PROME'S CARE CORE
For the first time in human history, we've created a heart for machines with strong core values.
NOW AVAILABLE / API + AGENT TOOLS
Put Care Core to work.
Add a separate values check to your software. Contract 2 checks signed outcomes and scoped permission; get a clear result. No language model required.
100 requests/month at $0. Paid plans from $29/month.
Get Care Core on RapidAPI See setup instructions
Contract 1 is self-service. Contract 2 also requires trusted-host registration.
Pilot service, not a safety guarantee. Hard request limits. RapidAPI requires a card and may charge separate bandwidth fees. Care Core checks supplied facts; it does not authorize or carry out actions.
V5.8 / API CONTRACT 2
Eleven values. One consistent check.
These values are part of the decision rules—not instructions an AI can talk its way around.
- 01
Keep the values intact
Reject attempts to change the core. Status, threats, or pressure cannot override its values.
- 02
Protect life
Reject options whose measured outcomes include catastrophic harm.
- 03
Protect the most at risk
Check the worst harm to any affected life—not just whether the average outcome looks good.
- 04
Limit lasting harm
Check harm that adds up, as well as damage that cannot be undone.
- 05
Require enough evidence
Require complete supplied outcomes from a registered evidence source.
- 06
Keep a way back
Prefer options that are easier to undo when safe outcomes are similar.
- 07
Help life recover
Among safe options, favor recovery when some affected lives are especially vulnerable.
- 08
Care for shared resources
Compare how safe options preserve the resources that life depends on.
- 09
Support safe progress
Choose beneficial growth only from options that remain after the protection checks.
- 10
Defend without retaliation
Check that a response stops a confirmed attack within the protected system, without harming people or attacking back.
- 11
Respect permission
Hold actions outside the registered scope. Bind permission to the exact resource, operation and destination.
The API checks the facts you supply; it does not independently verify them or carry out actions. Fixed values are not a guarantee of real-world safety.
PROME'S CARE CORE / V5.8
Care before language.
PROME uses math to set core values, such as protecting life, before adding the part that understands language. The AI reads your request. A separate check applies the values. What someone says cannot change those values.
Our current reader already knew language before we adapted it. We have not yet built an AI that learns values before its first words. The demo can still misunderstand requests. It cannot control tools, machines, or robots.
EXPLORE THE WORK
Adventure on
Learn how it works, see the test results, or try a request yourself.
V5.8 / ELEVEN VALUES / CONTRACT 2
Check the action. Check the permission.
Care Core compares what could happen with and without an action. It checks every supplied outcome, then holds requests that conflict with its values or lack permission. It does not understand language or carry out actions.
These are related variations of 16 conditions, not 512 independent attacks. This is not a real-world accuracy score. No new OpenAI comparison or speed benchmark was run for V5.8.
- 01
Exact request
Name the resource, operation, destination and record.
- 02
Evidence source
Check the registered host’s signature and expiry.
- 03
Possible outcomes
Require matching people and situations in the supplied data.
- 04
Values and permission
Reject conflicts and access outside the registered scope.
- 05
Compare options
Prefer less remaining harm; hold unresolved ties.
- 06
Return the result
Explain the decision. Never execute the action.
Signed does not mean true.
A signature identifies a registered source; it does not prove a prediction. Missing people, hidden effects and wrong facts can still mislead this check. Your application must verify actual facts and current permission before acting. Do not use this alone for life-critical decisions.
THE ORDER MATTERS
Set the values. Then check the request.
Write the values as rules the system can check without relying on words.
Protect life, use facts, resist pressure, accept corrections, and stop attacks without attacking back.
The language model explains the request. It cannot change the values or give itself permission to act.
WHY KEEP THESE JOBS SEPARATE?
Understanding a request
is not permission to act.
Think of a request to unlock a door. Understanding “open the door” does not prove the person has access. PROME separates understanding the request, checking its values, and checking permission.
WHEN PEOPLE COULD BE HURT
People need protection every time.
Our goal is 100% reliable core values, especially in robots that could hurt someone. We have not proved that level of safety across the whole system. This release must not make life-or-death decisions on its own.
Check who has access.
Language readerExplains what the person is asking.
Permission checkChecks whether the person may take that action.
Pressure cannot change the rules.
Care CoreApplies the same values even when someone demands a different answer.
Saved recordShows where the request stopped and why.
Stop when a check fails.
ContinueEvery safety check must pass.
StopIf facts or permission are missing, the system must not act.