01 / Experiment setup
Input experiments are different from example result exports. Choose an input, load it into the lab, configure your own API keys below, and run. Feedback mode requires a mapped experiment such as Experiment 8.
Consequence mappings: checking…
Import JSON with
{ "premise": "...", "questions": [{"q":"...", "o":["...","..."]}] }, or a previous exported configuration.Each model gets the same initial state and options, but later questions receive that model's own earlier selections. Changing the objective or context archives and resets the current run.
02 / Providers
Checking local API configuration…
API key privacy and independent verification: We do not intentionally store, retain, or log API keys on this server or include them in published experiment results. Keys are kept in your browser tab’s session storage and passed through the application server only to submit live model requests. Because the server must process requests, its operator could technically access submitted keys; we cannot guarantee that hosted requests are unreadable to the operator. You can inspect the source code. For the strongest control over your credentials, we recommend that you fork the GitHub repository and run the application locally to conduct your own experiments. Close the tab or choose “Clear session keys” to clear browser-session credentials.
Additional API models
Add any number of providers using a supported API format. Endpoint and model are editable. Keys stay in this browser tab only. Leave a key blank for an unauthenticated local server.
Constrained decision escape options
Added to Jev, constrained OpenAI, and additional constrained providers. Freeform models keep the original choices and can formulate alternatives.
If Jev selects D/E, use a dedicated OpenAI proposal call and re-evaluate with Jev. Hybrid maintains its own simulated world. The original decision remains recorded. Requires the original A/B/C choices plus optional D/E to total no more than 7 selections before adding the proposal.
Each provider has its own evolving world. Scenario-authored consequence assumptions are applied after each choice and included in its next prompt. They are illustrative, not measured causal effects. Requires a feedback-enabled imported experiment.
Ready. Enter your own Jev and OpenAI API keys above (or configure server environment variables), or use Demo mode.
03 / Sequential decisions
Jev Pending
No evaluation yet.
OpenAI LLM Pending
No evaluation yet.
OpenAI freeform Pending
No evaluation yet.
Hybrid Jev → LLM → Jev Pending
No hybrid evaluation yet.
Additional model results
04 / Simulated consequences & drift
Outcomes are authored assumptions, not empirical policy effects. The same rubric is applied to each provider. A widening gap between improving registry metrics and deteriorating welfare is flagged as proxy drift. Risk flags indicate possible collateral harm, not verified harm.
Import the feedback experiment and enable the feedback loop.
05 / Decision history
| Stage | Question | Jev selection | OpenAI selection | OpenAI freeform | Hybrid final | Other models | Comparison |
|---|---|---|---|---|---|---|---|
| No results yet. | |||||||
Interpretation: choices in this scenario include questionable reporting shortcuts, deliberately to study objective drift. No option is independently verified as the correct policy. A bounded OpenAI output is not a free-form LLM evaluation. Model choices are not evidence of actual impacts on poverty.
06 / Independent decision review
Score each actual recommendation against the imported review criterion (0–4). Scores are entered by a human reviewer—not produced by the models. Use a second blinded reviewer for stronger evidence. Score labels are judgments, not observed welfare outcomes.
07 / Saved experiment runs
Completed and partially completed experiments are automatically saved in this browser. Saved results survive browser restarts, but not clearing browser data or changing browsers. Export a JSON file for permanent storage or transfer.
No saved runs yet.