What you’ll make work
Configure a matched-control comparison, inspect two observations for each candidate, and make a follow-up request for researcher review.
New to AI agents? Start with seven short foundation lessons on models, tools, feedback, evidence, and boundaries. Then return here to apply the ideas.
Each run connects a goal to a tool result, a check, and a next action. You change a meaningful setting, watch what changes, then apply the idea to a new case.
- Goal
- Find the largest observed increase over its matched control.
- Tools
- Read the fictional observations and calculate matched differences.
- Check
- Use both observations and limit the claim to what they support.
- Stop
- Request another controlled observation for researcher review.
This is a guided simulation: its authored controller selects from a bounded set of actions, while its tools execute calculations and checks. A live model would choose actions dynamically.
Original worked example · fictional scenario
The higher reading can hide the smaller change
The fictional assay measures signal points. Each candidate reading has its own matched control. These points have no clinical meaning, and the dataset contains no real substances or patients.
| Observation | Candidate reading | Matched control | Difference |
|---|---|---|---|
| A1 | 30 | 29 | +1 |
| A2 | 31 | 30 | +1 |
| B1 | 24 | 19 | +5 |
| B2 | 26 | 21 | +5 |
compare_matched returns raw means of 30.5 for A and 25 for B, but mean matched differences of +1 for A and +5 for B.
What happens nextThe comparison changes the preferred follow-up from A to B. The agent requests another controlled replicate and pauses for researcher review.
Subtract each reading’s own control before comparing the candidates. A starts from a higher control value, so its larger raw reading does not establish a larger increase. B has the larger observed difference in these four rows. With only two observations per candidate, this exercise does not establish statistical significance or a health benefit.
Common mistakes worth catching
Ranking raw readings
Check the starting comparison for each result. A reading of 30 against 29 represents a smaller increase than 24 against 19.
Turning an observation into a treatment claim
The fixture contains no evidence about people. The next useful output is a follow-up request, not a claim that a candidate works in patients.
Choosing the one row that looks best
Include both supplied observations. Agreement across these rows is useful to inspect, but it still does not replace further testing.
What to take with you
- Specify the comparison before ranking the evidence.
- A tool result can change an agent’s next action.
- A promising observation and a verified conclusion are different achievements.
Finishing the activity gives you a record of the configuration you changed, the case you tested, and the result you observed. The new case checks your understanding separately from the guided run.
A few good questions
Do I need biology knowledge?
No. The activity supplies every number and rule. You compare fictional signal points and decide how much the evidence supports.
Does a larger difference prove that B works?
It supports only the narrower statement that B has the larger observed difference in this fixture. Further controlled observations and researcher judgment are still needed.
What if both candidates improve by the same amount?
Report a tie in observed differences. A higher raw reading alone does not break that tie; request more relevant evidence.
Research behind this lesson
Google DeepMind · May 19, 2026
Generates, critiques, compares, and refines hypotheses with scientists guiding and reviewing the work. Co-Scientist: a multi-agent AI partner to accelerate research ↗
This imaginary signal-point exercise borrows a process. It reproduces neither the research system nor a biomedical finding.
This school scenario is a simplified simulation. The research connection explains the method that inspired the activity; it does not establish this lesson’s effectiveness or imply an endorsement.
Content version: 2026-09-12.1 · Prepared September 12, 2026
Editorial review pending · No completed specialist review or review date recorded.