Why it works this way2 min read

Evidence before advice

The console has a language layer. It is deliberately the last layer, and it is never the source of a fact.

Rules decide, models narrate

Anything a rule can determine, a rule determines. Seven of them run over metadata Lance already reports, cost nothing, need no key and no network, and each carries the literal figures it was computed from — so a claim can be checked rather than believed.

A model, when configured, only ever narrates that. It rewrites findings into sentences and turns questions into filters. It cannot disagree with a finding, because the finding is the numbers.

This ordering is also why the tool is useful with nothing configured, which is the state most people first see it in.

A caveat travels with the number

Some numbers are right and would be wrong to act on. num_small_files is the standing example. So a finding carries four things: a claim, the evidence behind it, a caveat where the number lies, and a suggested action — which is sometimes leave it alone.

A check that could not run says so

The first version of the findings engine caught every exception and continued. That made a broken rule indistinguishable from a clean table — the one failure mode a panel whose entire job is honesty cannot have.

Failures are now captured per rule and reported. A partial analysis renders as its own state, on every panel, not as a variant of "nothing to report". And a summary written during one says it describes an incomplete picture.

Every uncertain state has a name

A remote connection that cannot be browsed. A query one version cannot answer. A model whose price we do not know. A scan that outran its timeout and is still running, because Lance offers no way to interrupt one. Each of those has its own wording, because collapsing them into a generic error — or worse, into a plausible-looking success — is how a tool teaches people not to trust it.

The model never acts

A translated filter lands in the filter box for you to read. Nothing runs until you say so, and the agent tool set contains nothing that writes or spends. Output is a draft; the decision stays yours.