Submission scope: this demo runs on 97 in-scope PDFs (5 NCCA policies + 5 NCCE artefacts + 87 English Leaving Cert syllabi + 1 sample). The full BIEP substrate (DLT + CocoIndex + BAML + Google ADK + Gemma 4 + Gemini 3.5 + A2UI + Gradio + Web SPA) is preserved. See docs/SUBMISSION_SCOPE.md for the full scope and the 8 jurisdictions that are deferred for this submission.

Google All Things Agentic 2026 Hackathon submission. Built with Gemini 3.5 (Vertex AI) + Gemma 4 26B-A4B (Unsloth Studio) + the Google ADK 2 agent framework. Deployed on Google Cloud Run with the editorial canvas on Hugging Face Spaces. The British Isles education system: Aistear → Bunscoil → MeanScoil → Scoil Sinsearach → Ollscoil across 9 jurisdictions (Ireland + England active; NCCE cross-jurisdiction for the showcase; NI / Wales / Scotland / IoM / Jersey / Guernsey as Phase 2).

The headline feature of the 2026-08-31 batch. The UK NCCE (National Centre for Computing Education) publishes its computing curriculum as structured row × column learning graphs — the canonical example of how every official syllabus should be modelled. The 5 lifted PDFs:

NCCE PDF → DLT resource → CocoIndex App → BAML extraction → Dagster asset → Firestore + Gradio SVG

1. DLT (dlt_pipelines/uk_ncce_learning_graphs.py) — emits 11 rows into official_documents
2. CocoIndex (cocoindex_flows/uk_ncce/learning_graphs_app.py) — Docling converter that preserves the row × column grid via the new _shared/_docling_grid_segmenter.py
3. BAML (baml_extracts/learning_graph.baml) — 8 classes + 9 functions; 6 per-subject extractors for Computer Science, Mathematics, English, Gaeilge, Chemistry, Geography
4. Dagster (orchestration/defs/3_model_lifecycle/uk_ncce_learning_graphs.py) — 11 assets; sensor on data/bi_ep/syllabi_raw/uk_ncce/
5. Gradio + HF Space + React (gemini_hackathon_gradio/an_learning_graph/, hf_spaces/gemini_hackathon_learning_graphs/, web/src/routes/learning-graphs/) — 4-tab studio: Render / Equivalencies / Generate from PDF / Pedagogy overlay

Every cell in every jurisdiction's learning graph has a pointer to its equivalent cells in the 7 other BI jurisdictions. Powered by ExtractCellEquivalencies (Change B). Visualised as a Sankey diagram.

Every cell is coloured by which of the 12 NCCE pedagogy principles it uses (PRIMM, pair programming, semantic waves, etc.). Principles are dynamically extracted from pedagogy_principles.pdf and cached to disk + Cognee (Change C).

[SCREENSHOT placeholder — the SVG grid of the Y8 Python learning graph with pedagogy overlay]

• gcp-first-iac-refactor-v1 — Phase 0: drops Komodo/Pangolin/Locket/Infisical for Cloud Run + Secret Manager + WIF + 11 Terraform modules
• observability-otel-completeness-v1 — Phase 1: ADK OTel OTLP path + OpenInference Langfuse instrumentor + 6 Stackdriver env vars
• cocoindex-pdf-pipeline-v1 — Phase 2: pdf_to_markdown App (Docling converter) + Dagster asset + MLflow benchmark
• retire-letta-wire-vertex-memory-bank-v1 — Phase 0 (memory): replaces Letta with VertexAiMemoryBankService + MarkdownMemoryService + InMemoryMemoryService

• uk-ncce-learning-graph-showcase-v1 — the headline change (see §1)
• learning-graph-equivalency-graph-v1 — extends ExtractEquivalencies from linear topics to cell-level (48 jurisdiction × subject cross-walks)
• pedagogy-overlay-renderer-v1 — dynamic extraction + disk + Cognee cache for the 12 NCCE pedagogy principles; renders the annotated SVG

The Celtic 5-element palette (Talamh / Uisce / Tine / Aer / Anam) was REPLACED with the 5-stage British Isles education palette (Aistear / Bunscoil / MeanScoil / Scoil Sinsearach / Ollscoil). The Hades Shadow-First dark theme is preserved.

See ARCHITECTURE.md for the full picture, or docs/TUATHA_CONSOLIDATION_MAP.md for what was absorbed from sruth/tuath and /dev/tuatha.

Every claim on every generated certificate cites a page from one of the 5 NCCA policy PDFs:

SC-L1-L2-Programme-Statement.pdf,
key-competencies-in-senior-cycle_en.pdf,
the-potential-of-online-learning-environments_en.pdf,
the-potential-of-technology-to-support-online-certification-and-reporting.pdf,
scr-advisory-report_en.pdf

The pipeline (7 stages): extract_criteria → decompose_outcomes → extract_paper+marking → search_official → generate_background → compose_certificate → save_to_provenance.

