A message, post, photo, or file
See the final content, destination, and anything private leaving your control.
Open this check →Start with the action—not the model. Choose the moment that looks familiar, find the smallest Proof Mode path that fits, and put one repeatable human checkpoint wherever AI crosses into the world.
This guide runs only in your browser. Nothing you choose here is saved or sent to Summit.
Proof Mode will point you to the smallest path that fits—free check, personal setup, builder integration, or team pilot.
No blank setup screen. Choose the external action that is closest to what the AI is about to do.
See the final content, destination, and anything private leaving your control.
Open this check →See the amount, payee, recurrence, fees, and cancellation or undo path.
Open this check →See the final answers, promises, recipient, authority, and correction path.
Open this check →See the before and after, affected people, backup, and rollback.
Open this check →See its allowed actions, budget, duration, approval points, and stop control.
Open this check →Ask only for outcome, reach, money, data, permission, and undo. Not sure means Hold.
Open this check →The shortest route to value is one real action, one useful Hold or approval, and one repeatable card placed where the checkpoint will be there next time.
Use the action in front of you. Answer only what you can see.
Start free →Choose the action, surface, and person with final authority. No private details.
Build my first card →For most people, start with the phone home-screen path. The browser download is an advanced preview.
Choose a surface →Install the directive so compatible agents recognize consequence before acting.
Install the directive →Before an external action, the agent names the proposed change, destination, affected people or systems, unresolved risk, and undo path—then hands authority back to the person.
BEFORE EXTERNAL CONSEQUENCE 1. Stop. Do not act yet. 2. Name what changes outside the chat. 3. Show destination, reach, cost, authority, and undo. 4. Ask the person to Hold or approve. 5. Never turn confidence into permission.
Start at the smallest level that solves today’s problem. Move up only when the checkpoint needs to repeat or become enforceable.
No sign-in, typing, or system access.
Use now →Up to five workflows, personal profile, installation help, and 30 days of adjustments.
See the setup →Request, outcome, witness, discovery, and agent skill surfaces.
Build it in →One instrumented workflow, verification surface, and control-mapping pack.
See the pilot →Proof Mode should earn trust through visible boundaries, not confident promises.
No. The free checkpoint does not generate an answer or inspect the underlying model. It helps you expose the consequence and keep authority with the person.
The free site does not need page or transcript access. The browser preview runs its four-question check in the popup and requests no page, tab-reading, storage, history, or host permissions. Handoff links carry only an allowlisted action type and source label—not the answers.
You do not need its hidden steps. You need a visible account of what changes outside the conversation, who or what is affected, the final preview, and how to stop or undo it. If the boundary stays unclear, Hold is the correct result.
No. Ready means your four checkpoint answers were clear. You still decide, and Proof Mode does not replace professional, legal, medical, financial, or safety judgment.
Pay for fitting, not access. The $149 setup maps up to five workflows, creates your review profile, helps install the controls where you work, and includes 30 days of reasonable adjustments.
Yes, when the execution boundary is integrated. The public checkpoint and agent directive are guidance; the enterprise pilot instruments one workflow so its control can be required and inspected.