You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Status: Open — reproduced on main; no fanout guard in the ReAct answer path.
Impact: Detail joins can multiply an owner-level metric even when the SQL runs faithfully, which affects totals and business comparisons.
Evidence
Neither run_sql nor submit inspects join cardinality.
Reproduced on main with the default model. Asked "For each supplier, what is the total sourcing share percentage and the total quantity shipped?", the agent LEFT JOINed sourcing and asn on supplier and summed both. For the largest supplier it reported 107,994,000 and 573,044,580 instead of 6,000 and 9,550,743: each measure was multiplied by the other table's row count (17,999 ASNs and 60 sourcing rows).
Actions
Detect owner-to-detail fanout in submit from the join keys and the catalog's uniqueness: declared keys, sampled uniqueness, and #272. Return it to the model as a tool error so it projects at the owner grain, uses a unique-owner subquery, or pre-aggregates the detail.
Acceptance
Test owner metrics joined to 1, 2, and 19 detail rows. Require invariant owner values across direct projection and valid pre-aggregation, and reject SUM-after-fanout variants. Cover grouped and ungrouped SQL, cardinality uncertainty, and independent keyed-row oracles.
Review (a630ebd9): Reproduced on a SQLite supply-chain dataset; the oracle is each table aggregated separately.
changed the title [-][P0] Run-metric fanout and stale-lineage filtering (OWNER-GSF-RUN-METRIC-FANOUT-006)[/-][+][P0] Run-metric fanout across detail joins (OWNER-GSF-RUN-METRIC-FANOUT-006)[/+]on Sep 14, 2026
changed the title [-][P0] Run-metric fanout across detail joins (OWNER-GSF-RUN-METRIC-FANOUT-006)[/-][+][P0] Run-metric fanout across detail joins[/+]on Sep 14, 2026
@tomer-levin-nv@liavnave, flagging this as unowned. It's reproduced on main (details in the issue): "For each supplier, what is the total sourcing share percentage and the total quantity shipped?" returned 107,994,000 and 573,044,580 instead of 6,000 and 9,550,743. Nothing in flight covers it; I checked the open ReAct PRs and branches. The smallest fix I can see is a deterministic check in submit that flags an aggregate over a measure joined through a non-unique key, and returns it to the model. It sits in code you're actively changing, so who should own it? Happy to supply the fixture as a regression test.
Activity
[-][P0] Run-metric fanout and stale-lineage filtering (OWNER-GSF-RUN-METRIC-FANOUT-006)[/-][+][P0] Run-metric fanout across detail joins (OWNER-GSF-RUN-METRIC-FANOUT-006)[/+][-][P0] Run-metric fanout across detail joins (OWNER-GSF-RUN-METRIC-FANOUT-006)[/-][+][P0] Run-metric fanout across detail joins[/+][-][P0] Run-metric fanout across detail joins[/-][+]Run-metric fanout across detail joins[/+]cafzal commentedon Nov 18, 2026
@tomer-levin-nv @liavnave, flagging this as unowned. It's reproduced on
main(details in the issue): "For each supplier, what is the total sourcing share percentage and the total quantity shipped?" returned 107,994,000 and 573,044,580 instead of 6,000 and 9,550,743. Nothing in flight covers it; I checked the open ReAct PRs and branches. The smallest fix I can see is a deterministic check insubmitthat flags an aggregate over a measure joined through a non-unique key, and returns it to the model. It sits in code you're actively changing, so who should own it? Happy to supply the fixture as a regression test.