Output: PNG (~80 KB) + PDF (~700 B) with the awarding-body palette background + competency strip + provenance footer + UNOFFICIAL banner.

gemini_hackathon/
├── pyproject.toml + requirements.txt + Makefile
├── gemini_hackathon/                       ← the Python package
│   ├── agents/                             ← W7 — 5 stage coordinators + 5 pillars
│   ├── assets/                             ← the image-gen router
│   ├── call_llm.py + backend.py            ← the LLM gateway
│   ├── certificate/                        ← W14 — the LC/JC certificate pipeline
│   ├── cli.py + __main__.py + compare.py
│   ├── knowledge_graph/                    ← W8 — hybrid search (LanceDB + FalkorDB)
│   ├── ledger/                             ← W9 — skill-progression ledger
│   ├── memory/                             ← W8 — MarkdownMemoryService
│   ├── models/ + model_registry.py
│   ├── observability.py
│   ├── ocr.py + ocr_ensemble.py
│   ├── progression/                       ← existing LC/JC certificate schema
│   ├── session/
│   ├── sources.py + theming.py
│   └── subnations.py                       ← W11 — 9 jurisdictions
├── gemini_hackathon_gradio/                ← W3 — the 5 editorial studios + 2026-08-31 an_learning_graph
├── gemini_hackathon_assets_fibo/            ← W10 — the FIBO image-gen pipeline
├── gemini_hackathon_backend/                ← 2026-08-30 — the ADK 2 Cloud Run service
│   ├── agents/ + memory.py + observability.py + main.py
│   ├── catalog/                            ← the Lake namespace adapter
│   └── tests/
├── baml_extracts/                          ← BAML contracts (existing + 2026-08-31 learning_graph.baml)
├── baml_extracts_education/                ← W4+W5 — lifted from cianfhoghlaim
├── cocoindex_flows/                        ← W5 + 2026-08-30 + 2026-08-31
│   ├── _shared/                             ← shared_lifespan + _vector_target + _docling_grid_segmenter
│   ├── ireland/                             ← per-stage embedding apps
│   ├── pdf/                                 ← Phase 2b Docling converter
│   ├── uk_ncce/                             ← 2026-08-31 NCCE learning graph App
│   ├── equivalency/                         ← Phase 4a equivalency graph
│   ├── education/                           ← W14 BAML extraction target
│   └── _factory/                            ← bi_jurisdiction + four_stage
├── dlt_pipelines/                           ← W5 + 2026-08-30 + 2026-08-31
│   ├── ireland/                              ← 6 NCCA subject pipelines
│   ├── uk_ncce_learning_graphs.py            ← 2026-08-31 (jurisdiction #9)
│   ├── official_doc_fetcher.py
│   ├── pdf_downloader.py
│   ├── corpus_downloader.py
│   └── _shared.py + _base/ + _subject_base.py
├── data/
│   ├── ireland/ncca_policy/                ← W2 — 5 NCCA policy PDFs
│   ├── bi_ep/                              ← 2026-08-30 PDF→Markdown substrate
│   │   ├── syllabi_raw/                      ← raw PDFs (gitignored)
│   │   │   ├── ireland/                       ← NCCA + SEC + DES PDFs
│   │   │   └── uk_ncce/                       ← 2026-08-31 (5 NCCE PDFs)
│   │   ├── syllabi_md/                       ← Docling-converted Markdown
│   │   └── extracted_syllabi.sqlite          ← the BAML extraction target
│   ├── leabharlann/                         ← W6 — leabharlann corpus (manifests only)
│   ├── equivalencies/, jurisdictions/, marking_schemes/, policies/, sources/, syllabi/
├── themes/                                  ← awarding-body palettes (NCCA, AQA, OCR, ...)
├── web/                                    ← Firebase-native TanStack Start web (post Firebase migration)
│   └── src/routes/learning-graphs/         ← 2026-08-31 NCCE landing page (embeds HF Space)
├── hf_spaces/                               ← W13 + 2026-08-31 (6 headline demos)
│   ├── _generate.py
│   ├── gemini_hackathon_{aistear,bunscoil,junior_cycle,leaving_certificate,editorial_studio}/
│   └── gemini_hackathon_learning_graphs/   ← 2026-08-31
├── orchestration/                           ← 2026-08-30 + 2026-08-31
│   ├── defs/                                ← 5-layer Dagster defs tree
│   │   ├── 1_ingestion/
│   │   ├── 2_materials/
│   │   ├── 3_model_lifecycle/               ← NCCE learning graph assets (2026-08-31)
│   │   ├── 4_asset_generation/
│   │   └── 5_agent_ops/
│   └── storage/                             ← BigQuery client + DuckLake client
├── openspec/changes/                       ← W16 — 26 openspec changes (23 + 3 NEW 2026-08-31)
├── docs/                                   ← W15 — the docs (TUATHA_CONSOLIDATION_MAP, KNOWN_ISSUES, LEARNING_GRAPH_SHOWCASE, ...)
├── notebooks/                              ← the 14 marimo + ipynb walkthroughs (00_theming … 13_pedagogy_overlay)
├── cloud/                                  ← GCP-native IaC
│   └── terraform/
│       ├── cloud_run_adk.tf + cloud_run.tf + cloud_run_jobs.tf + cloud_run_journey.tf
│       └── modules/                          ← 11 Terraform modules (per Phase 5 GCP-first IaC)
├── cloudbuild.yaml + Makefile + README.md
└── firestore.{json,rules,indexes.json}     ← Firebase-native schema + NCCE collections

