The path to 100% AI accuracy.

An undefined business term.

A business term can mean more than one thing, and the model picks one without telling you.

“Rank the accounts by VPN abuse.” The schema doesn’t define a VPN abuser in any column or value, and the benchmark only says what qualifies if somebody asks. Without Ava (Ask, Verify, Answer), the model invents a threshold. Ava asks the user, or the agent that sent the question, what qualifies before anything computes. Canary will add that answer to its test run, so “VPN abuser” resolves the same way every time.

An unstated ranking measure.

The question asks for a ranking but doesn’t say what to rank by.

“Give me a ranked list of stars in the galaxy.” The database is an astronomy catalog, and size, distance and brightness are each a different calculation. Since the question didn’t name a measure, the result came back sorted alphabetically by name. Ava asks which measure you mean before computing. Canary will then check on each run that a ranking request with no measure still makes Ava ask.

A metric that lives in a document.

The formula’s written out in a document and isn’t defined as a metric, so the model has to guess the SQL.

“How reliable is this equipment?” The knowledge base defines the Composite Reliability Index (CRI) as a formula, so there’s no metric to query. A data engineer adds CRI to the semantic model, computed from the RELIAB_IDX and BSCSI columns. ACE (the AI Computation Engine) holds the SQL that computes CRI, and CRI will join the Canary suite so each run checks the result.

A join that fans out.

The answer takes more than one pass, and a wrong pass still returns a number that looks right.

“What share of each doctor’s patients take their medication as prescribed?” One table holds patient visits, and another holds the treatments given at each visit. Joined naively, a visit with three treatments counts three times. In the data, 123 of 351 visits had more than one treatment, and 42 of 157 doctors came out wrong. The semantic model sets the visit as the unit to count. ACE catches the fan-out: it counts each visit once, then rolls visits up to each doctor. Canary will check that a visit still counts once, however many treatments it carries.

A region spelled several ways.

One place is stored several ways, so filters and groupings miss rows that should match.

“How many households are in each region?” “Chicago,” “CHICAGO” and “chicago” with a stray tab on the end are three stored values for one real place, and a GROUP BY can’t tell they’re the same, so it counts three regions. The semantic model maps every variant onto one region. Canary will check the spellings in the data on each run, and Semantic Modeling reconciles any new variant it finds.