DECISION-FIRST AI EXPERIMENT
Ahmadreza Samadi | Start with the decision.
https://a-samadi.com/about.html
Companion to https://a-samadi.com/tools/before-you-build-ai.html

FICTIONAL PLANNING EXAMPLE — NO EXPERIMENT WAS RUN
All answers below are copied from the existing fictional card.
The six-report plan is illustrative, not a sample-size recommendation.
Read Reading-Notes.txt for source-use and interpretation limits.

Save a private copy and type below each prompt. Keep unknowns explicit.
This text file sends nothing and contains no executable code. Your editor
and storage settings determine where your saved notes go.
Proposed planning aid, not a validated assessment. Prepared with AI assistance.

Who uses the result?
Fictional researcher reviewing public sources

Experiment ID / date
FICTIONAL PLAN / 17 Oct 2026

1. Decision
What choice should the output help someone make?
Trial AI for private summaries
of public housing reports?
Decide on a supervised pilot only.

2. Action and fallback
What follows the decision? What stays with a person?
A person opens sources and checks every claim.
Unverifiable claims return to manual work.
Nothing publishes automatically.

3. Evidence and gaps
Sources, dates, permitted use and what remains unknown.
Fictional setup: public reports; permitted use unverified.
Record publisher, URL, date and page before starting.
Actual sources and coverage are not yet selected.

4. Current baseline
How is this done without AI?
Read reports and write
summaries manually.
Measure effort first.

5. Acceptance criterion
What counts as an accepted result?
All claims source-supported;
no material caveat omitted.
Reviewer accepts or rejects.

6. Review and stop rules
Who checks errors? When must the experiment stop or hand back?
Researcher checks every output before use.
Stop for confidential input or unclear reuse rights.
Reject unverifiable claims; log failures and retries.

7. Test before acting
Cases, comparison, time and full costs.
Plan 6 reports; same brief.
Log attempts, effort, costs.
No pilot run yet.

8. Observed result / next decision
Record evidence; unknown stays unknown.
Unknown: no pilot run.
No savings or quality data.
Next: select eligible sources.

Keep original sources, acceptance decisions, failures and full cost records
with your private experiment. This card does not establish fairness,
correctness, savings or permission to use a source.
