Skip to content

Learning-Goal Book and Evidence-Review Pipeline

Generated artifact. Do not edit manually.

Generated by: scripts/build_compendia.py Regenerate with: python3 scripts/build_compendia.py --compendium skill-graph-architecture --web-and-tex Source of truth: docs/concept/skill-graph/architecture-source/

Language / Sprache: English (EN) · Deutsch (DE) | PDF: English PDF · Deutsches PDF

Compendium: §1 General Goal System · §2 Node Types · §3 View Projection · §4 Source & Resource Links · §5 Content Integration · §6 Source Rationales · §7 Goal Visualizations · §8 Learning-Goal Book · §9 Gymnasium Rollout · §10 MEM/FWU Plan · §11 Dual Package Releases

8. Learning-Goal Book and Evidence-Review Pipeline

The learning-goal book (Lernzielbuch) and positive understanding-evidence pipeline make every curricular-atomic goal inspectable by humans and multi-AI reviewers without bloating learner-facing goal descriptions (teacher timeline planning on top of the GoalBook is specified in Goal-Book Time Planning).

8.1. Shared BookModel, One-Goal-Per-Page Atlas, and Evidence Profiles V2

One deterministic BookModel (contracts/goal-book/v1/goal-book-model-1.1.schema.json) drives both the read-only /lernzielbuch WebGUI and the A4 portrait PDF:

  • Current nationwide review editions (30 Aug 2026): 792 unique curricular-atomic pages for Gymnasium Mathematics and 461 pages for Physics across all 16 German states (publicationMode: review, carrying QA-ledger candidate images without claiming human image approval; coach runtime and Mastery writes remain unchanged).
  • Fourteen binding invariants: Concise canonical title/description; separate fingerprint-bound V2 evidence profiles (contracts/goal-evidence/v2/goal-evidence-profile.schema.json capturing expectations, coverageExpectations, variationAxes, and applicationCaseBriefs; V1 retained as archived audit history); one physical page per ordinary curricular-atomic target goal displaying full Lernziel-ID, chapter path (canonical-composition-view-v1 + Bundeslandspezifische Ergänzungen with gap-free treeOrder), stable Kahn topological order from requires, clickable named PDF destinations (goal-<uuid>), fail-closed overflow/legibility checks, incremental Skia PDF 1.4 bookmark outlines, state-primary applicability tuples (Bundesland -> [{stage, durationModel, courseProfile}] — never Cartesian G8/G9 cross-products), and on-demand /lernzielbuch/<bookId>.original-sources.json supplements.
  • Application-based recognition review (4 Nov 2026): Assessments supporting global competence recognition (Assessment-Gated Completion) require static review against compensating-point bypasses before release.

8.2. Multi-AI Review Loop, Evidence Gates, and Moderated Feedback Custody

Quality iteration combines a fast blind multi-AI loop with strict generalization gates and a slow, privacy-minimized public feedback channel:

| Scale | Levels & Meaning | | –- | –- | | Evidence (E0–E5 + NORMATIVE) | E0 unanchored opinion \(\to\) E1 anchored observation \(\to\) E2 reproducible minimal counterexample \(\to\) E3 independent multi-model/prompt triangulation \(\to\) E4 controlled field evidence \(\to\) E5 replicated cohort evidence. | | Claim scope (G0–G4) | G0 single case \(\to\) G1 one goal \(\to\) G2 goal archetype (requires \(\ge\) E3 across \(\ge\) 2 chapters) \(\to\) G3 subject/stage (stratified sample + replication) \(\to\) G4 global Coach/Mastery rule (Math + Physics + field canary). | {cols="30%,70%"}

  • Two-round V2 campaign: Calibrated on 16 Mathematics goals (5–10, E–Q4, pilot goal 8dd9f210-2683-5902-acab-e3be22725232) before the 792-goal batch: (1) freeze Atlas V1, (2) blind Codex V1 review (keep | revise | split_review | block), (3) deterministic check + goal-by-goal human adjudication, (4) controlled bilingual V2 authoring + downstream lane re-runs, (5) blind external-provider V2 review, (6) dissent-preserving synthesis. Neither AI agreement nor silence ever auto-approves canonical changes.
  • Two-phase /lernziel-feedback custody transfer: Default-off in production behind SKILLPILOT_GOAL_FEEDBACK_OPERATOR_TOKEN; accepts privacy-minimized V2 envelopes without learner/SkillPilot IDs, contact fields, or attachments. POST /api/operations/goal-feedback/v1/export-batches exports deterministic batches + server-derived publication snapshots into owner-only tmp/goal-feedback/inbox/ (trusted-context.jsonl separated from untrusted-feedback.jsonl); only after local byte/digest verification does idempotent DELETE .../{exportId} with If-Match clear live rows while retaining a content-free batch receipt.
npm --prefix app run quality:goal-evidence:check
npm --prefix app run build:goal-book-model -- \
  scripts/config/goal-books/de-gym-math-national-atlas.json
npm --prefix app run build:goal-book-original-sources
npm --prefix app run check:goal-book-original-sources