DATACAREERHUB — APPLIED AI
Nine original lessons: semantic retrieval, structured extraction and factual support.
All CedarLab policies/invoices/claims are fictional. They are NOT DataCareerHub service terms.
Public payments and registration remain disabled. No website DB or live chatbot is added.

CORE / NO INSTALL / NO NETWORK
Extract the entire ZIP and open a terminal inside applied-ai-kit.
Python 3.10+ (tested 3.12.14):
python3 lab.py
python3 -m unittest discover -s tests -v
Windows: use py or your chosen Python executable where appropriate.
Open lesson-guide.html for all lessons; worked-report.html for visual evidence.
This command replays recorded vectors from an actual local pretrained encoder; it does NOT
encode arbitrary new text. Unknown query IDs and changed source hashes fail explicitly.
Extraction is deterministic parsing; factual support uses a human rubric and authored labels.
The fixture callback in provider_contract.py is NOT a live LLM/provider integration or sandbox.
The caller remains responsible for what any supplied Python callback does.

OPTIONAL ACTUAL LOCAL SEMANTIC ENCODER
Requires Python 3.12 and compatible wheels; tested macOS arm 64 (numpy wheel macOS 14+).
Create a fresh virtual environment in your extracted practice folder:
python3 -m venv .venv
macOS/Linux commands (Windows uses .venv\Scripts\python.exe):
.venv/bin/python -m pip install -r requirements-semantic-lock.txt
.venv/bin/python fetch_model.py --destination ./model
.venv/bin/python encode.py --model-dir ./model --query "Can I hand my workshop credit to someone else?"

The explicit fetch downloads about 91 MB of public model/tokenizer/card files, verifies
immutable revision and hashes and retains them locally. Pip also requires internet for
public packages. No documents or query text are uploaded to any model/provider.
After setup, inference uses only local files. No GPU, cloud/API key or paid service.
Do not upload model/, .venv/, caches or the installed packages to Aruba.
New query above ranks D3 first on the executed CPU reference; scores are not probabilities.

REGENERATING YOUR OWN FIXTURES
Keep original files and results. For a deliberate new corpus/query experiment, copy the
kit, update labels before measuring, then run:
.venv/bin/python encode.py --model-dir ./model --write-fixtures
python3 lab.py
--write-fixtures explicitly replaces embeddings.json in that practice copy.
Normal report commands never download or regenerate silently.
Inputs exceeding 256 wordpieces fail; design chunks with provenance rather than hide truncation.
The dependency inventory records one tested environment, not a cross-platform compatibility
or security guarantee. Core lessons remain usable if optional model setup is unavailable.

CONTENTS
corpus.json:8 short fictional passages,7 current and 1 retired draft, including injection text.
queries.json:3 dev,3 answerabletest,1 unsupported. Tiny authored evaluation, not a general benchmark.
embeddings.json:384-dimensional actual model outputs with provenance/hashes.
invoices.json:6 narrow extraction cases;2 accepted,4 rejected by declared contract.
claims.json:8 claims, human reference labels/reasons. Attempt reviews-blank.json before gold.
reviews-example.json:imperfect hand-authored reviewer;6/8 accuracy,1/3 supported precision.
No automatic entailment model is implemented. Exact quote checks are not truth judgments.
model-manifest.json, requirements-semantic-lock.txt, MODEL-NOTICE.txt record provenance.
extraction-prompt-template.txt gives a hypothetical structured-AI prompt and application checks.
exercises.txt, worked-solutions.txt, support-rubric.txt, privacy-cost-contract-template.txt,
privacy-cost-contract-worked.txt, evidence-memo-template.txt and capstone-rubric.txt support work.

LIMITS
No generative extraction model, live provider adapter, autonomous factual judge, production
vector DB, private document permissions, OCR, financial service or deployed assistant.
Actual encoder was executed locally; replay and model timings are distinct.
All finite-case metrics are teaching evidence, not commercial accuracy guarantees.
No grading, certificate or employment guarantee. Cite supplied code and your own contribution.
