Skip to content

[Workstream M] PR 1 of 2: Score-margin gate #182

Description

@zaxbysauce

Part of #100 (EPIC). Workstream M. One issue = one PR.
Source: zmem deep audit 2026-09-10 (https://claude.ai/code/artifact/39012f75-25f0-441b-978f-cc90a3bd3a9d) and issue plan v3 (https://claude.ai/code/artifact/59e74242-17da-4416-9b00-0b0cf347837c); dependency DAG computed over all open PRs (edge kinds: Depends on = merged artifact, Schema after = migration lane, Gated on = #100 measurement/decision gate).

Labels: enhancement, severity:medium. Depends on: none. Gated on: G-3 (#155). Blocks: none.

Design. After the selective gate: if (top − second) / top < ZMEM_INJECT_MARGIN (default 0, i.e. off; recommended 0.05) and neither row is decision/constraint, keep only the top row; log margin=<value>; explain verdict margin_pruned.

Acceptance criteria. (a) Fixture scores 0.80/0.79/0.60 with margin 0.05 → one row; margin 0 → three. (b) constraint at 0.79 kept. (c) tests/test_eval_inject_runner.py negative controls unchanged (committed expected JSON equal); eval_replay --compare-baseline eval/baseline-replay.json --fail-under miss_delta=0 exits 0 with margin 0.05.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestseverity:mediumMedium-severity finding from codebase review

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions