERDICT are produced per-run by a Claude model reasoning over the recovered peak table (it is never shown the scenario's ground truth). THE FORGE invents genuinely novel adulteration scenarios per click.
Simulated: spectra are synthesized from published Raman peak positions of real APIs, excipients, adulterants, and packaging polymers; LENS inversion is an animated stand-in for the neural-operator model to be trained under Phase 1.
This file runs as a Claude artifact with zero configuration. For public deployment on the standard portfolio stack (GitHub → Netlify), route the two AI calls through a Netlify function proxy (ESM, native https, ~8.5s internal timeout, max_tokens ≤ 1000) — the repo build tries the bundled Netlify function first and falls back to the direct API in artifact preview.