Curriculum Release Model Conformance
DPK-004 compiles the real German Gymnasium Mathematik authoring state into a strict, unpacked release model. DPK-004a established the Runtime portion; DPK-004b adds authoritative publication evidence. Together they are the executable bridge between repository-specific sources and the later full-standalone-v1 JSON ZIP.
Current result: DPK-004 through DPK-007a are passed: the real profile, fixtures, independent validation, production/adversarial cases, byte-identical model and ZIP builds, finished-package gates, secure content-addressed provisioning/CAS activation, package-authoritative Runtime services, the hermetic whole-application consumer, final visualization import, and Core/registry lane-alignment gate pass.
At a Glance
| Question | Current answer |
|---|---|
| Is real curriculum data compiled? | Yes: 1 Mathematik landscape, 1,079 goals, 88 views, 12 decks, 128 cards, and 803 resource links. |
| Are runtime payloads strict? | Yes: every compiled runtime role has a closed Draft 2020-12 schema; unknown fields fail. |
| Is dependency closure explicit? | Yes: typed definitions and schema-registered hard references are closed to a fixed point; the real model has no unresolved hard reference or external runtime dependency. |
| Are images part of semantic identity? | Yes: all 734 active package images, totaling 1,603,749,500 bytes, contribute byte length and SHA-256 records to the shared contentDigest. |
| Is publication evidence part of semantic identity? | Yes: mappings, official sources, source-goal references, and quality evidence are four normalized logical artifacts in the same contentDigest. |
| Is publication evidence in the Runtime closure? | No: all four package roles declare runtimeRequired: false; the closure contains 2,403 Runtime definitions and 18,820 references, with zero publication roles. |
| Is publication quality honestly releasable? | No: 731 of 754 atomic scope goals have an active image, 23 are deliberately provider-deferred, and only 155 active visualizations are human-approved; 576 active reviews remain open, so the evidence artifact reports publicationStatus: not-ready. |
| Are image files copied into this output? | Not into the unpacked conformance-model directory. The companion ZIP materializes all 734 byte-bound package images. |
| Is this output a standalone package? | The unpacked directory deliberately remains conformance-model-only-not-a-package. The DPK-007a freeze derives a structurally valid full-standalone-v1 ZIP; DPK-006a–c safely provision, activate, load and serve it, and DPK-007 proves hermetic whole-application consumption. |
The short implementation workboard and remaining sequence are maintained in Dual Curriculum Package Implementation Status. The target architecture is Dual Curriculum Package Releases.
Trusted Inputs
The compiler does not discover release semantics from filenames or goal titles. Its trusted authored inputs are explicit:
- the Mathematik release-model build profile, which pins source and output paths, package identity, views, decks, resource policy, path relocations, contracts, and exact counts;
- the Mathematik ontology profile, whose own SHA-256 and canonical namespace-map SHA-256 are pinned by the build profile and which binds the Core-first mapping to the exact FWU repository, checkout path, commit, source path, ontology IRI, and file digest;
- the Mathematik semantic-kind ledger, which carries a current source fingerprint for every goal and forbids title-, ID-, or path-based kind inference;
- the field-semantics registry, normalization profile, strict payload schemas, and canonical definition-digest profile under
contracts/curriculum-package/v1/.
The semantic-kind ledger covers all 1,079 goals:
| Semantic kind | Count |
|---|---|
curricularAtomic |
754 |
curricularArea |
192 |
practiceAssessment |
113 |
programStructure |
7 |
memory |
6 |
runtimeSupport |
5 |
orientation |
2 |
The field registry contains 454 explicit entries: 323 across the DPK-004 Runtime roles, 130 across the four publication normalization roles, and one pre-existing generic legacy entry. It remains unchanged at SHA-256 2e536c3f8d63e2acf45690375ace69ec0c6a6e92787bc8a16957b80120c4ca48. Ontology and publication profiles 1.1.1 align Core-position and application-order predicates with the registry's declared lanes; compiler and independent validator reject profile/registry drift. The validator also walks every reachable field in the closed payload schemas in both directions: every schema field needs one effective registry mapping, and dead registry paths are rejected. The generated field-coverage.json separately reports which entries occurred in this concrete model; schema-only entries remain visible instead of being presented as instance-covered.
