SEEDING STRATEGY

Mobius Encyclopedia Seeding Strategy¶

This playbook explains how the Librarian cron job chooses questions, how frequently it should run, and how we keep the encyclopedia clean as it grows.

Seed Sources (Curriculum Pillars)¶

1. Civics & History — Haymarket, suffrage, civil rights, constitutional basics.
2. Science & Environment — climate fundamentals, vaccines, energy mix, LUCA, water cycle.
3. Ethics & AI — Henrietta Lacks, consent, provenance, civic OS, democratic AI.
4. Math & Numeracy — Arabic numerals, base-10, compound interest, inflation literacy.
5. Culture & Language — multilingual education, anti-intellectualism counterpoints, minority viewpoints.

Seeds can be stored as { question, topics[] } objects. The first version can be a static JSON list; future versions should track coverage metrics (how many entries per topic) and bias (ensure minority/alternative perspectives are represented).

Schedule Examples¶

The script infra/cron/encyclopedia_seed.sh reads ENCYCLOPEDIA_API_URL and ENCYCLOPEDIA_CRON_SECRET, posts to /v1/encyclopedia/seed, and logs the new entry.id + entry.question.

Guardrails¶

1. GI threshold filter (v2) — store everything now, but flag giScore < 0.93 for review instead of serving publicly.
2. Topic coverage — avoid over-indexing on a single domain. Track counts per topic and favor underrepresented ones.
3. Bias & plurality — include prompts that explicitly ask for multiple perspectives (“what are the main disagreements about X?”) and minority voices.
4. Review loop — low-GI entries default to pending_review. Admin tools let humans promote to approved, add commentary, or retire stale answers.

Flow Recap¶

[CRON] → POST /v1/encyclopedia/seed
   → pickSeedQuestion()
   → brokerClient.deliberate(question)
      → Sentinels debate
      → ECHO.reinforce(a1, a2, question)
   → encyclopediaRepo.save(entry)
   → ledgerClient.attestEncyclopedia(entry)
   → entry available for GET /v1/encyclopedia queries

The Librarian habit is simple: - Good questions (high GI, good sources) get reinforced, re-used, and maybe “canonized”. - Weak questions remain flagged until humans intervene. - Every failure is a signal: it tells humans “please add better context here,” aligning with “If an intelligence can reflect and see its flaws, then it knows it can proceed.”