Design a small evidence-based recall evaluation #98

Design deliverable

Create a small non-sensitive scenario set: resume a task, recall a decision, detect superseded guidance, avoid duplicates and answer a project-specific question. Compare current recall to proposed lexical relevance + validity + project scope + linked evidence ranking within the same token budget.

Measures

Useful evidence retrieved, missing essential evidence, unrelated content, token count and latency/cost. Include empty/small projects, multiple scopes, expired memories, permissions and query synonyms.

Acceptance

• Baseline, expected evidence and pass/fail criteria stated before choosing improvements.
• Graph context expansion has a bounded budget and preserves authorization.
• Embeddings remain a candidate only if measured failures justify their cost.
• No telemetry or paid external service activated by this task.

Dependencies

Design only under #77; do not claim results before running an explicitly scoped evaluation.