Level 3 · Responsibility
Uncertainty, model requests, and human decisions
Weak evidence does not force one path. A decision may be accepted, preserved as unresolved, sent for one bounded model opinion, or reserved for human judgment according to impact and policy.
Decision routing
Four possible outcomes
evidence + alternatives + impact
if policy allows
if required
The lower branch is not a mandatory sequence. A low-impact ambiguity may remain unresolved; a high-impact contradiction may go directly to a person.
Language model · planned
One constrained question, not document rewriting
A request would contain the relevant source fragments or crop, limited neighboring context, a precise question, and an allowlist of answers including unresolved. The response must identify the evidence used and satisfy a schema before it can influence resolution.
- Store provider and model version, request and response hashes, resource use, alternatives, and final disposition.
- Treat self-reported model confidence as advice, not a measured probability.
- Bound retries and context expansion; budget exhaustion creates a review item.
- Retain responses for deterministic resolver replay; measure fresh-call variance separately.
Human review
Corrections are versioned semantic decisions
A reviewer sees the source location, proposed value, supporting observations, alternatives, reason, method, uncertainty, impact, and review state. A correction updates that semantic decision and regenerates its views. It does not silently replace the source observation.
Human review may occur after conversion or, in a future interactive system, while unrelated pages continue. High-impact source contradictions, ambiguous reference expectations, and release acceptance remain human responsibilities.
Current implementation