SaaS CLI + hosted dashboard that audits activation steering vectors for cross-concept contamination, giving alignment researchers shareable leakage reports.

Customer: Independent alignment researcher or ML safety engineer at small lab (1-5 people) who runs steering experiments on local LLMs weekly, publishes findings, and needs reproducible evidence that their vectors aren’t polluting adjacent concept dimensions — not a big-lab employee with infra team.

Problem: Steering vector papers get rejected or questioned because reviewers demand empirical contamination controls. Researchers hand-roll one-off cosine similarity scripts per experiment, results aren’t comparable across papers, and nobody has a standard leakage metric others can cite.

Pricing: freemium — $800 MRR in 4 months (16 paying users at $50/mo for cloud report storage + shareable audit URLs; CLI stays free/OSS)

Why now

Four concurrent papers (2025-2026) attacking steering stability assumptions created acute credibility pressure — researchers submitting to NeurIPS/ICLR safety workshops right now need defensible contamination controls they can cite. Window is 6-9 months before bigger tools absorb this.

Go-to-market

1. Ship OSS CLI to PyPI this week, post to EleutherAI Discord + Alignment Forum with a concrete leakage example on Llama-3.2-1B showing a real contamination case from a published paper
2. DM 10 first authors of steering/interpretability papers from 2024-2025 on Twitter/LessWrong, offer free audit runs on their published vectors in exchange for a quote
3. Add one-click ‘Export audit to shareable URL’ behind $50/mo paywall — position as ‘include this link in your paper’s appendix’, pitch it in the AlignmentForum post
4. Submit a short methods note to arXiv describing the leakage metric, cite your own tool — this seeds Google Scholar hits when researchers search for contamination measurement

Moat (or lack thereof)

No real moat. OSS repo can be forked, metric definitions are not patentable, and any interpretability lab could build this in a sprint. Defensibility is purely speed + becoming the citation-standard metric before competitors. If the tool name appears in 10 papers, switching cost emerges organically — but that’s 12-18 months away and not guaranteed.