Rubric Memory

Rubric memory is a local scorecard knowledge-base convention for giving Plexus agents and reports relevant policy history, scripts, meeting notes, emails, and other source material while keeping the active ScoreVersion rubric as the official authority.

Authority Model

The official policy authority for score-level analysis is the active score version, usually the champion version referenced by Score.championVersionId. Corpus material can explain interpretation, history, rationale, exceptions, or gaps, but it cannot silently override the official rubric.

Plexus is moving toward rubric as the domain term. Some storage fields are still named guidelines, and the rubric-memory boundary translates those fields into rubric terminology.

Folder Structure

Knowledge-base folders live beside the pulled score artifacts. They use the same sanitized score file names as local score YAML and Markdown files, so the relationship between a score and its knowledge base is visible in the file tree.

<SCORECARD_CACHE_DIR>/
  <Scorecard Name>/
    scorecard.knowledge-base/
    <Prefix>.knowledge-base/
    <Score Name Stem>.knowledge-base/
    <Score Name Stem>.yaml
    <Score Name Stem>.md

For example, with SCORECARD_CACHE_DIR=dashboard/scorecards, the score Example Score in Example Scorecarduses this score-level knowledge-base folder:

dashboard/scorecards/Example Scorecard/Example Score.knowledge-base/

Scope Levels

A retrieval request combines the canonical roots that apply to the score:

  • Score: exact score-level material from <Score Name Stem>.knowledge-base/.
  • Prefix: shared material from matching folders such as Information Accuracy.knowledge-base/.
  • Scorecard: broad material from scorecard.knowledge-base/.

Ranking prefers score-specific evidence first, prefix evidence second, and scorecard evidence third. Prefix folders are optional overlays, not fallback search locations.

Prefix Knowledge Bases

Prefix knowledge bases support groups of related scores. This is useful when a scorecard contains a composite concept and many sub-element scores, such as Information Accuracy: High-Pressure Sales Tactics.

Example Scorecard/
  Information Accuracy.knowledge-base/
  Information Accuracy- High-Pressure Sales Tactics.yaml
  Information Accuracy- High-Pressure Sales Tactics.knowledge-base/
  Information Accuracy (Composite).yaml

A prefix folder matches when its stem is a leading prefix of the sanitized score name at a clear boundary, such as a space, hyphen, or parenthesis. Exact score folders and scorecard.knowledge-base are excluded from prefix matching to avoid duplicates.

Date Folders

Temporal source metadata comes from folder names. Put raw files under the date of the meeting, email, chat, or document. Use unknown-date when no date is known.

Example Score.knowledge-base/
  2026-04-24/
    meeting-notes.md
    client-email.txt
  unknown-date/
    pasted-notes-without-date.md

Plexus infers source_timestamp from the nearest ancestor folder matching YYYY-MM-DD, including nested paths such as 2026-04-24/client/source.md. Unknown-date files remain retrievable, but they do not contribute to chronological history ordering.

Prepared Corpora

Raw knowledge-base folders are the source of truth and are never rewritten. At runtime, Plexus prepares a working Biblicus corpus under ignored local storage:

tmp/rubric-memory/prepared/<stable-cache-key>/

The prepared cache stores copied source files, sidecar metadata, source roots, file counts, retriever id, schema version, fingerprint, and prepared timestamp. The same preparation path runs just in time during retrieval and through the prewarm command:

SCORECARD_CACHE_DIR=dashboard/scorecards plexus rubric-memory prewarm \
  --scorecard "Example Scorecard" \
  --score "Example Score"

Two Products

Retrieval-Only Citation Context

This is input to an LLM call. Plexus retrieves official rubric authority and relevant corpus snippets, assigns citation IDs, separates chronological memory from relevance-ranked evidence, and injects the context into an existing agent or report prompt. It does not add another synthesis agent.

Rubric Evidence Pack

This is an optional structured analysis output. It uses retrieval plus a Tactus synthesis step to produce a deeper explanation with rubric reading, evidence classification, history of change, likely disagreement reason, confidence, and open questions. It is appropriate for drill-through analysis, not default high-volume per-item report voting.

Using Evidence Responsibly

Agents should cite rubric-memory evidence when they make policy claims. If corpus evidence answers a question but the rubric is unclear, the right SME question is not "What is the policy?" It is "Should the rubric be updated to explicitly say this?"

Sparse, conflicting, or undated evidence should lower confidence and produce open questions instead of confident policy conclusions.

Next Steps