hired.alignment.elicitation

The honesty loop: rank unknowns into the most informative questions first.

This module is deterministic scaffolding around a judgment the analyst makes. It computes which UNKNOWN requirements are worth asking about, in what order (by expected information gain ≈ profile-uncertainty × requirement-criticality), and whether the analysis has become decision-stable enough to stop.

See misc/docs/DESIGN.md §8.

Functions

info_gain(record)	Expected information gain of resolving this requirement, wrt the verdict.
is_decision_relevant(record)	A requirement is worth asking about only if resolving it could move the verdict — i.e. it's not a low-value-add nice-to-have we can ignore.
is_decision_stable(records)	True when no remaining open unknown could flip the verdict.
rank_clarifications(records, *[, ...])	Produce clarifying questions for askable records, highest info-gain first.
should_ask(record)	Whether to route this record to a clarifying question.

hired.alignment.elicitation.info_gain(record)

Expected information gain of resolving this requirement, wrt the verdict.

Heuristic: uncertainty × criticality. An impact override (1-5) scales criticality when the analyst has a sharper view than the requirement class.

Return type:

float

hired.alignment.elicitation.is_decision_relevant(record)

A requirement is worth asking about only if resolving it could move the verdict — i.e. it’s not a low-value-add nice-to-have we can ignore.

Return type:

bool

hired.alignment.elicitation.is_decision_stable(records)

True when no remaining open unknown could flip the verdict.

Concretely: there are no askable (decision-relevant, open, unknown) requirements left. This is the sharpest stop criterion from the research.

Return type:

bool

hired.alignment.elicitation.rank_clarifications(records, *, max_questions=None, min_info_gain=0.0)

Produce clarifying questions for askable records, highest info-gain first.

The question text is left to the analyst; this returns the prioritized skeleton (requirement, reason, info_gain) so the caller can fill question.

Return type:

list[Clarification]

hired.alignment.elicitation.should_ask(record)

Whether to route this record to a clarifying question.

Ask only when it’s UNKNOWN on an OPEN field and decision-relevant — the “don’t assume absence from silence” rule. CONTRADICTED facts and closed-field unknowns are real gaps, not questions.

Return type:

bool