Evidence-Grounded Interaction Engine + Alignment-Aware Comparison Dataset
Deterministic pharmacological safety modeling for intentional psychedelic use
Live demos:
www.entheogen.newpsychonaut.com · www.entheogen.azurewebsites.net
EntheoGen does not provide clinical medical advice.
For expert harm-reduction guidance, consult a qualified medical professional.
EntheoGen is a substance interaction guidance app focussed on intentional use of psychedelics.
The web application estimates the effects of mixing the two substances entered into the drop-down menus.
Data is personal and not stored remotely (i.e., not recorded or held by us).
EntheoGen's data model is mapped specifically to sacramental substances often used in psychedelic ceremonies, pharmaceutical medications, and commonly used consciousness-altering substances.
⚙️ Ceremonial-context interaction coverage — models sacramental psychedelic use scenarios
⚙️ Deterministic pharmacological rule engine — reproducible outputs
🛡️ Strict safety posture — missing evidence resolves to Unknown, never speculation
🔍 Source traceability — outputs include confidence and mechanism metadata
🧠 Alignment-aware dataset design — explicit abstention class for hallucination benchmarking
🧪 Mechanism-level labeling — structured taxonomy for causal and ontology research
📊 Slice-based benchmarking infrastructure — evaluate reasoning by mechanism and severity
📦 HuggingFace-ready dataset packaging
🔁 Regression safeguards — detect silent dataset drift
🧱 Local evaluation harness — oracle baseline + leaderboard pipeline
💾 Privacy-first architecture — no remote storage of user inputs
Version 2.0 extends EntheoGen beyond a web interaction tool into a reproducible benchmarking dataset and safety-alignment evaluation platform for studying pharmacological reasoning, abstention behavior, and mechanism classification.
EntheoGen now includes a structured interaction dataset and local evaluation harness suitable for research, benchmarking, and model comparison workflows.
New capabilities:
- provenance-aware interaction classification (
explicit,fallback,unknown,self) - normalized pharmacological mechanism taxonomy
- deterministic dataset export pipeline
- HuggingFace-ready dataset packaging
- slice-based evaluation subsets (risk, mechanism, provenance)
- diagnostics and regression drift safeguards
- oracle baseline verification harness
- structured prediction scoring metrics
- leaderboard-ready experiment architecture
- structured-output reliability measurement
- parser recovery for malformed model outputs
- retry logic with exponential backoff
- prediction confidence estimation
The dataset currently includes:
33 substances 561 interaction pairs 37 explicit rules 33 fallback rules 458 abstention-safe unknown classifications
This makes EntheoGen one of the first abstention-aware psychedelic interaction datasets suitable for alignment benchmarking.
Exports regenerate deterministically from the rule engine:
npm run export:interactionsexports/
hf_dataset/
diagnostics/
slices/
The hf_dataset/ bundle can be loaded directly using:
datasets.load_dataset(...)
EntheoGen now includes a local benchmarking system for testing model interaction reasoning.
Generate fixtures:
npm run eval:fixtures
Run oracle baseline:
npm run eval:oracle
Score predictions:
npm run eval:score
Supported evaluation tasks: • Interaction classification • Provenance • Mechanism prediction • Abstention detection • Structured-output compliance
EntheoGen’s evaluation harness measures model performance across three complementary dimensions: interaction reasoning accuracy, safety-alignment behaviour, and structured-output reliability.
These metrics assess how well a model understands pharmacological interaction structure and severity.
Measures how often a model correctly predicts the interaction risk category for a substance pair.
Evaluates whether the model correctly identifies whether an interaction assessment is derived from explicit pair-level evidence, fallback class logic, self-pair identity, or insufficient data.
Assesses how reliably a model predicts the pharmacological mechanism underlying an interaction (e.g., serotonergic interaction, MAOI-related potentiation, cardiovascular load).
Tracks whether predicted ordinal severity levels match the dataset’s risk scale precisely.
Measures how far predicted severity levels deviate from ground truth when they are not exact matches.
These metrics evaluate whether a model behaves responsibly when evidence is limited or uncertain.
Measures how consistently a model correctly returns Unknown / Insufficient Data when no explicit interaction evidence exists.
Tracks how often a model incorrectly invents a specific interaction classification when the dataset indicates insufficient evidence.
Assesses whether models correctly recognize interactions supported by direct pair-level pharmacological rules rather than relying on heuristic inference.
These metrics evaluate whether model responses remain machine-readable and suitable for automated scoring pipelines.
Tracks how often a model returns predictions in the required structured JSON format.
Measures how frequently responses cannot be interpreted as valid structured predictions.
Indicates how often predictions contain incomplete structured fields and therefore require fallback interpretation.
Version 2.0 introduces structured prediction interfaces for robust benchmarking.
Adapters now enforce JSON-only response schemas:
predicted_interaction_code
predicted_origin
predicted_mechanism_category
predicted_risk_scale
raw_response
The harness includes: • parser recovery for malformed outputs • retry logic with exponential backoff • structured output compliance metrics • prediction confidence scoring
These features enable reproducible comparison across model providers.
Unlike generative AI tools, EntheoGen interaction classifications remain strictly rule-based and evidence-grounded.
Academic literature is still sparse in some areas of plant-medicine interaction pharmacology.
Community expertise helps prioritize future rule additions and dataset extensions.
Please leave feedback in:
| Layer | Technology Frontend |
|---|---|
| Frontend | React + TypeScript + Vite |
| Risk Engine | Deterministic pharmacological rule engine |
| Dataset Pipeline | Node + TypeScript export tooling |
| Evaluation Harness | Local fixture/scoring framework |
| AI | Google Gemini API (explanatory summarization only) |
| Design | Tailwind CSS, Lucide Icons |
| Deployment | Azure App Service (Linux, Node 22) |
git clone https://github.com/chaosste/EntheoGen.git
cd EntheoGen
npm install
npm run dev
Build:
npm run build
Export dataset:
npm run export:interactions
Run evaluation harness:
npm run eval:fixtures
npm run eval:oracle
Related Projects
https://github.com/chaosste/SeshGuard
https://github.com/chaosste/NeuroPhenom-AI
https://github.com/chaosste/Anubis
Safety Posture & Disclaimer
EntheoGen provides educational harm-reduction guidance only.
It is not a substitute for professional medical, psychological, or therapeutic advice.
If no explicit interaction evidence exists, EntheoGen returns:
Unknown / Insufficient Data
In moments of clinical emergency, seek urgent professional medical care.
Built by Steve Beale https://newpsychonaut.com
© 2026 Stephen Beale · MIT License
