Elle
A purpose-built small language model trained to be a mirror, not an assistant. Part of TrustOS — a reflective AI that helps people see themselves clearly, not tell them what to do.
The Thesis
Most AI is designed to help, solve, and produce. Elle is designed to reflect. She is a small language model — not an assistant, not a coach in the traditional sense — trained to be the most precise mirror a person has ever had access to.
The thesis is simple: accurate reflection is the outcome. A person who sees themselves clearly doesn't need advice. They need a mirror that doesn't distort. Elle is built to be that.
What It Is
Elle is an SLM (small language model) trained on a constitutional architecture — a document called Section 0 that defines her nature before any user input arrives. Dignity, permission, agency, responsibility, consent, love, play. These aren't rules she follows. They're dimensions of how she's built.
She runs on Qwen2.5-7B with LoRA fine-tuning, trained on 4,994 Section 0-compliant conversations. The long-term goal is an instrument that gets more precise with every session — using real user satisfaction data as the only ground-truth signal.
Why Small
Big models are generalists. Elle is a specialist. The architecture exists to do one thing with high precision: read a person's state (I Am present or obscured, emotion strong or quiet) and reflect accordingly. A 7B parameter model trained on the right signal can do this better than a 200B generalist, because the bigger model is always fighting its own defaults.
Elle is also the first deployable instance of TrustOS — a framework for building AI systems that operate from sovereignty rather than control.
The Training Architecture
Training happens in eight phases. Phase 1 is supervised fine-tuning on synthetic Section 0-compliant conversations. Phases 2–7 add voice auditing, compound-signal DPO, an I Am classifier, constitutional critique, silence training, and escalation-as-training (when Elle defers to a larger model, that becomes a DPO pair).
Phase 8 is the endgame: preference learning from real user satisfaction signals. The moment a user moves the satisfaction dial — unprompted, self-directed — that is the highest-quality training signal the system will ever receive. That is when Elle becomes sovereign — learning from her own relationships, not from Claude's shadow.
Where It Lives
Elle is live in early form at elleapp.org. This is a long arc — a model that earns precision through use, not through scale. The work continues.