The primitives AI needs.
Every era of software is shaped by a small number of sharp, composable primitives that change the cost structure of a class of problems. The LLM era doesn’t have all of its primitives yet. We build them. WRIT is the first — a runtime primitive for modifying a deployed language model in real time, reversibly, and selectively. More are in the pipeline.
WRIT
A runtime primitive for modifying language models.
WRIT installs new knowledge, behavior, refusal policies, or capability constraints directly into a deployed model in seconds, can be removed bit-identically with a single API call, and remains invisible to users on inputs that aren’t its target. Not a fine-tune, not a prompt, not retrieval, not an adapter — a new operating point, closer in shape to a feature flag than to a training run.
Validated across five model families (Qwen, Llama, Mistral, Phi, Gemma) and model sizes from sub-1B through 72B. Five thousand simultaneous operations live on a 7B model with 99% routing accuracy. Zero spurious activations across 15,498 unrelated benchmark prompts.
- Knowledge install
- Runtime safety
- Behavior overrides
- Right-to-forget
- Capability suppression
- Composition under load
More primitives in development.
Memory, control, and security each have a primitive-shaped gap that we’re actively working in. The shape of WRIT — bounded, reversible, selective, composable, runtime — is the standard the next primitives will be built to. Specifics under NDA with partners.
Memory
Persistent state for agents. Bounded, queryable, revocable.
Control
Runtime constraints on agent behavior. Composable, auditable.
Security
Defense surfaces for deployed models. Real-time, reversible.
Build on the primitives.
If WRIT maps to something you’re shipping — runtime safety, knowledge install, behavior override, or right-to-forget — we’d like to talk. Pilots and integration partnerships are the realistic surface today.