Generated Model
The output is a disposable directory below repository tmp/:
data/
canonical/mathematik.landscape.json
views/index.json
views/*.json
cards/card-index.json
cards/*.json
resources/resource-index.json
mappings/source-to-canonical.json
sources/source-index.json
sources/source-goal-references.json
runtime/catalog.json
runtime/dependency-closure.json
runtime/migration-aliases.json
assessment-sources/...
metadata/
quality/release-quality-evidence.json
semantic-content-index.json
field-coverage.json
build-inputs.json
release-model-conformance.json
The compiler performs only declared relocations. In particular, deck references below /data/ become data/cards/..., and Mathematik assessment-source references become data/assessment-sources/.... It preserves the relative source path instead of collapsing files to basenames.
The historical DPK-007a freeze turned an earlier state of this directory into skillpilot-curriculum-de-gymnasium-mathematik-0.1.0-conformance.3.json.zip; its 914-entry, 757-image byte bindings remain immutable historical evidence. The current authoring-state full-package Builder instead plans 893 entries, 891 manifest-inventoried files, and 734 binary assets bound to the current semantic digest below. Manifest and SHA256SUMS are the only profile-declared self-referential files outside the manifest inventory. Package-local schemas and semantic contracts, license evidence, source-/redistribution-review evidence, and every active image byte are copied into a newly built staging archive; no historical DPK-007a/008 artifact is overwritten or silently rebuilt.
The implementation-independent finished-ZIP validator checks the physical archive, inventory and actual payloads without importing Builder code. The v1 trust profile exposes its JSON resource limits directly: 64 MiB per JSON entry, maximum nesting depth 128 and at most 5,000,000 parsed nodes. Thus a package cannot pass the Builder but encounter an undocumented stricter JSON limit only in this validator.
The caller-selected output path is handled lexically and must be a strict descendant of repository tmp/. Existing symlink components are rejected. Compilation happens in a fresh private sibling staging directory, followed by rename-based promotion; recursive deletion is limited to private staging or backup paths created by that process. The conformance wrapper attacks both a symlink target and a symlink parent and proves that unrelated sentinel directories survive.
Before compiling payloads, the compiler proves that tmp/lehrplan-ontologie is the declared Git worktree, has exactly the bound origin and HEAD, and contains an unchanged regular src/ontology/components/lehrplan-core.owl whose committed and working-tree bytes match the profile hash and ontology IRI. Every compact Core term used by the mapping profile must expand through the pinned namespace map and occur in that OWL module. metadata/build-inputs.json records the complete profile, namespace, term-set, repository, commit, source-file, byte-count, and SHA-256 binding. The independent validator repeats these checks without importing compiler code.
dependency-closure.json inventories typed Runtime definitions and every schema-classified Runtime reference. Identical definitions can be deduplicated by canonical definition digest; the same stable identity with a different definition is a hard conflict. Publication evidence is deliberately not seeded or traversed: evidence about a goal must roundtrip and affect the content digest, but it must not become a navigation dependency. The initial Mathematik release has no predecessor, so migration-aliases.json uses an explicit initial baseline and an empty rule set. Later releases must compare against a pinned stable baseline and use the registered migration relation and mastery/history policy for every identity change.
semantic-content-index.json is package-path-neutral. It binds 111 normalized logical records—including all four publication artifacts—and 734 image binary records into one deterministic contentDigest; generated digest fields are excluded through the versioned normalization contract to avoid self-reference. For the current 0.1.0-conformance.3 authoring state the result is sha256:830b0e3b98b5ade7a4efc412fd39feb378e98302367fe57fa6c49e23e50a28c4. The older DPK-007a/008 release freeze remains bound to its historical digest and must be supplied as a frozen artifact rather than rebuilt from current sources. The closure digest is sha256:087aed6c6df2421d95c79f1b16e51d7ba8962c6fc463b766f642da8d3a15c523; the definition-index digest is sha256:2d15fa563e0ac3e9f5e2ea772913cf5dc378cfe06fd19a1da320cc5b47f8eeee.
