For the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to page URLs.
Skip to content

Verified Actions Over the World Model

Why Concurrence treats a record read, a write, and a confirmation as separate checked steps, what each establishes, and how to verify an action yourself.

The Problem

In a healthcare conversation, three easy assumptions are wrong. A phone number that matches a record doesn't prove who is speaking. A world-model observation doesn't change the scheduling system. An agent saying "you're booked" doesn't mean the booking happened. Each assumption, left unchecked, leads to a disclosure or a confirmation the organization can't stand behind.

Our Design

Concurrence splits an action into steps that each produce their own evidence.

  1. Verify before reading. A caller matched by phone number or a supplied entity_id starts unverified. Built-in record reads return verification_required until the caller's full name and date of birth match exactly one record through world_confirm_patient_identity. Repeated failures lock verification for the conversation. After verification, built-in reads return only the caller's own record or a household member's that names them as guardian or under a power of attorney.

  2. Scope every write. Supported agent write tools receive a server-constructed write scope that limits the workspace and entity they can target, whatever identifier the model supplies.

  3. Declare what a write does. An integration endpoint declares its writes, such as booking, cancellation, reschedule, or message. Only a declared write that succeeded backs a claim that the action is done.

  4. Gate a write on a person where needed. In a text conversation, an integration with an approval policy, or an approval_required skill, parks the call until an authorized reviewer approves or rejects it. Approval is a separate record from the external result.

  5. Keep an unknown outcome unknown. A write with no answer that settles it, such as a timeout, reports "outcome": "unknown", and the platform doesn't send it again in that conversation.

  6. Confirm at the destination. The endpoint's own result, or a read-back from the target system, is the evidence of what changed there.

The world model holds source-linked observations and derives read views from them. It does not deliver observations to external systems: an external change happens only through a tool the agent calls.

What It Guarantees

  • Built-in record reads on a caller-facing session return no record before verification and nothing outside the caller's own or household records after it.

  • A write that got no settling answer isn't resent in the same conversation.

  • In a text conversation, an approval-gated call doesn't run before an authorized decision.

What It Doesn't Cover

  • Approval isn't completion. An approved write can still fail or have an unknown outcome at the destination.

  • Integration reads aren't gated. Patient reads through your own integration endpoints are not code-gated by caller verification. The agent is instructed not to share another person's details from any tool, which is model guidance.

  • Approval gates are text only. Voice calls, simulations, voice notes, companion skills, and trigger actions don't offer approval-gated tools.

  • Projections lag. A world-model write that was accepted may not be queryable yet.

  • No universal idempotency. Use the operation's documented idempotency key or the destination's own protection where you need it.

Verify It Yourself

  1. In a test conversation, ask for an appointment before verifying, and confirm the tool call returned verification_required. Verify with the synthetic caller's name and date of birth, and confirm the same read returns the record.

  2. Ask for another person's record after verifying, and confirm the read returns not_caller_record.

  3. Run the integration endpoint's test operation against a synthetic appointment, read outcome and wrote, and read the target record back independently.

  4. Request the same change in a conversation, then compare the tool call's result, the agent's statement, and the target record.

  5. Follow Verify an Integration Action for the full procedure, including a failure and an unresolved write.

Last updated

Was this helpful?