Spaces:
Running
title: README
emoji: 👁
colorFrom: gray
colorTo: indigo
sdk: static
pinned: false
The AI engine for business.
Market Standard is a full-service technology studio. We design, build, and ship production AI, computer-vision, and LLM systems — from custom models and data pipelines to the Standard Suite of developer tools. On Hugging Face we publish our open-source research, beginning with Panoptes.
Panoptes — calibrated evidence for AI-text attribution
Our flagship open-source release reframes AI-text detection as evidence, not verdicts. Rather than a binary “AI or not” label, Panoptes estimates how much of a document is AI-authored, tells you how much to trust that estimate, and measures whether the evidence still holds up when the population shifts.
What it gives the community
- A participation score (0–100% AI-authored) with split-conformal coverage — an estimate with valid uncertainty, not a false-precision verdict.
- A frozen, preregistered open-weights detector (three-seed DeBERTa-v3 ensemble) with verified hashes — a strong, reproducible baseline anyone can run.
- An evidence-transportability harness — cross-dataset leave-one-out and calibration-transfer matrices that quantify how scores and calibration generalize across MAGE, RAID, DeFactify, CoAuthor, M4GT, and EvoBench.
- A license-audited dataset registry and split manifest, so others can add cohorts and extend the corpus on solid legal footing.
- A preregistered protocol with signed cards, so every reported number is reproducible.