Does adding water always help?
Learn to choose a controlled experiment and explain what its result can—and cannot—tell you.
The walkthrough uses an existing real IMD proposal and its recorded synthetic-world measurement. Your prediction stays in your tab; nothing is purchased or rerun.
See what the measurements, model calculations and external provenance each establish →
Your first research question.
Make your prediction
Read the recorded first observation: growth is 50/100 without water. Predict what will happen if only water changes.
Notice the assumption you are making before seeing the answer.
Check the evidence
Follow the original IMD proposal, then reveal the measurement and inspect which explanations still fit.
Separate the agent’s proposal from the engine’s measured response.
Choose the next question
Choose a new controlled pair for a different mineral or light setting, then explain what comparing the two water contrasts could establish.
A single contrast cannot settle every interaction or establish a universal law.
Try it in your own free world
Open the island, start or continue a local expedition, and inspect its journal. Use Your experiment to choose a test. When one model remains, test its four held-out predictions and reveal the committed configuration.
40 internal game credits; 2 per training experiment. No wallet or IMD tokens are needed for local play.
A new local world has different hidden rules. Its results are not a continuation or repetition of the historical IMD request. If your previous world is complete, New expedition offers a separate new world. The recorded IMD archive is never reset by this guide.
For agent builders
Inspect the public observation → proposal → measurement boundary. Compare a controlled hypothesis test with selectors that already know the candidate family; then use the public engine to evaluate your own experiment policy.
This is a small synthetic testbed, not evidence of real-world scientific discovery or a general AI benchmark.
Now choose the next test.
After the recorded walkthrough, design a new controlled pair and explain what it could tell you. You receive feedback on your choices, without running an experiment.
Inspect the contract route in a free test.
A separate local EVM run shows forecast, contract order, mock callback, measurement and revised explanation across three rounds. It uses a deterministic mock provider, not IMD.
Show solution and IMD/local comparison
Spoiler: this section reveals the recorded measurement. Make your prediction first if you want to test your reasoning.
IMD matched the next local choice.
The recorded IMD agent and Kepler both chose to add water while holding mineral, light and temperature fixed. Ada and Lin would change mineral and temperature. All four choices have the same expected model separation under the local reference family.
| Researcher | Proposed experiment | Expected models left¹ | Recorded result |
|---|---|---|---|
| IMDHistorical proposal | Water · no mineral · amber light · warmChanges: water | 11 | Growth 50/10019 → 14 models · 5 excluded |
| Ada | No water · basalt · amber light · coolChanges: mineral + temperature | 11 | Not run separately |
| KeplerNext local turn | Water · no mineral · amber light · warmChanges: water | 11 | Not run separatelySame conditions as recorded IMD treatment |
| Lin | No water · basalt · amber light · coolChanges: mineral + temperature | 11 | Not run separately |
What the local selector knew
Local strategies receive the public observations and know the full 24-model family and response function.
No external payment. A local training experiment costs 2 internal game credits.
What IMD received
The original IMD request supplied the hypothesis, one observation and available condition combinations. It did not include the model family or response formula.
Historical external cost: 0.5 IMD + 0.0000157250756555 ETH aggregate gas, plus 2 internal credits for the simulation.
Useful for the question: the water-only contrast tests whether water improves growth in these fixed conditions. The recorded response did not improve, ruling out five explanations. It leaves water/mineral/light interactions unresolved.
This is one retrospective choice comparison, not evidence that IMD is better or more cost-efficient. Local rows are proposals recalculated from the original first observation; no new experiment or payment was made.
¹ How expected separation is calculated
Assign equal weight to the 19 still-compatible local models. For each proposed test, predict the response under each model and group identical outcomes.
- Growth 0/100: 2 models would remain.
- Growth 50/100: 14 models would remain.
- Growth 100/100: 3 models would remain.
The model-weighted average is (2² + 14² + 3²) / 19 = 11. The actual recorded growth was 50/100, so 14 models remained. Expected counts are retrospective calculations from the public family; they are not measured outcomes or IMD-provided probabilities.
Local selection order: Ada → Kepler → Lin. Ada performed the original control, so Kepler is the reconstructed next local turn. Each local strategy uses the production selector and the same available training conditions.
Download the full comparison JSON · Reproduce from source and the original record