DocNexus.ai · case study
Self-verifying answers on any warehouse
Attaching a verified / unverified verdict to every aggregate answer, without gold labels.
For customers who bring their own warehouse, an agent cannot be checked against a known answer. So the answer is checked against itself, and the joins it relies on are proven from the data rather than approved by hand.
Methods and relative results only. Client figures, internal names and scale are not published.
Metamorphic self-checks
Each aggregate answer is re-derived through rewritten variant queries (partition sums, distinct-count bounds, join fan-out, summary reconciliation) within a fixed budget. A check that does not finish means unverified, never failed. The method states its own limit: it verifies consistency, not intent.
Joins proven by evidence
A survey of commercial and academic join layers found that shipping products rely on people to approve joins; only research systems test value overlap against the data. The design profiles keys, then accepts a candidate join only when value containment clears ≥0.9, and learns from SQL that actually executes.
On the TPC-DS benchmark with foreign keys hidden, join recovery went from 0% to 100%. The honest number is the held-out one: on a schema the rules had never seen, it recovered 62% before fixes.