The 4-path OCR/VLM ensemble (BAML + Docling + EasyOCR + Tesseract) with RAGAS voting + MLflow observability. Scanned-PDF detector (is_scanned_pdf()) routes to the recommended backend via select_ocr_backend(). The 22 VLM + 6 classical OCR entries live in MODEL_REGISTRY:ocr_vision (qwen3-vl-8b + qwen3-vl-4b added in 2026-08-25-tuatha-vision-models-v1).

• _shared/{_lifespan, _vector_target, _vertex_embedder, _docling_grid_segmenter} — canonical lifespan + dual-backed vector target (Firestore default / Vertex AI Vector Search) + VertexEmbedder + grid-preserving Docling segmenter (2026-08-31)
• ireland/ — per-stage embedding apps for the 6 NCCA LC subjects
• pdf/pdf_to_markdown_app.py — Phase 2b Docling converter
• uk_ncce/learning_graphs_app.py — 2026-08-31 NCCE learning graph App
• _factory/{bi_jurisdiction, four_stage} — the 114-App 4-stage BI factory
• equivalency/equivalency_graph_app.py — Phase 4a cross-jurisdiction graph
• education/lc6_extraction_app.py — W14 BAML extractor target

• 6 NCCA LC subjects + cross-jurisdiction + jurisdiction sensors
• pdf_downloader.py (Phase 2) + corpus_downloader.py + official_doc_fetcher.py
• uk_ncce_learning_graphs.py (2026-08-31 — jurisdiction #9)
• DLT 1.30 features: refresh + dlt[hub] + pipeline.dataset()
• DuckDB → BigLake Iceberg (prod) via orchestration/storage/bigquery_client.py:BIGQUERY_DATASET = "biep"

orchestration/defs/{1_ingestion, 2_materials, 3_model_lifecycle, 4_asset_generation, 5_agent_ops}. The 2026-08-31 NCCE learning graph assets (11) + equivalency assets (48) + pedagogy overlay assets (6) + 1 sensor live under 3_model_lifecycle/. Wired via dg CLI (Dagster 1.13+).

• baml_extracts/learning_graph.baml (2026-08-31) — 8 classes + 9 functions for the structured learning graphs
• baml_extracts/extract_equivalency.baml — topic-level + cell-level equivalencies (extended in 2026-08-31 Change B)
• baml_extracts_education/{celtic_curriculum, subjects/*}.baml — per-subject BAML contracts with strand + BloomLevel metadata

pgvector/pgvector:pg17 + lake-keeper-rust + lance-namespace + cognee + falkordb + memgraph + clickhouse + garage (S3) + olake (CDC). Langfuse v3 :3001 + MLflow v2.20 :5050 (the dual-purpose local observability stack).

• Cloud SQL Enterprise HA (Postgres)
• Memorystore Standard M3 (Valkey)
• BigQuery dataset biep (via orchestration/storage/bigquery_client.py)
• Lance namespace iceberg backend → BigLake Iceberg REST (Lakekeeper)
• GCS bucket for the PDF→Markdown raw + md pairs

cloudrun_service, cloudrun_secret_mount, cloudsql_postgres, memorystore_valkey, gcs_bucket, bigquery_dataset, firestore_database, artifact_registry_repo, workload_identity_gha, cloudbuild_trigger, iam_gcp_ai_agent_adk.

• VertexAiMemoryBankService (prod, when DEPLOYED_AGENT_ENGINE_ID is set)
• MarkdownMemoryService (dev/offline, when GH_MEMORY_DIR is set)
• InMemoryMemoryService (default fallback)
• before_agent_callback → add_session_to_memory
• letta_agent_id field renamed to memory_namespace (semantics change, same shape)

• ADK OTel OTLP path → Cloud Trace + Cloud Logging + Cloud Monitoring
• 6 Stackdriver env vars (per the Stackdriver AI Agent ADK doc)
• OpenInference Langfuse instrumentor (wraps every ADK call as a nested Langfuse span)
• Langfuse v3 :3001 + MLflow v2.20 :5050 (local dev)

aistear          dawn-orange       #e8915c
bunscoil         sea-blue          #1e80c6
meanscoil        meadow-green      #28955e
scoil_sinsearach harvest-gold      #cc9966
ollscoil         scholarship-indigo #5a4fcf

Hades Shadow-First is preserved as the dark-mode foundation.

8 NCCA + 6 NCCA-adjacent subjects. The 2026-08-31 NCCE showcase adds a 9th routing key (learning_graph) for the structured cross-walk view.

The 5 ADK 2 stage coordinators (gemini_hackathon/agents/stages/) and the 5 reusable workflow pillars (gemini_hackathon/agents/workflows/) remain the runtime substrate.