Scope · Anchor · Audit · Loop. The diagram I draw at every kickoff. A way to talk about an LLM product as a system of seams — not a model with a wrapper.
Just as SaaS abstracted software delivery and PaaS abstracted infrastructure, SAAL defines a model for delivering AI capability as a structured, reliable layer within modern organizations.
SAAL is not about building models — it's about building the systems that make AI trustworthy, observable, and composable at scale. It provides a framework for teams who need AI that actually works in production, not just in notebooks.
Full framework documentation coming soon.
What this system is for. What it isn't. The boundary you'll defend in code review.
The eval set, the source of truth, the human who decides when prod disagrees with bench.
Continuous traces, sampled. Re-labeled quarterly. The discipline of looking at what you shipped.
The escape hatch and the feedback channel. How the model gets less wrong over time, with humans in it.
// in detail
Most LLM products fail at scope before they fail anywhere else. "Customer support assistant" is not a scope. "Answers questions about a customer's active plan, using only documents in the billing index, and escalates anything else" is a scope. The discipline is to write it down, in one sentence, and defend it in code review.
The anchor is your ground truth — the eval set, the source documents, and the human who decides when production disagrees with the bench. Without an anchor, every change feels like an improvement and every regression is debatable.
Audit is the discipline of looking at what you shipped after you shipped it. Every quarter, sample 200 real production traces, hand-label them, and compare against the anchor. Most teams stop here. Audit is the cheapest 10× improvement on most LLM products.
The loop is how the system gets less wrong with humans in it. That means an escape hatch — a wired-to-a-real-person feedback channel — and a cadence for closing it. The loop is what turns your AI product into a learning organization. Without it, you have a feature that decays.
If you can name the four pieces and your team can't, the framework hasn't shipped yet. Tape it to the wall.
The team won't say no to a use case. Scope creeps. The model is asked to do five things and does each badly.
The bench is a notebook somebody made once. Nobody owns it. Nobody trusts it. It is wheeled out before launches and ignored after.
Audit isn't on the roadmap. Engineers want to ship features; auditing last quarter's feature feels like overhead. It is the work.
The escape hatch was a Slack channel. Nobody read it. Users learned to bypass the AI surface entirely within a quarter.
SAAL is released under CC-BY 4.0. Use it in your decks, your docs, your team wiki — no permission needed. A link back is appreciated but not required.