Skip to content

Re-tune significance_weight: the salience term is underpowered #7

Description

@kevindechang

bench/ (regime D, experiment F) shows the retrieval scorer's salience term cannot express an informative salience signal. Significance enters our score at weight 0.03 — a 3% ceiling — while it enters Park et al.'s additive form at 1.0.

Sweeping that single parameter, with the multiplicative composition untouched, lifts regime-D MRR from 0.858 to 0.960, against the additive baseline's 0.968. Meanwhile the recency multiplier ranges over 0.4–1.0 and can inject far more noise than salience can remove.

Why this is not a one-line change: significance_weight is part of the pinned cross-runtime contract. Changing the default updates expected values in shared/golden/kernel_golden.json and requires the workflow in docs/parity-contract.md — a reviewed fixture change passing in both runtimes in the same PR, plus a version decision.

What a good change looks like:

  1. Sweep the weight on more than one regime, not just D, and report where it hurts.
  2. Decide whether rehearsal_weight and rehearsal_cap need the same treatment.
  3. Update the fixture, both runtimes, docs/foundations.md section 6, and bench/RESULTS.md.

Good first issue for someone comfortable reading a benchmark; the analysis is already written down.

Metadata

Metadata

Assignees

No one assigned

    Labels

    good first issueGood for newcomersresearchNeeds an experiment or an evidence review

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions