# Solo AI Lab — experiment card

Status: TEMPLATE_NOT_RUN

This is a blank planning resource, not a measured case or a completed product
test. Instructions and empty fields are supplied; no customer, revenue,
conversion, or performance results are implied. Keep credentials and personal
customer data out of this file.

## 1. State one decision

Experiment title:

Decision this experiment should inform:

Responsible role:

Planned start / end:

Hypothesis:

Write a falsifiable claim. State the task, intended audience, and observable
change. Choose one primary outcome before collecting results.

## 2. Set the boundary

Included task:

Excluded tasks:

Permitted data and source of permission:

External tools or services involved:

Actions that need explicit human approval:

Recovery or rollback procedure:

## 3. Record the evidence

| Claim or assumption | Primary source URL | Checked date | What the source supports | Remaining uncertainty |
|---|---|---|---|---|
| | | | | |
| | | | | |

Separate a vendor's description from your own observation. Reopen consequential
sources before relying on them. Label an estimate as an estimate.

## 4. Choose a small test

Starting state or comparison method:

Inputs and sample selection:

Success condition:

Failure examples to include:

Maximum duration, attempts, and spending:

Stop condition:

Human review checkpoint:

Write the stop condition before the run. Include a way to stop external actions
and recover data. A persuasive demo alone is not a successful test.

## 5. Keep estimates and observations separate

| Measure and unit | Planned estimate | Observed result | Evidence location | Limitation |
|---|---|---|---|---|
| Accepted outputs / attempted outputs | | | | |
| Total cost and currency | | | | |
| Human review time | | | | |
| Errors or exceptions | | | | |

Attach the original cost assumptions and provider usage records when applicable.
Include failed attempts and correction time. Do not replace missing observations
with a forecast, a vendor claim, or an invented number.

## 6. Decide after review

Actual execution status: NOT_RUN

Evidence reviewed:

Outcome against the original condition:

Unexpected effects:

Decision: [ ] Continue narrowly  [ ] Revise and repeat  [ ] Stop

Reason and next permitted action:

Review date:

Keep NOT_RUN until execution has actually happened. A saved file, a working
preview, and human acceptance are different milestones. Preserve the original
card when repeating the experiment so that a later result cannot silently change
the initial hypothesis or success condition.