Publication Evidence
Generic package roles remain stable for manifest inventory and consumers, while the profile binds each one to a specific normalizationRole for schema validation, field-registry lookup, and canonical hashing:
| Package role | Normalization role | Runtime required |
|---|---|---|
mapping |
source-to-canonical-mappings |
no |
source-index |
official-source-index |
no |
source-goal-reference-index |
source-goal-reference-index |
no |
quality-evidence |
release-quality-evidence |
no |
The mapping artifact is compiled from authoritative review decisions, not from the older legacy carrier. Its 10,021 decisions yield 33,382 exact/partial edges. Match type is fail-closed: an explicit decision value wins; otherwise the edge-specific reviewed mapping row supplies it. This resolves 2,599 edges without an unsafe missing => exact assumption and leaves zero unresolved types. The edge rows contain 33,334 edges: 48 decision edges are absent there, and 23 explicit decision-versus-row conflicts are resolved by the decision truth. Legacy files remain an authoring/audit compatibility lane only; their 2,539 rows include 695 source IDs outside historical membership and eight dangling targets and cannot override the decisions.
Source evidence covers 31 collections, 55 official documents, 9,977 reviewed source-aligned goal records, and 16 jurisdictions. A hash-bound role table classifies 52 actual curriculum documents as Core Lehrplan and three official implementation/restriction/profile documents as application-level OfficialSourceDocument; unknown, unused, or mismatched roles fail closed. Source goals conservatively use the generic Core Curriculares Element, while source-to-canonical edges use Core CE-Verweis. Exact/partial match type is the narrow mapping extension. Quality evidence is application-only and publishes 754 semantic-atomicity decisions, 754 memory-goal decisions, 64 active-card decisions, and 731 active visualization decisions—2,303 decisions in total. The atomic visualization scope contains 754 goals; its 23 missing image IDs are explicit provider-deferred gaps rather than active-resource decisions.
sourceText is the reviewed authored wording carried by the extraction/review lane, not a claim that every value is a verbatim PDF quotation. The release compiler preserves and hashes that wording exactly after the declared boundary-whitespace projection. The DPK-005a source-verification lane now proves 9,493 contiguous authored-carrier matches and five additional, replayable PDF-projection matches. The remaining 479 fingerprint-bound records require human classification. No machine match is a human or legal approval; DPK-004 proves lossless package projection, not independent quotation certification.
Strict scalar validation also exposed real PDF/OCR residue in source extraction: C0 controls and unpaired UTF-16 surrogates were corrected at the authoring source. The compiler rejects such unsafe values fail-closed; it does not sanitize them into a different release meaning.
Commands
Compile and independently validate the real model from the repository root:
python3 -B scripts/compile_curriculum_release_model.py \
--profile contracts/curriculum-package/v1/profiles/de-gymnasium-mathematik-release-model-v1.profile.json \
--output tmp/curriculum-release-model/mathematik-a
python3 -B scripts/validate_curriculum_release_model.py \
--profile contracts/curriculum-package/v1/profiles/de-gymnasium-mathematik-release-model-v1.profile.json \
--release-root tmp/curriculum-release-model/mathematik-a
The validator is implementation-independent from the compiler. Its normal invocation runs 47 named production/adversarial cases spanning strict fields and order, ontology/profile and Registry-Core pins, publication-role separation, source-role/semantic-type overclaims, fail-closed Mapping-Truth reconciliation, source joins and lineage, package-internal scope bindings, independently derived quality status, stable-key uniqueness, quality fingerprints and assets, the recorded FWU Core binding, legacy kompetenzen hard references, program-unit parent cycles, rejection of non-regular filesystem nodes, semantic-kind fingerprints, definitions and closure, migration policy, resource and record hashes, and reproducibility-sensitive bindings.
