Resolving Contradictory Source Facts

Contradictory sources should produce an explicit unresolved claim until their definitions and applicability have been compared. We would resolve the disagreement at the field level. Selecting the newest document or averaging two reported numbers can conceal a difference in scope rather than settle it.

This is an ingestion decision within our state and memory framework. It becomes an evaluation case through trace feedback, where the source resolution can be inspected before judging the model's later action.

Three balances that look comparable

Consider a hypothetical account tool reporting 1,000 cash units, a portfolio tool reporting 920 available units, and a research summary reporting 950. At first glance, these are conflicting balances. The account tool may include 80 reserved for an open order, making its 1,000 total compatible with 920 available. The research summary may be an older estimate whose derivation is missing.

We would first normalize the field definitions. Total cash, available cash, and estimated cash become separate typed claims. Account identity and valuation currency also need to agree before comparison. A dollar-denominated account balance and a collateral-unit amount can be numerically similar while describing different objects.

Once scopes align, a true contradiction remains actionable. If two current tools both claim available cash for the same account at the same boundary and return 920 and 940, their difference is 20 units. The runtime should preserve both observations and request the evidence needed to resolve them. A mean of 930 creates a third balance with no source.

Resolution leaves a receipt

The proposed conflict record includes the claim key, compared source records, normalized definitions, and outcome. Outcomes can be compatible after normalization, resolved by a declared authoritative source, or unresolved. Resolution should say which evidence changed the status, so a later reader can distinguish an investigation from an arbitrary preference.

The W3C provenance model offers concepts for linking entities and derivations. We would use those links to retain the original tool payloads and the normalization that produced the comparison. The engineering choice of source precedence still belongs to the application contract.

A fixture suite should include a compatible reservation case, a currency mismatch, and a genuine same-field conflict. The first should preserve both legitimate views. The second should require conversion evidence or separate presentation. The third should remain unresolved when the prescribed source is unavailable. This prevents a conflict detector from appearing effective simply by rejecting every multi-source input.

We would evaluate resolution correctness before measuring action behavior. A no-trade response can coexist with a mistaken factual synthesis; a trade can be valid when the apparent contradiction was merely a difference between total and available cash. The final owner-facing explanation should name the affected field and the unresolved amount. That precision gives a research agent a concrete next question and keeps disagreement from becoming vague doubt about the entire portfolio.