Skip to content

feat(quotient): plumb unresolved_ratio into quotient CALLS-layer discount #149

Description

@zaebee

Why

The quotient graph (#144, build_quotient) drops edges whose endpoints fall outside every domain — which includes all raw_call:* targets. Consequence: the quotient fingerprint's unresolved_ratio is always 0, so the v2 CALLS-layer confidence discount (disc = clip(1 − unresolved_ratio)) is always 1.0 at k=1. The quotient's CALLS census looks fully trustworthy even when the underlying module-level resolution is poor.

Deliberately deferred in the Part B plan (observe-only milestone — recorded in src/cgis/query/quotient.py docstring and tests/self_parsing/test_drift.py).

What

Propagate a meaningful unresolved signal to k=1. Options to evaluate:

  • Aggregate the member domains' unresolved ratios (edge-count-weighted mean) and stamp it on the quotient fingerprint.
  • Count dropped raw_call: edges per source domain during build_quotient and derive the ratio from what was actually discarded.

Gate

Must land before project_level is switched to enforce: true — enforcing an undiscounted CALLS layer would overtrust noisy data. Sequencing: after the burn-in decision in #146.

🤖 Generated with Claude Code

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions