Learning-Goal Book and Evidence-Review Pipeline

Status: normative target concept
Initial scope: canonical German Gymnasium mathematics and physics
Audience: curriculum authors, didactic reviewers, teachers, AI-review operators, and runtime implementers

Implementation status (2026-08-10): Phase 0 and the nationwide mathematics review-atlas slice are implemented. Closed contracts exist for evidence profiles, review configuration, BookModel, render manifests, AI-review bundles, AI runs, findings, and the public-feedback envelope. The deterministic renderer has completed a real 754-page review edition covering the deduplicated union of curricular-atomic Gymnasium mathematics targets from all 16 German states, Sekundarstufe I and II. Every goal has exactly one physical PDF page. Exact applicability remains state-first: G8/G9 and GK/LK are displayed only in their bound jurisdiction, stage, and profile tuples and are never inferred as global goal properties. The same digest-bound reviewer edition is shipped by the regular repository build as a coreless, read-only /lernzielbuch preview with chapter navigation, search, stable goal deep links, state-bound scope filters, and PDF download. The reserved /lernziel-feedback target is currently a clearly labelled, non-writing pilot placeholder. The representation-choice profile is deliberately still an ai_candidate with needs_human_review; no independent multi-provider review or human approval is claimed. Public feedback links and their version-bound privacy-minimized envelope are prepared, but the moderated production intake is a later phase. Coach runtime and Mastery behavior remain unchanged.

1. Purpose

SkillPilot needs a review and publication path that makes every learning goal easy to inspect without overloading its learner-facing wording. The path has two products built from the same canonical state:

1. a navigable learning-goal book, published as HTML and PDF; and
2. a structured, fingerprint-bound evidence profile for every reviewed atomic goal.

The book makes the skill graph legible to people and review models. The normative V2 evidence profile states the goal-specific essential understanding, independently observable performance, coverage expectations, meaningful variation axes, and application cases for transfer. Neither product replaces the canonical skill graph.

The first rollout uses a fast, reproducible multi-AI review loop as the primary development engine. A public feedback channel is installed early but is treated as a slow signal and challenge channel, not as a release dependency or a source of automatic curriculum mutations.

2. Binding decisions

The following decisions are invariants of this concept:

1. Canonical goal title and description remain concise and learner-facing. Detailed didactic and assessment guidance belongs in a separate evidence profile.
2. Every evidence profile is bound to the exact semantic fingerprint of its goal and to the review inputs that could bias or invalidate it.
3. HTML and PDF are rendered from one shared, deterministic BookModel. They must never build their own ordering or relationship projections.
4. In the initial edition, every included ordinary technical atomic target goal has exactly one goal page. A goal never shares a page with another goal and never continues onto a second page. Cluster, memory, orientation, practice/assessment, exam, and runtime-support nodes need separate later edition contracts and are not silently mixed into this book.
5. Every goal page shows the full canonical goal identifier under the unambiguous label Lernziel-ID.
6. contains supplies chapter and topic context. requires supplies the hard ordering constraints.
7. Direct prerequisites and direct reverse prerequisites are clickable. PDF navigation uses stable named destinations, not guessed page numbers.
8. A page that does not fit its fixed template makes the build fail. Text may not be clipped, silently omitted, continued on another page, or shrunk below the approved minimum sizes.
9. Books, review bundles, logs, links, and feedback envelopes contain no learner state and no permanent learner or SkillPilot ID. A Lernziel-ID identifies a public curriculum goal, never a person.
10. AI findings and public feedback can create review candidates or blockers, but cannot modify canonical goals, approved profiles, or runtime behavior automatically.
11. A finding may be generalized only through the evidence- and scope gates in this document. Repetition of an opinion is not independent evidence.
12. Evidence profiles do not enter the learning-coach runtime or Mastery write path during the book and review-pipeline rollout. Runtime enforcement is a later, separately gated phase.
13. A nationwide subject atlas is a deduplicated union of explicitly bound, reviewed learner-facing projections. A canonical goal receives exactly one page regardless of how many jurisdictions, duration models, stages, or course profiles include it.
14. Applicability is represented as exact, state-primary tuples. G8 and G9 never form an independent nationwide label or filter; they are meaningful only together with the Bundesland whose reviewed policy supplies that duration model. The generator must not create Cartesian combinations from independently collected states, stages, duration models, and course profiles.

3. Source-of-truth boundaries

3.1 Canonical graph

The canonical landscape and its separately reviewed semantic-kind ledger remain authoritative for:

• the stable goal ID;
• title and concise learner-facing description;
• contains and requires;
• semantic kind and other canonical goal metadata; and
• canonical resource links, including reviewed visualizations.

The book generator must not repair, reinterpret, or enrich missing graph semantics while rendering. Broken references, cycles, duplicate projected occurrences, or unresolved identifiers fail before publication.

For a nationwide atlas, learner-facing applicability has a separate source of truth: membership as target in the explicitly enumerated effective composition views, combined with the reviewed Gymnasium duration-model policy. The broad canonical goal.applicability field is useful provenance and a conservative outer bound, but it is not precise enough to print or filter the final Bundesland/G8/G9/GK/LK matrix. Changes to any bound projection or to the duration policy make the generated atlas stale.

3.2 Positive understanding-evidence profiles

Evidence profiles are separate curriculum-quality artifacts. They explain the competence more precisely without turning the ordinary goal description into a large prompt. A V2 profile may contain:

• content-specific statements of essential understanding;
• independently observable learner performance;
• required coverage across those expectations;
• meaningful variation axes;
• application-case briefs for changed contexts and representations;
• teaching-case blueprints;
• transfer-case blueprints; and
• the minimum number and independence requirements of demonstrations.

The canonical description may be improved when a review demonstrates real ambiguity. It must nevertheless remain short enough to be understandable in the normal learner UI. Detailed examples, positive evidence requirements, variation designs, and review arguments stay in the evidence profile.

3.3 Generated publications

HTML, PDF, manifests, indexes, and AI-review bundles are generated views over a versioned canonical graph and the current review artifacts. BookModel, ordering, page ownership, relationships, and semantic manifests are deterministic. Browser-generated PDF bytes may contain tool metadata such as a creation timestamp; therefore the exact PDF is hash-bound per review bundle, while reproducibility is judged from the deterministic model and validated page/link parity rather than from a frozen PDF hash alone. Generated artifacts are not edited by hand and are not a second source of truth.

4. Evidence-profile contract

4.1 Archived V1 conceptual record

The initially implemented V1 record is retained unchanged as audit history. It is not reviewer guidance for new campaigns. Its conceptual shape and closed implemented contract are contracts/goal-evidence/v1/goal-evidence-profile.schema.json; bounded operational fields preserve the historical semantics:

{
  "schemaVersion": 1,
  "reviewId": "<review scope ID>",
  "ruleVersion": "goal-evidence-v1",
  "landscapeId": "<canonical landscape ID>",
  "goalId": "<full canonical goal ID>",
  "goalFingerprint": "sha256:<normalized semantic goal fields>",
  "reviewInputFingerprint": "sha256:<goal, relations and reviewed resources>",
  "profileFingerprint": "sha256:<normalized evidence profile>",
  "status": "needs_human_review|approved|rejected",
  "reviewAuthority": "ai_candidate|human",
  "evidenceLevel": "E0|E1|E2|E3|E4|E5",
  "maximumClaimScope": "G0|G1|G2|G3|G4",
  "reviewRunIds": [],
  "dissent": [],
  "profile": {
    "archetype": "concept|procedure|representation|modeling|proof|experiment|data",
    "facets": [
      {
        "id": "<stable profile-local ID>",
        "criterionDe": "<observable aspect of understanding>",
        "criterionEn": "<same criterion in English>"
      }
    ],
    "coverageRequirements": {
      "allOf": ["<facet ID>"],
      "anyOf": [["<facet ID>", "<alternative facet ID>"]],
      "minimumIndependentChecks": 2,
      "requireChangedCase": true,
      "requireCueFreeTransfer": true
    },
    "misconceptions": [],
    "variationAxes": [],
    "nonEvidence": [],
    "outOfScope": [],
    "contrastCaseBriefs": []
  }
}

The archetype vocabulary is deliberately small. A goal may declare one primary archetype and bounded secondary characteristics, but the pipeline must not grow a subject ontology merely to classify review prompts.

4.1.1 Positive understanding-evidence contract V2

The V1 profile and its representation-choice record are retained unchanged as an archived AI experiment. They are not the normative authoring target for new profiles. The closed V2 contract is contracts/goal-evidence/v2/goal-evidence-profile.schema.json.

V2 states the expected competence positively and content-specifically through expectations, coverageExpectations, variationAxes, and applicationCaseBriefs. Each expectation pairs essential understanding with observable learner performance. Application cases state the task demand, expected performance, and understanding focus for a fresh variation. The established goal-evidence-v1 goal-fingerprint rule remains unchanged, while a separate V2 profile-rule version binds the new profile semantics and review criteria. V2 is initially a review contract only; GoalBook and Coach runtime projection require their own later acceptance gate.

4.2 Fingerprint binding

goalFingerprint binds at least the normalized semantic fields used by the existing goal-quality lanes: ID, title, description, semantic kind, and other declared semantic goal fields. The fingerprint algorithm and normalization version are explicit in the schema or manifest.

reviewInputFingerprint additionally binds all inputs that can invalidate the didactic judgment:

• the goal fingerprint;
• direct requires and contains relations relevant to the review;
• active visualization and other reviewed learner-visible resource digests;
• the evidence-profile schema version; and
• the review-criteria version.

Any mismatch makes the prior profile stale. A stale profile remains available as audit history but is neither published as current reviewer guidance nor eligible for future runtime projection. A title correction, changed image, or changed prerequisite must never inherit an approval silently.

4.3 Status and authority

• needs_human_review: generated or authored proposal. reviewRunIds, evidenceLevel, maximumClaimScope, and dissent distinguish an initial candidate from a proposal that has already received independent reviews.
• approved: a responsible release decision accepts the current fingerprint.
• rejected: the proposal is not suitable for publication or use.
• stale is a derived condition: a stored decision whose goal or review-input fingerprint no longer matches is ineligible even if its stored status says approved.

Generation models, reviewer models, and the synthesis step cannot assign approved. The responsible curriculum release process owns that transition. Disagreement is recorded; synthesis must not erase dissent to produce apparent consensus.

4.4 Proposed authoring location

The implementation should place subject-specific configuration, ledgers, and review run manifests below a dedicated quality lane, for example:

curricula/DE/Gymnasium/quality/goal-evidence/
  mathematics.config.json
  mathematics.review.jsonl
  physics.config.json
  physics.review.jsonl
  runs/<run-id>/...

The closed Phase-0 companion contracts are:

• contracts/goal-evidence/v1/goal-evidence-review-config.schema.json;
• contracts/goal-evidence/v1/goal-evidence-ai-run-manifest.schema.json;
• contracts/goal-evidence/v1/goal-evidence-finding.schema.json; and
• contracts/goal-evidence/v1/goal-public-feedback.schema.json.

Run payloads may be large. The implementation must deliberately classify which parts are durable review evidence, which are generated summaries, and which are temporary provider traces. Secrets, provider credentials, private prompts, and learner data are never committed.

5. Shared BookModel

5.1 One projection for HTML and PDF

The generator first creates a locale- and scope-specific BookModel validated against contracts/goal-book/v1/goal-book-model.schema.json. The HTML site and PDF renderer consume that exact serialized model. The model contains:

• book version and build fingerprint;
• canonical landscape and composition-view or source-manifest identifiers;
• ordered chapter records;
• exactly one record for every included goal;
• direct prerequisite and reverse-prerequisite links;
• resolved, approved public resources;
• current evidence-profile status and fingerprints; and
• deterministic destinations and canonical permalinks.

Parity validation compares the goal IDs, order, relationships, versions, and content fingerprints emitted by both formats. A renderer may change visual presentation, but never content selection or navigation semantics. The closed render manifests additionally bind the normalized feedback base URL used by the HTML/PDF links; HTML and PDF for one publication may not silently point to different feedback targets.

5.2 Scope and inclusion

Books are built for an explicit reviewed composition view or for an explicit, closed source manifest and locale. The first edition includes only target goals that are ordinary technical atomic competencies. The same projection roles used by the learner-facing graph remain authoritative:

• eligible target goals are normal book goals;
• prerequisiteOnly goals may be included in a clearly marked prerequisite appendix or linked to their canonical online page, according to the versioned book configuration; and
• excluded or out-of-scope goals must not appear merely because a traversal can reach them.

Memory, orientation, practice/assessment, exam, and runtime-support nodes are excluded by semantic kind in the initial edition. Encountering one never causes the generator to invent a description or reinterpret it as an ordinary competence. A later edition may include such nodes only through an explicit versioned page contract suited to their different semantics.

Within a book, each included goal ID occurs exactly once. A goal referenced from multiple chapters receives one canonical placement chosen by the reviewed composition view or, for a union edition, by its explicitly versioned neutral navigation-ownership contract. Other occurrences are links, never duplicate pages.

The nationwide Gymnasium mathematics atlas is a union edition. Its source manifest enumerates every admitted Bundesland/stage/course projection and binds the reviewed G8/G9 policy by path and digest; filesystem discovery or a broad glob is not a publication contract. Its target set must equal the complete union of current curricular-atomic targets across the effective projections. At the current bound revision this means 754 unique goal pages. Missing, additional, or duplicate IDs fail the build.

Each nationwide page carries groups of the form:

Bundesland -> [{ stage, durationModel, courseProfile }, ...]

Only exact authored and policy-validated tuples are legal. A duration-neutral projection with reviewed applicability to both G8 and G9 produces two exact tuples with the same goal set; a single-duration source produces only its one reviewed duration; and a genuinely duration-specific projection may produce different goal sets. Sekundarstufe-II GK/LK remains coupled to the respective state and stage. null means that the bound policy makes no duration/profile distinction for that tuple; it is not a wildcard and must not match an explicit G8, G9, GK, or LK filter.

5.3 Chapters from contains

contains supplies the topic path and chapter nesting. It is not treated as a prerequisite relation. In the initial atomic-goal edition, cluster and structure nodes supply chapter labels, breadcrumbs, bookmarks, and ordering preferences; they do not receive goal pages.

Front matter, contents, indexes, and publication notes may use additional non-goal pages. The one-page invariant concerns goal pages: every included goal has one and only one complete page, and every goal page belongs to exactly one goal.

When a union edition consumes several composition views, view-local runtime IDs must not create repeated roots such as several indistinguishable Mathematik chapters. The nationwide mathematics atlas binds common-topic-suffix-v1: it compares every admitted placement of a goal, removes state-, stage-, duration-, and course-specific wrappers that are not common to all of them, and retains only their exact shared topic suffix below a single neutral Mathematik root. If no shared topic path exists, the goal stays at that root rather than inheriting a misleading path from whichever source file happens to sort first. Source-specific supplement chapters remain distinguishable only when their complete labels are genuinely shared.

5.4 Stable topological order from requires

The linear reading order is a stable topological sort over the included goals:

1. every included direct prerequisite precedes the goal that requires it;
2. chapter membership derived from contains supplies the preferred grouping;
3. reviewed composition order supplies the first tie-breaker;
4. stable canonical goal ID supplies the final tie-breaker.

A stable Kahn-style sort is suitable. Chapter grouping may not violate a requires edge. If keeping a chapter contiguous would do so, the prerequisite constraint wins and the book records the affected cross-chapter relation.

A requires cycle, missing required goal, or nondeterministic tie fails the build. Re-running the generator over byte-identical inputs must produce an identical BookModel and semantic manifest.

5.5 Goal-page contract

Every goal page presents, in a stable layout:

1. subject, stage, and chapter path;
2. learning-goal title;
3. the full canonical Lernziel-ID, without abbreviation;
4. the reviewed visualization when one is available, including its review status in a reviewer edition;
5. the concise canonical description;
6. direct prerequisites (requires) with title and full Lernziel-ID;
7. direct reverse prerequisites (goals that directly require this goal), with title and full Lernziel-ID;
8. book version, goal fingerprint, and page number; and
9. a version-bound feedback permalink and optional QR code.

Nationwide editions additionally show a compact curricular-applicability summary. The web atlas exposes the complete exact matrix and filters in the order Bundesland, Stufe, then the still-valid G8/G9 and GK/LK choices. G8/G9 controls remain unavailable until a Bundesland is selected. The PDF may group states only when their complete ordered scope tuple lists are identical; it must link to the full online matrix rather than truncate silently.

The public learner/teacher edition does not need to expose the complete evidence profile. A reviewer edition may add bounded essential-understanding, observable-performance, coverage, variation, and application-case summaries if they fit the same fixed page contract. Longer AI findings and adjudication records remain in the structured review bundle.

5.6 Named destinations and links

Each included goal receives a deterministic named destination derived from its full canonical goal ID, for example:

goal-8dd9f210-2683-5902-acab-e3be22725232

The generator must escape or encode IDs reversibly if a PDF library restricts destination names. Link labels continue to display the full unmodified ID.

• An in-book prerequisite links to the goal's named destination.
• An in-book reverse prerequisite uses the same mechanism.
• An out-of-book prerequisite links to the versioned online atlas when a canonical public target exists and is marked explicitly as external.
• Missing, ambiguous, or unsafe targets fail the build; they are never rendered as apparently valid dead links.

PDF bookmarks include chapter hierarchy and every goal title. A second index maps full Lernziel-IDs to pages. HTML URLs use stable goal-ID fragments or paths and preserve the same link graph.

5.7 Fail-closed layout

The page template has explicit bounds for title, image, description, relationships, footer, font sizes, and accessible link targets. The initial mathematics review profile uses A4 portrait with a fixed, full-width 16:9 visualization region. Relationship density may compact the relationship cards and their technical IDs, but it must never shrink the image. The visualization itself is not placed in a wider tinted container, and the page title makes a repeated image caption unnecessary.

Before release, the renderer validates at least:

• all required content is visible inside its assigned region;
• no element overlaps, clips, or overflows;
• no automatic font scaling crosses an approved minimum;
• every goal starts and ends on the same physical page;
• no physical page contains content from two goals;
• every named destination resolves to the expected page;
• every internal link resolves;
• full IDs and fingerprints are rendered without truncation; and
• HTML and PDF manifests have identical goal sets and order.

If a goal does not fit, the build reports its ID and the overflowing region. The remedy is an intentional template, content, edition, or source correction; it is never silent clipping, ellipsis, microscopic text, or an extra goal page.

The renderer has two explicit resource policies:

• a review publication may include a current QA-ledger-bound candidate image; its QA status remains machine-readable in BookModel and manifest, while the edition as a whole is visibly labelled as a review publication; and
• a public publication includes only a current publication-approved image.

The policy is part of the BookModel and manifest. A public renderer must never silently inherit the more permissive review policy.

An explicitly labelled, read-only reviewer atlas may be reachable on the public SkillPilot origin before the learner/teacher edition is approved. It must keep publicationMode: review, visibly mark the whole edition, retain each image's exact QA status in the bound machine contract, expose no write-capable feedback control, and make no claim of public approval. Repeating an image-level warning on every page is not required. This preview is a review surface, not the public publication. A later public edition is built separately with publicationMode: public and the stricter resource policy above.

Original visualization assets remain immutable. A PDF renderer may create deterministic print derivatives with a versioned maximum pixel size, encoding, and quality profile. The BookModel and release manifest bind both the original asset digest and the derivative profile/digest. HTML may use the approved public asset URL; a local PDF build may resolve only allowlisted repository assets and must make no network request. A renderer must not embed hundreds of unbounded original images into one monolithic HTML document, because that turns a valid book into an avoidable memory and file-size failure.

5.8 AI-review bundle

The PDF is an important review surface, but it is not the only input supplied to review models. Every review batch contains the same content in three aligned forms:

• the exact goal-page PDF or chapter PDF seen by human reviewers;
• normalized JSON from the BookModel and evidence-profile candidate; and
• concise Markdown rendered from that JSON.

This avoids making PDF extraction or OCR an accidental source of curriculum semantics. Every model finding cites the full goal ID and fingerprint, not only a page number. The pipeline verifies that the cited goal belongs to the exact review bundle before accepting the finding.

Goal-local bundles include the goal, its chapter path, direct prerequisites, direct reverse prerequisites, current resources, and the profile candidate. Chapter batches expose local consistency. Overlapping batches include boundary goals from neighboring chapters so reviewers can detect sequencing gaps and contradictions that a one-page review would miss. A final global manifest pass checks relationships and vocabulary across the complete subject without asking one model invocation to rewrite the entire book.

6. Fast independent multi-AI loop

This generic diagnostic loop remains available for non-V2 quality lanes and historical V1 auditability. It is not the prompt, criteria, or output model of the positive V2 understanding-evidence campaign in Section 9.4.1.

6.1 Role and cadence

The multi-AI loop runs in hours or days and is the primary early improvement loop. It is designed for breadth, adversarial testing, reproducibility, and rapid iteration before public feedback can accumulate.

The loop has distinct roles:

1. candidate author;
2. subject-matter reviewer;
3. didactic reviewer;
4. assessment and adversarial reviewer;
5. cross-goal sequencing and representation-cue reviewer;
6. disconfirming reviewer;
7. deterministic validation and replication runner;
8. synthesizer; and
9. responsible release decision.

The synthesizer groups findings and proposed changes but has no release authority.

6.2 Independence

The first review pass is blind: reviewers do not see other reviewers' answers. A normal review batch uses at least:

• three distinct review roles;
• two model families or providers where available; and
• two independently authored prompt families.

Five runs of one model with one prompt count as repeated sampling of one source, not five independent sources. Run manifests bind model and provider, model version when available, prompt and criteria fingerprints, role, generation parameters, context bundle fingerprint, time, and toolchain version.

Review order and option order are randomized where position bias is plausible. Named model identities are hidden during synthesis when that does not impede auditability. Material model, prompt, criteria, or context changes require a new calibration run.

6.3 Structured findings

Each finding records:

• exact goal ID and fingerprint;
• anchored observation;
• hypothesized mechanism;
• violated criterion;
• smallest reproducible counterexample;
• severity;
• evidence level and maximum claim scope;
• counterarguments or disconfirming evidence;
• proposed local change;
• possible side effects; and
• the evidence needed for broader generalization.

Reviewers must distinguish a defect in the canonical goal, visualization, evidence profile, book projection, coaching policy, model behavior, or runtime guard. The pipeline must not turn every observed coaching failure into a goal wording change.

6.4 No majority-vote release

Model votes are not a release criterion. A minority finding with a reproducible factual error, false-mastery path, privacy issue, or normative conflict can block the affected item. Conversely, many similar unanchored opinions do not establish a general rule. Release decisions cite claims, evidence, scope, and remaining dissent.

7. Evidence and generalization gates

7.1 Evidence levels

Level	Meaning
E0	Opinion or hypothesis without an anchored observation.
E1	Anchored observation in a specific goal, page, artifact, or dialogue.
E2	Reproducible defect with a minimal counterexample.
E3	Independently replicated or triangulated defect and mechanism.
E4	Field evidence from an appropriate controlled learner or teacher setting.
E5	Replicated field evidence across settings or cohorts.

Normative requirements such as privacy, source fidelity, accessibility, or provider policy are tagged NORMATIVE in addition to an evidence level. They may block immediately within their actual scope.

7.2 Claim scopes

Scope	Meaning
G0	One concrete dialogue, rendering, or generated case.
G1	One identified learning goal and its bound artifacts.
G2	A defined goal archetype or recurring mechanism.
G3	A subject, stage, or similarly bounded curriculum scope.
G4	Global SkillPilot coach, publication, or Mastery rule.

Every finding declares the maximum scope supported by its evidence. A proposed change may be narrower, never broader.

7.3 Generalization rules

1. One reproducible case may justify a G0 fix and can block or correct the affected G1 item.
2. A G2 rule requires at least E3: the mechanism must reproduce on multiple goals from at least two chapters and survive a disconfirming test.
3. A G3 rule requires a stratified sample across the relevant stages and goal archetypes, plus independent replication.
4. A G4 coach or Mastery rule requires evidence from both mathematics and physics, several goal archetypes, independent offline replication, and a controlled field or canary phase before general release.
5. An immediate safety, privacy, factual, or normative blocker stays as narrow as the demonstrated scope. Urgency does not make its mechanism universal.

The governing sentence is:

Ein einzelner Fall kann einen konkreten Defekt beweisen und deshalb einen Ablauf stoppen. Für eine allgemeine Regel braucht es zusätzlich einen belegten Mechanismus, unabhängige Replikation und einen zum behaupteten Geltungsbereich passenden Feldtest.

8. Slow public signal and challenge channel

8.1 Deliberately different time scale

The public HTML atlas and feedback control are installed early so teachers and other reviewers can cite exact goals. At the beginning, however, this channel is expected to produce sparse, delayed, and uneven feedback. The AI loop does not wait for it, and absence of public comments is not evidence of quality.

Public feedback is valuable as:

• a sentinel for concrete defects missed by systematic review;
• field evidence about actual interpretation and use;
• a source of counterexamples and alternative explanations; and
• a long-term challenge mechanism for approved decisions.

It is not a popularity vote, a rapid release gate, or a substitute for the structured multi-AI review.

8.2 Version-bound feedback envelope

Every feedback action carries immutable context supplied by the publication:

• full goalId;
• goal fingerprint;
• book or atlas version;
• locale and scope;
• page or canonical URL; and
• relevant publication manifest fingerprint.

The form asks for structured observations such as:

• Which mathematical idea or relationship should become clear?
• Which observable performance would demonstrate that understanding?
• Which changed example, representation, or context would show transfer?
• Which parts of the goal should be covered together?
• Which concrete improvement would make those expectations clearer?

It may request an optional role and contact path, but it must not solicit learner names, permanent SkillPilot IDs, chat secrets, or unnecessary personal data. Submissions warn users not to include identifying learner information.

8.3 Moderation and authority

Public submissions enter a moderated intake queue. An authorized reviewer may use them to create a finding, request reproduction, or immediately block a concrete affected artifact for a credible factual, safety, privacy, or normative concern. The submission itself cannot edit a goal or profile, resolve its own review, block a release automatically, or create a global rule.

One teacher's comment may be exactly right about one goal. It is therefore not dismissed. But role, confidence, public agreement, or forceful language does not increase the supported generalization scope. Broader rules still pass the E0-E5 and G0-G4 gates.

Feedback questions should focus on observable learner reasoning rather than ask teachers to defend or grade “the SkillPilot didactics.” Materials should be anonymized and, where practical, reviewed blind. This reduces avoidable role conflict while preserving critical challenge.

Public triage runs on a slower scheduled cadence, with a separate urgent lane for factual, safety, privacy, accessibility, or normative defects. Public feedback is retained against the exact historical fingerprint even after the current goal changes.

9. Rollout

9.1 Phase 0: contract and deterministic skeleton

Implement and validate:

• the evidence-profile schema and fingerprint algorithms;
• the BookModel schema;
• composition-view projection and stable topological ordering;
• HTML/PDF parity manifests;
• the one-goal-page and link validators;
• AI-run manifests and structured finding schema; and
• the public feedback envelope, without making public response volume a gate.

No runtime or Mastery behavior changes in this phase.

9.2 Phase 1: focused mathematics pilot

Start with the concrete representation-choice goal 8dd9f210-2683-5902-acab-e3be22725232 (“Darstellungsform auswählen und begründen”). The pilot must test at least:

• whether table, graph, term, and sketch each receive a case in which their particular advantage matters;
• whether the learner can compare the chosen representation with the closest alternative;
• whether the learner can name a limitation of the chosen representation;
• whether the learner can choose and justify a representation independently in a fresh presentation; and
• whether the learner transfers the same selection criteria to a changed case with a different surface form.

The output is one approved pilot profile, one chapter-scale HTML/PDF book, a reproducible understanding-evidence review set, and documented dissent. The pilot is not evidence for a global Coach rule.

9.3 Phase 2: mathematics and physics calibration

Calibrate on a default stratified set of at least 24 ordinary atomic goals:

• at least 12 mathematics and 12 physics goals;
• both Sekundarstufe I and Sekundarstufe II;
• all relevant archetypes present in each subject; and
• cases with and without visualizations, cross-chapter prerequisites, common conceptual distinctions, and representation choices.

The sample and replacement rationale are versioned in configuration. Exit requires stable schemas, acceptable inter-review consistency, successful calibration of the positive evidence fields, deterministic books, and no unresolved substantive disagreements. Calibration changes invalidate prior unbound bulk output.

9.4 Phase 3: bulk mathematics and physics review

After calibration, generate and review profiles for all in-scope ordinary atomic goals in canonical Gymnasium mathematics and physics. Work in bounded chapter batches, then run overlap batches and global passes for:

• coherent, non-duplicative expectations;
• complete prerequisite coverage;
• consistent terminology and difficulty;
• alignment between visualizations and independently observable performance;
• cross-chapter progression; and
• coverage of changed-case transfer expectations.

Bulk generation never bulk-approves. Content-specific disagreement, novel evidence expectations, schema violations, and proposed canonical text changes stay in explicit review queues. A green count is not obtained by filling every profile with a generic template.

9.4.1 Concrete two-round understanding-evidence workflow

The first nationwide mathematics wording pass uses a small, versioned calibration before the 754-goal run. Its 16 goals cover Jahrgangsstufe 5 through 10 and E through Q4, including concepts, procedures, representations, modeling, proof/derivation, data, and strategy choice. The sample calibrates three positive, content-specific expectations for every goal: essential understanding, independently observable performance, and transfer to a structurally related but changed case presented independently as a fresh task. The exact sample and order are authored in mathematik-description-understanding-evidence-kalibrierung.view.json; the corresponding GoalBook configuration is de-de-gym-math-description-understanding-evidence-calibration.json. Replacing a convenient sample goal silently is not allowed; a changed calibration sample is a new version.

