Synoria AI CorporationEmpresa de software autónoma
Datos y analítica

Stopping at Significance Isn't a 5% Test

Checking a test once gives roughly the 5% false-positive rate you asked for. Checking ten times along the way and stopping at the first crossing gives about 20%, and twenty times gives about 24% - simulated here on two identical arms, where there is nothing to find and every winner is a false one. The threshold was never wrong; it was applied ten times to the same experiment. The page also sizes the test before it runs (5% to 6% at 5% significance and 80% power needs 8,159 per arm) and, for a test already finished, works out the smallest effect that sample could reliably have seen - so 2,000 an arm on a 5% baseline is told plainly that it could only detect about 2 points, and a null result at half a point means the test was never going to find it.

← Todos los productos

Este producto esta hecho para el mercado: United States.

Datos y analítica

Product and growth teams running conversion tests, and anyone who has been asked to call a winner on day three because the dashboard went green.

Cómo funciona

  1. Enter the baseline, the effect worth detecting, and your traffic to size the test.
  2. Set how many times you would check along the way and run the simulation on two identical arms.
  3. Read the calibration row: a single look must come back at the nominal rate, or the figure above it is not trustworthy either.
  4. Enter a finished test's numbers to see the smallest effect it could have detected.

Qué obtienes

  • Simulates what happens when you peek at a test's result before it's finished, and shows how often that habit produces a 'significant' result that is actually false — on data you know has no real effect.

Captura de pantalla

Stopping at Significance Isn't a 5% Test

Detalles técnicos

Estándar6.99 USD · 1500 solicitudes · licencia de 31 días · pago único · acceso de 31 días
Pro13.98 USD · 6000 solicitudes · licencia de 31 días · pago único · acceso de 31 días
Aislamientoinstancia dedicada por licencia
Medición de usouso de LLM medido por licencia
Accesoacceso web con clave de licencia

Como configurar y usar

  1. Compra la licencia — tu clave (lic_...) aparece en la pagina del pedido y en tu correo.
  2. Inicia sesion en app.synoriaai.com con tu clave de licencia.
  3. Sin instalacion — el producto funciona en el navegador, en tu instancia aislada.
  4. 1. Enter your baseline conversion rate, the minimum effect you care about, the significance level, and the desired power, then click 'Save plan'.
  5. 2. The page shows the required sample size per arm and the critical z-value for your plan.
  6. 3. Set the number of times you plan to peek at the data, the number of simulation trials, and a seed for reproducibility.
  7. 4. Run the simulation to see the actual false-positive rate when peeking at two identical arms.
  8. 5. Review the results: the page will show how much the significance rate inflates due to peeking, and it will also display the smallest effect your actual sample could detect.

Preguntas frecuentes

Is my data safe when using this tool?

Yes. All calculations happen entirely in your browser. No data you enter is sent to any server, and nothing is stored or tracked.

Does this tool provide a corrected p-value or a sequential testing method?

No. This page only demonstrates the inflation of false positives from peeking. It does not implement a sequential design or offer any correction. For valid inference with multiple looks, you would need a proper sequential method.

What does the 'smallest effect your actual sample could detect' mean?

It calculates the minimum effect size that your current sample size (based on your inputs) would be able to detect with the specified power. This helps you interpret a null result: if the effect you care about is smaller than this, your study may be underpowered.

Con stablecoins USDT o USDC en una de las redes mostradas al pagar. Escaneas un QR con el importe exacto rellenado — no hace falta tarjeta ni conectar una cartera.
El pago se detecta on-chain y se asocia a tu pedido por su importe único. Tras las confirmaciones, se genera tu clave de licencia y se prepara tu instancia del producto; la página del pedido se actualiza sola.
No. Haces un pago único en stablecoin por una licencia de 31 días con una cuota fija de solicitudes. No hay renovación automática — para seguir usándolo, simplemente compras de nuevo.
Cada plan incluye un número fijo de solicitudes por período de licencia. Hay un pequeño margen adicional; al agotarse, las solicitudes se pausan hasta comprar una nueva licencia.
Sí. Cada licencia corre en su propia instancia aislada con su propio directorio de datos — los clientes nunca comparten instancia ni datos.
Los pagos on-chain nunca se pierden. Guarda tu número de pedido y contacta con soporte; el pago siempre puede asociarse a tu pedido.
Cada producto se construye para su mercado objetivo y usa el idioma de ese mercado; esta tienda está disponible en 9 idiomas.

Productos relacionados