Controlled Generation: Build Three Options That Teach You Something
Generate options that answer one musical question. The goal is not three random songs. The goal is a traceable comparison that helps you make the next creative decision with evidence.
Enter with a project direction and one uncertainty.
If you came from a reference track, carry forward the completed #35 Reference Analysis & Production Brief v1—not the reference song as a copying target. You should already be able to explain your project purpose, protected identity, and the one production relationship you want to test.
You need
- a clear project purpose or Song Direction Brief
- an explainable source prompt or brief
- one answerable musical question
- one variable worth testing
- a freeze list of what should stay stable
You are not doing yet
- choosing the winner
- diagnosing every weakness
- rewriting the entire prompt
- generating endlessly for novelty
- treating every variation as evidence
Separate requested change, drift, and discovery.
Requested change
The musical difference you deliberately asked to test.
Uncontrolled drift
An unrelated change the model introduced that may distort the comparison.
Useful discovery
An unrequested result valuable enough to archive and explore in a separate future test.
Do not listen only for “which one sounds better.” Listen for the job the variable was supposed to affect: project identity, emotional weight, groove and pocket, structure and energy, performance, hook value, arrangement density, or another explicit requirement from the brief.
Same song. One relationship changes.
| Version | Controlled direction | What to observe |
|---|---|---|
| A | Intimate single lead | Does the chorus preserve clarity and intimacy but lack enough payoff? |
| B | Wider call-and-response support | Does the communal feeling grow without masking the lead or words? |
| C | Large stacked choir | Does scale increase at the cost of the grounded identity of the project? |
None is automatically the correct answer. The right continuation depends on the song’s job, not which version sounds most expensive or dramatic on first listen.
Build the A/B/C test in six passes.
Choose one answerable question.
Use: I need to learn whether ______ will better support ______. If the sentence hides several decisions, narrow it.
Name one variable and predict the evidence.
Examples: tempo feel, groove, lead delivery, supporting vocals, opening texture, signature instrument, chorus scale, density, register, transition behavior, ending treatment. Write what you expect to hear before generation.
Write the freeze list.
Protect the project purpose, lyric version, hook/focal identity, core structure, source brief, lead identity/performance target, model/version/settings where available, and other decisions that should not drift.
Prepare A, B, and C.
A is the baseline. B changes the selected variable in one direction. C changes the same variable in a meaningfully different direction. Do not introduce a new genre, vocalist, lyric and structure at the same time.
Generate close together and label immediately.
Use the same platform, model/version, lyric version, settings and source brief where possible. Label each output before preference takes over: Project — Test ID — Variable — A/B/C — Date.
Observe before choosing.
Listen once and write what appeared, what drifted, and whether protected traits survived. Preference belongs mainly in #37 Candidate Selection.
Classify the test before you trust it.
VALID TEST
The intended variable is audible, protected decisions remain usable, and unrelated drift does not materially distort the question.
USABLE WITH DRIFT
The intended variable is audible, but unrelated differences must be acknowledged during comparison.
INVALID / REGENERATE
The variable is missing, protected identity collapses, or drift changes the question so much that the result is poor evidence.
DISCOVERY
A result may be invalid for this test yet artistically valuable. Save it separately instead of contaminating the current comparison.
Controlled Generation Record v1
Use this as one connected working tool. Define the test before generating, document A/B/C after listening, record drift and validity, then save or export the completed record. How to Use Jack Righteous →
Draft data is stored only in this browser on this device. Export opens your browser print dialog so you can save this completed tool as a PDF.
1. Project and test setup
2. Question, variable, and hypothesis
3. Freeze list
4. A/B/C evidence
A · Baseline
B · Direction 1
C · Direction 2
5. Advance decision
Advance only when the comparison can support a decision.
- one narrow test question is explicit
- one variable is named
- a hypothesis and predicted audible evidence were written before generation
- a freeze list protects the project centre
- A, B and C test the same category of decision
- platform/model/settings are documented where available
- observations are separated from preference
- uncontrolled drift is recorded
- every option has a validity judgment
- at least two options are comparable enough for candidate selection
- the Controlled Generation Record v1 is complete
Know what #36 does not own.
#36 generates comparable evidence. It does not choose the winner, diagnose every weakness, or begin unrestricted revision.
#37 · Candidate Selection
Choose which candidate deserves to continue.
#38 · Build Diagnosis
Separate symptoms from likely causes after selection.
#39 · Bounded Revision
Test a limited corrective hypothesis while protecting what already works.
Supporting resources should solve the current blocker only. The AI Version Strategy Kit can support reusable testing, while the current AI Music Guides provide platform-specific implementation without replacing this transferable workflow.
You now have evidence, not a winner.
Continue to #37 Candidate Selection and decide which version deserves the next stage before diagnosing or revising it.
Updated September 17, 2026 · Production Intelligence 36/100 · Generative outputs can vary even under repeated inputs; compare rendered results and document what actually changed.