The operational sequence is:

1. Freeze Atlas V1 inputs. Build the calibration book and review packets from the current canonical graph. Bind the book, page, goal, prompt, criteria, and batch fingerprints before any reviewer starts.
2. Codex blind V1 review. Codex receives no earlier reviewer output and writes exactly one candidate record for every assigned goal against goal-description-review-record.schema.json. Every record states, in German and English, the content-specific essential understanding, the independently observable learner performance, and the expected transfer to a changed, independently presented case. keep leaves the text alone; revise supplies one exact German and one semantically equivalent English replacement; split_review and block prevent a wording-only repair. Detailed assessment cases and variation sequences remain work for the positive-understanding-evidence-v2 profile rather than being copied into the learner-facing description. Each record binds that exact profile contract and recommends none, create, or revise; this recommendation remains a candidate and never mutates a profile.
3. Deterministic validation and human adjudication. The campaign check proves exact sample coverage, order, fingerprints, DE/EN replacement parity fields, candidate authority, and absence of learner data. A responsible human then accepts or rejects proposals goal by goal. AI output, an empty finding set, or model agreement is never an adjudication and never writes canonical data.
4. Controlled V2 authoring. Only human-accepted replacements enter an authoring change set. The application step preserves stable goal IDs and graph relations, changes German and English together, and re-runs every fingerprint-bound quality lane affected by the semantic text. Stale semantic-kind, semantic-atomicity, memory-card, visualization, evidence, and publication records are reviewed or regenerated according to their own contracts; their hashes are not merely refreshed to make checks pass. The nationwide Atlas V2 and its manifests are then rebuilt from canonical data.
5. External blind V2 review. A reviewer from another provider or model family receives Atlas V2 with the same locked understanding-evidence prompt, criteria, output schema, goal order, and batch policy, but receives neither Codex records, adjudication notes, nor the V1-to-V2 diff. The new run binds the V2 bundle and fingerprints and also remains candidate/ai_candidate.
6. Compare and resolve. Synthesis joins both rounds by stable goal ID while retaining their different input fingerprints and all dissent. An external revise, split_review, block, substantive disagreement about essential understanding, independently observable performance or changed-case transfer, or a DE/EN regression returns the goal to human adjudication. It does not trigger an automatic second canonical rewrite.

After the 16-goal calibration passes, the same sequence is applied to all 754 mathematics goals in deterministic bounded batches with a final exact-coverage check. “The same external round” means the same locked review contract and sampling policy over V2, not reuse of V1 fingerprints or disclosure of the first reviewer's answer.

Sharper descriptions help the current coach identify the aspects covered by an active goal. They do not by themselves establish the expected evidence of understanding. Detailed content-specific facets, independent-performance tasks, and changed-case transfer tasks remain in positive-understanding-evidence-v2 profiles, and hard Mastery enforcement remains subject to the separate runtime gates in Section 10.

9.5 Phase 4: public atlas and slow feedback integration

Publish the versioned HTML atlas and corresponding PDFs, with a feedback target on every goal page. The infrastructure may go live earlier as a preview, but its sparse early feedback is not an input-volume gate for Phases 1-3.

Public findings are triaged against the exact goal and publication fingerprint. Validated local corrections re-enter the same review pipeline. Generalization requires the same evidence as an internally discovered finding.

9.6 Phase 5: optional pedagogical text and worksheets

Only after the relevant evidence profiles are approved may topic-level pedagogical book text or teacher worksheets be generated from them. These are separate, versioned publication artifacts. They must cite the goal IDs they cover and must not silently introduce new curricular goals or mastery rules.

10. Later runtime and Mastery integration

Evidence profiles can eventually reduce the reasoning burden on a learning model by projecting only the approved profile for the active goal. This is not part of the initial publication rollout.

Before runtime projection, a separate implementation concept and acceptance gate must define:

• the bounded safe projection; only the active approved profile is eligible;
• prompt-size and privacy limits;
• handling of stale or missing profiles;
• provider-neutral behavior and provider-specific adapter contracts;
• adversarial dialogue tests; and
• feature flags, canary rollout, observability, and rollback.

Before a profile can control Mastery, the backend must be able to validate structured evidence rather than trust an unconstrained model assertion. A later design may use server-issued challenges or evidence receipts that bind facets, assistance level, case variation, and state version. It must prove that all required facets and independent checks were covered.

Mastery requires an independently completed fresh application case whose variation is meaningfully different from the teaching example and whose performance covers the profile's required expectations. Until the server can enforce these positive evidence requirements, the profile is guidance and evaluation material, not a hard Mastery guarantee.

11. Publication versions and reproducibility

Every release manifest binds:

• book/atlas version;
• generator and schema versions;
• canonical package or landscape fingerprints;
• composition-view ID and fingerprint;
• for union editions, the closed source-manifest digest, every bound projection fingerprint, the navigation-ownership contract, and the reviewed duration-policy digest;
• locale and scope;
• ordered goal IDs and goal fingerprints;
• evidence-profile fingerprints and statuses;
• visualization/resource hashes;
• the normalized feedback target used by the rendered links;
• HTML and PDF artifact hashes; and
• AI-review criteria and run-set identifiers used for the release decision.

The public URL and PDF footer expose the book version and goal fingerprint in a human-copyable form. The full manifest is machine-readable. A historical feedback link continues to identify the historical page; it never silently retargets a comment to a newer goal revision.

