A person should always be able to ask a machine why, and get an honest, complete answer. That sentence is easy to print on a wall. It is expensive to build.
This paper treats legibility as an interface property — not a documentation afterthought, not a post-hoc explanation layer bolted onto an opaque core. If the answer to “why?” cannot be produced from the same representation that produced the action, the system is not legible. It is narrating.
The cost of honesty
Complete answers require systems that keep intermediate structure: which inputs mattered, which constraints bound the decision, which alternatives were considered and discarded. That structure has a memory cost, a latency cost, and a design cost. Teams that optimize only for output quality quietly erase it.
We propose a minimum bar: every consequential action must expose a why-trace that a non-expert can inspect without opening a second product. If the trace is only available to engineers, it is not an interface property. It is an internal log.
What we are not claiming
We are not claiming that every weight in a model can be made intuitively meaningful. We are claiming that the reasons the product acts — the ones that affect a person’s next ten minutes — can and should be inspectable. Those reasons live closer to the product’s decision surface than to the full parameter space.
The work ahead is mostly unglamorous: schemas for traces, language that does not overclaim, and product decisions that refuse to ship when the why-trace would be fiction.