What does (unverified) or ENSEMBLE INCOMPLETE mean in an answer?

When a language model writes the answer, the panel checks each number in it against what the agent fetched this turn: the tool results, the packet on screen, your own question and any fact you asked it to remember. A number that matches none of them gets (unverified) added straight after it, in italics. The check allows for rounding and for unit conversion between mph, kt, km/h and m/s, and percentages are allowed. Times such as 07:00, years, and whole numbers from 1 to 3 are ignored. The mark is per number, not per answer, so the rest of the sentence may be fine.

Under each answer is an evidence line. It lists each source the agent used with a UTC time, for example the forecast or the nearest station, and says Workspace packet when it used what is on your screen. After that it shows who wrote the words: a narrating model, the engine with no model, or a standard refusal. If any numbers were marked it reads, for example, "2 numbers unverified". If none were, it says the numbers were verified against tool output. That is a check against the agent's own fetched data. It does not mean the forecast itself is right.

Do not use that number. Ask again with the place, the height and your limit in the question, which gives the agent something to fetch against. You can also read the same figure in the instrument at /app, where every value is labelled with its source. The agent's own contract is that every figure comes from a tool or the packet, so a mark usually means it did arithmetic or rounding of its own, quoted a figure from an earlier turn, or repeated a number from memory. Earlier turns are context, not data, and are not accepted as sources.

The ensemble is the set of members from three ensemble systems, GEFS, ECMWF ENS and ICON-EPS. The agent expects a known number of members at your height and hour. When fewer arrive, you see ENSEMBLE INCOMPLETE with received over expected, such as 28/31. It appears in the readout in the inspector, in the first sentence of the agent's answer, and in the engine's answer. Fewer than 10 members at an hour means no odds are given for that hour at all. Odds are plain member fractions, so with fewer members each one counts for more. Read the percentage as rougher, and look at the models chips to see which ensemble is short. See reading P(>limit).

If the narrating model is not configured, times out, errors or returns nothing, the agent switches to an engine. The answer then starts with a line saying the model is unavailable and that this is the engine speaking, a deterministic read of the same data with no narration. The engine looks up the place, the wind at your height, the ensemble and the nearest observation. For a question about your sites, such as which to shut first, it compares your saved sites if you are signed in. It reads place, height, limit, hour and "tomorrow" from your wording, so a question without a place gets a reply asking you to name a place, a height and your limit. The engine cannot save a rule, remember a fact or change the brief, and its numbers are not run through the check.

If even the engine fails you see Agent unavailable followed by the reason, in red. The agent allows 12 questions a minute from one connection, for everyone, signed in or not. Past that the reason reads that the rate limit is reached and to try again in a minute. There is no daily cap and no plan limit on questions. A question must be 2 to 600 characters, and a very large request is refused. A single tool can also fail inside an otherwise good answer. That shows as a line in the list of what the agent checked, with a cross, the tool name and "failed". The answer is then built on what did arrive.

Name the place, the height and your limit, for example Clonmel, 85 m, 25 mph, 07:00 tomorrow. Ask one thing at a time. If you open the panel from /app the agent already sees your place, height, limit and hour, so ask about that directly. If a number is marked, ask for just that figure. If the ensemble is short, ask again later. The instrument at /app shows the same members, with the count of members received next to each ensemble model.

Every figure in an answer should match what the instrument shows for the same place, height and hour. The agent compares the forecast to your limit and stops, so the decision is yours. Treat the instrument as the place to check the working: the hero readout, the exceedance fan, the agreement spine and the observation line. If the agent and the instrument disagree and you cannot see why, trust the instrument, and contact us with the question and the time.