Nine valid release-model fixture documents are validated separately against the closed local schemas and current registry/definition-profile trust bindings; 22 targeted negative cases cover both Runtime fixed-point semantics and the four publication payloads. Their cross-document ownership and embedded hard-reference fixed point—including an exact goal.kompetenzen edge to a scoped competency definition—run with all remote schema retrieval disabled. A lightweight compiler probe invokes the same dependency emitter used by the real build and checks that compatibility edge byte-for-byte before the large model is compiled:
python3 -B scripts/validate_curriculum_release_model_fixtures.py
python3 -B scripts/compile_curriculum_release_model.py \
--self-test-dependency-emission
The required focused JSON gate combines those fixtures, destructive output-path attacks, the lightweight pinned-Core provenance binding recorded by the JSON profile, the real-model validator, current redistribution/source review bindings, and a byte-identical second model build. It deliberately stops before real ZIP materialization:
bash scripts/run_curriculum_release_model_conformance.sh
The optional full-package gate starts with that same release-model proof, then performs the real 1.7-GB Builder invocation, independent finished-ZIP validation, secure provisioning and activation, and an exact not-ready-incomplete Readiness assertion. The evaluator runs the policy-pinned consumer itself: a real React/Chromium flow and Catalog-404 fail-closed case execute in the checkout-hidden namespace, while retained assembly/evidence manifests allow independent tree-digest verification:
bash scripts/run_curriculum_full_package_conformance.sh
# or
./run_ci.sh package
It does not generate RDF, run ROBOT, validate OWL, invoke a reasoner, or perform an ontology roundtrip; those operations live in ./run_ci.sh owl. The scheduled/manual optional GitHub workflow runs the full package gate and the bounded OWL/roundtrip lane together.
The wrapper is strict by default: every referenced, locally retained source PDF
must be a regular non-symlink file and its real bytes are hashed. Clean hosted
checkouts do not contain the deliberately gitignored official PDFs, so the
GitHub workflows explicitly set
SKILLPILOT_SOURCE_PDF_MODE=committed-bindings. In that mode only an actually
missing PDF may fall back to the byte count and SHA-256 already bound by the
committed source-verification ledger. A present file is always hashed, and a
changed file, symlink, directory, missing ledger binding, update, or PDF replay
still fails closed. The full-standalone package builder forwards the same mode
as an explicit argument to its internal source-verification preflight. This
hosted mode verifies committed metadata and all derived repository evidence; it
is not a replacement for the strict local PDF or replay checks.
For a required per-commit checkpoint, targeted conformance checks are followed by the normal repository gate:
./run_ci.sh
The required gate covers the real 0.1.0-conformance.3 model, profile/registry lane-alignment mutations, all fixtures and adversarial cases, and deterministic model builds. The optional package gate additionally covers deterministic ZIP builds, Finished-ZIP validation, provisioning/activation, package-authoritative services, and the hermetic browser consumer. Generated models, ZIPs, stores, reports, assembly, and evidence trees remain under tmp/ and are not committed.
What Remains
| Step | Remaining outcome |
|---|---|
| DPK-008 | Produce and independently validate the Core-first FWU-OWL package, then reconstruct a valid JSON package in isolation without the original ZIP or source checkout. |
| DPK-009 | Compare the reconstructed normal form and binary assets, issue the real equivalence report, and bind the reproducible variant pair. |
The existing Subject Export ZIP remains not-ready-legacy. The current authoring projection for candidate .3 is technically operable but remains not-ready-incomplete: 576 active fachliche image reviews, 23 provider-deferred visualization gaps, 737 redistribution reviews, and 479 source-text reviews are still open. The historical .2 candidate was superseded by the immutable .3 DPK-007a/008 content freeze, whose byte bindings remain historical and are not rebuilt from current authoring sources. The dual JSON/FWU release does not exist until DPK-008/009 pass. The unpacked conformance-model directory itself remains non-installable.