# Advanced GEO engineering teaching package

Version: 2026-10-02 / benchmark 2.0.0. Author: Zhihe Growth.

Synthetic teaching material, not client data or new commercial AI tests. Labels are author-defined, without independent double review. Reference code is not a claim of a deployed customer-production toolchain.

## Executed scope

- controls.py: character-budget checks, query constraints, citation arithmetic, sample planning, version approval, read permissions, export fields, missing events and temporary-directory rollback.
- assessments.py: four HTML fixtures, five language mappings, four schema/fact records, six synthetic repeat attempts and three synthetic attribution records.
- media_labs.py: owned two-page PDF extraction, twelve bilingual facts and twelve paired queries producing 48 language-direction rankings, plus five CLIP queries over four geometric images.
- review-templates.json: blank human, video, joint-citation, sampling and lead-review forms. Null means unperformed, not a hidden successful result.
- architecture-examples.json: JSON-LD, graph and illustrative GraphRAG design for the same synthetic object; no full GraphRAG execution.
- topic-map.json: actual directory membership and five retention decisions; no URL consolidation or redirects performed.

## Installation and execution

Use an isolated Python environment. The recorded environment uses Python 3.12 and CPU. requirements-lock.txt records installed versions; lab-dependencies.txt pins direct dependencies. Model weights are not redistributed. The first model run downloads pinned revisions from their publishers; consider disk and network requirements before running it.

```sh
python -m venv .venv
. .venv/bin/activate
python -m pip install -r lab-dependencies.txt
python controls.py
python assessments.py
python media_labs.py
python media_labs.py --models
```

Scripts overwrite their result files in the package directory. The rollback exercise uses temporary directories only. No WordPress, client portal or production connection occurs. Model revisions come from model-revisions.json. Preserve original outputs before rerunning because execution timestamps change.

## Review the evidence

Check SHA256SUMS.txt first. Inspect the PDF and rendered PNGs, then compare text, tables and coordinates in document-results.json. Read relevance labels in benchmark.json before examining all rankings in model-results.json.

Cross-language metrics use ten answerable questions; two unanswerable questions retain rankings but are excluded from those denominators. Directional Top-1 counts are 8/10, 7/10, 8/10 and 8/10; Recall@3 is 10/10 in each. CLIP logits are not calibrated probabilities, product compatibility or real-photo generalization. The repeat fixture's 2/4 is not Zhihe Growth's current citation rate.

Video review, independent human annotation, commercial AI testing, sequential tests, full GraphRAG, production MCP and multi-tenant publication are not implemented here. Each helper has limited scope: trace_gaps checks event presence; translation_ready checks versions, not semantics; production origin must come from trusted application context.

## License and privacy

Authored synthetic data, PDF and geometric images are CC0-1.0. Package code is MIT; see LICENSE.txt. Third-party libraries, models and references retain their own terms. No client files, credentials, real personal data or model weights are included. Do not populate templates with confidential data and then publish them.

## References

- https://sbert.net/docs/sentence_transformer/pretrained_models.html
- https://huggingface.co/openai/clip-vit-base-patch32
- https://github.com/jsvine/pdfplumber
- https://pypdf.readthedocs.io/en/stable/user/extract-text.html
- https://json-schema.org/learn/getting-started-step-by-step

Knowledge directory: https://zhihezengzhang.com/en/knowledge-hub.html