Inside an AI build, what the L4 corpus actually shows.

"AI-analyzed" (L4) means we have looked at a specific project against a published methodology, what was built, how it was evaluated, and whether it reached production. This piece reads across that corpus to answer one question buyers keep asking: what actually separates the systems that ship from the demos that never do? As the corpus grows the specifics will sharpen; the patterns below are already clear.

A demo proves capability; production proves discipline

Almost any competent team can stand up an impressive demo in a fortnight. The gap between that and a system real users depend on is not intelligence, it is discipline across a set of unglamorous concerns that demos are allowed to ignore.

The five things production systems have that demos don't

• An evaluation harness. Production systems are measured continuously: retrieval quality against a golden set, task success rates, regression checks that run before every deploy. Demos are judged by vibes. The presence of a real eval harness is the single strongest predictor that a build reached production.
• Cost and latency budgets. A demo can call the largest model for every request. A production system has to hit a latency target and a per-request cost, which forces real engineering: caching, smaller models for easy cases, retrieval that actually narrows the context.
• Failure handling. Production systems assume the model will sometimes be wrong and design for it, confidence thresholds, fallbacks, human-in-the-loop on the high-stakes path. Demos assume the happy path.
• Monitoring. Shipped systems know when they degrade, because someone instrumented them. The teams that skip this find out about problems from users.
• A data and retrieval layer that was actually engineered. In the RAG systems we have analysed, quality lives or dies in retrieval, not in the model. The production builds invested in chunking, indexing, and evaluation of retrieval as a first-class problem; the demos bolted a vector search onto a model and hoped.

What this means for buyers

When you evaluate a partner, you are really trying to tell which population they belong to: the teams that have operated the five disciplines above, or the teams that can produce a convincing prototype and have not. The questions that separate them are concrete, *show me your eval harness; what is your latency budget; what happens when the model is wrong*, and they are exactly what the L4 analysis looks for.

The corpus is young, and each new AI-analyzed project makes the reading sharper. But the headline is already stable: production AI is not a smarter demo. It is a different discipline, and it leaves evidence.