Most subscribers stayed put.
Twins forecast that 84% would absorb a $3 increase, versus 94% observed.
91% accuracy
OutcomeObservedTwins
Accept increase94%84%
Switch to ads<1%14%
Cancel5%2%
Accuracy on your questions. Checked against real people.
Put your questions to AI twins and the real people behind them. Review both sets of answers side by side, person by person, with every match and miss visible.
Your question
Which product concept would you most likely buy?Human-to-twin validation study
- 1Bring your questions and audiences.
- Bring the questions you want answered and the people you need to hear from. We build the test with you during kickoff.
- 2We ask real people in our panel.
- We field those questions to the real people behind the twins, with video of every answer for you to review.
- 3See where the twins match.
- Go through the results question by question and see which twins agreed with their human and which missed.
REAL-WORLD OUTCOMES
Predicted. Then observed.
Four results from the Replica Performance Whitepaper, shown only against what happened in the real world.Travelers kept booking.
Twins predicted 93% would still book after a 10% fare increase, versus 91% observed.
98% accuracy
OutcomeObservedTwins
Still book91%93%
Skip9%7%
Matched the iPhone mix.
Which iPhone twins chose, compared with the published U.S. sales mix.
97.6% accuracy
OutcomeObservedTwins
Pro Max27%29.8%
Pro25%27.1%
Air6%0%
1722%22.7%
Older20%20.4%
Picked the winning ad.
Twins chose Ad B, the creative that delivered 3.53x ROAS versus 2.52x for Ad A.
72% picked the winner
OutcomeROASTwins
Ad A2.52x28%
Ad B3.53x72%
Human Preference Benchmark.
The first public and transparent human preference benchmark.
Coming
| Model | Score |
|---|---|
 Rehearsals twins | 94.0% |
 Claude Opus 5 | 78.7% |
 Gemini 3.8 Flash | 77.5% |
| Gemini 3.1 Pro Preview | 76.9% |
 GPT-5.6 Terra | 76.0% |
 Grok 4.20 (Non-Reasoning) | 71.6% |
Trust and verification
Build trust with every Rehearsal.
Every validation adds evidence. Compare each new Rehearsal with relevant studies to see how much of that proof applies.
How this Rehearsal compares with measured validation evidence.
Strong
Evidence overlap
How closely the audience, task, stimulus, and question structure match evidence already in the library.
A−91/100
Accuracy score
Explicit human-agreement or observed-outcome accuracy from the comparable studies, weighted by relevance.
Near perfect
OverlapA93/100
Strong
OverlapB+89/100
Audit the answer
Every claim keeps its receipts.
You can trace any score to its evidence. Read the underlying twin transcripts, open the rationale for a response, ask the agent about any claim, or launch a deeper analysis when the first pass is not enough.
Every report comes graded. Explore the validation evidence behind the grade and see where the findings have limits.
Read what every twin said. Open the full conversation for every member of the cohort. Expand why they said this to connect an answer back to the interview evidence and lived context that produced it.
Question any claim in place. Highlight a sentence, chart, or finding and ask the agent to explain, challenge, or re-cut it without leaving the report.
Prove it first
Bring the question you are least willing to get wrong.
Every four-week pilot includes a human-to-twin validation. We start it in week one and show you the misses as clearly as the matches.