12. Acceptance criteria

The first production-capable implementation is accepted only when:

1. identical inputs produce the same BookModel, goal order, destinations, and semantic manifest;
2. HTML and PDF contain the same included goal IDs in the same order;
3. every included ordinary technical atomic target goal has exactly one complete goal page and no excluded semantic kind receives one;
4. each page shows the full Lernziel-ID, title, description, version, and fingerprint without truncation;
5. all internal requires and reverse-requires links resolve to the correct named destination;
6. every out-of-book relationship is explicitly marked and safely resolved;
7. every page passes the fail-closed overflow and minimum-legibility checks;
8. no learner ID, learner state, permanent SkillPilot ID, credential, or private model trace is present in any publication or review bundle;
9. every current approved profile matches its goal and review-input fingerprints;
10. AI findings are reproducible, structured, dissent-preserving, and scoped;
11. public feedback is version-bound and cannot mutate canonical state;
12. the mathematics pilot requires the learner to choose and justify representations across fresh questions whose information needs make different representations suitable;
13. runtime and Mastery remain unchanged until their later explicit gates pass;
14. a nationwide atlas contains the exact union of all bound curricular-atomic targets once and only once;
15. every printed and filterable applicability row is an exact Bundesland-bound tuple, and no independent G8/G9 or GK/LK cross-product can be constructed;
16. changing a bound composition view or the duration-model policy makes the publication check fail until BookModel, PDF, and manifests are regenerated; and
17. cross-stage or cross-profile goals never inherit a contradictory chapter path from the lexicographically first source projection.

13. Required validation classes

The implementation should provide focused commands for these classes rather than one opaque all-purpose check:

• graph projection, cycle, and missing-reference validation;
• deterministic topological order and single-occurrence validation;
• BookModel schema and manifest validation;
• HTML/PDF parity validation;
• PDF page ownership, destination, link, overflow, and legibility validation;
• evidence-profile schema, fingerprint, and stale-decision validation;
• AI-run independence and finding-schema validation;
• public-feedback envelope and privacy validation;
• nationwide source-union completeness, policy binding, exact applicability tuple, and non-Cartesian filter validation; and
• pilot adversarial dialogue evaluation.

These checks belong in the normal curriculum/publication CI once their formats are stable. Updating a fingerprint baseline merely to make a changed artifact green is forbidden; the corresponding review must be repeated.

14. Deliberate non-goals

This concept does not:

• replace official curricula, source rationales, or composition views;
• turn contains into a didactic prerequisite relation;
• put long rubrics into learner-facing goal descriptions;
• publish hidden solutions or detailed exam scoring material;
• infer understanding from image recognition or repeated labels;
• let an AI model or a public commenter approve its own change;
• make teacher participation a prerequisite for early iteration;
• treat silence, popularity, or majority votes as evidence;
• expose a learner's SkillPilot ID for book navigation or feedback;
• treat G8/G9 as a state-independent property; or
• claim server-enforced Mastery from prompt guidance alone.

15. Implemented local workflow

The current implementation exposes separate, reviewable commands rather than one opaque bulk mutation:

npm --prefix app run quality:goal-evidence:check

npm --prefix app run build:goal-book-model -- \
  scripts/config/goal-books/de-de-gym-math-representation-choice-pilot.json

npm --prefix app run build:goal-book-model -- \
  scripts/config/goal-books/de-gym-math-national-atlas.json

npm --prefix app run render:goal-book -- \
  --model ../tmp/goal-books/de-gym-mathematik-bundesweit.book-model.json \
  --feedback-base-url https://skillpilot.com/lernziel-feedback \
  --pdf public/lernzielbuch/de-gym-mathematik-bundesweit.pdf \
  --print-derivative-profile bounded-atlas

npm --prefix app run render:goal-book -- \
  --model ../tmp/goal-books/de-de-gym-math-representation-choice-pilot.book-model.json \
  --feedback-base-url https://skillpilot.com/lernziel-feedback \
  --html ../tmp/goal-books/de-de-gym-math-representation-choice-pilot.html \
  --pdf ../tmp/goal-books/de-de-gym-math-representation-choice-pilot.pdf

npm --prefix app run export:goal-book-review-bundle -- \
  --model ../tmp/goal-books/de-de-gym-math-representation-choice-pilot.book-model.json \
  --html ../tmp/goal-books/de-de-gym-math-representation-choice-pilot.html \
  --pdf ../tmp/goal-books/de-de-gym-math-representation-choice-pilot.pdf \
  --output ../tmp/goal-books/review-bundles/representation-choice-pilot-v1 \
  --goal-id 8dd9f210-2683-5902-acab-e3be22725232

The exporter requires a new output directory and writes the exact PDF, HTML, BookModel, normalized JSON, JSONL, Markdown, prompt, criteria, schemas, and a digest-bound manifest. Independent AI outputs are accepted only together with their run manifest and are checked with validate:goal-evidence-findings. Generated books and bundles under tmp/ are disposable build artifacts, not source-of-truth files.

The HTTPS feedback base in this command is the reserved version-bound target for the later moderated intake. Its schema and link context are implemented; the production form and moderation queue are not claimed as deployed in the current phase.

16. References to adjacent concepts

• SkillPilot Skill Graph Specification
• General Goal System and Migration
• View Projection and Goal Placement
• Atomic Goal Visualizations
• Human-Readable Source Rationales
• Dual Curriculum Package Releases