Known
- The local setting is reversible.
- The active policy requires an exact host grant.
- One external-call unit is reserved for reconciliation.
Draft Agent Runtime Profile · complete simulated reference loop
Take the driver's seat
You are sitting where an AI agent sits. It starts from a bounded view of the situation, holding only what this task needs, and sees only the actions open to it right now. Authority, assurance, cost, unknowns, and evidence stay visible before any plan exists.
This is a teaching simulation. The situation, moves, and costs are fixed in the page, every effect is simulated, and nothing leaves your browser.
accretive@0.1.0 metered simulated calls no live effects
One view, six answers
situation.demo.releaseWhat do I need to know before acting?
Unknown
- Whether dispatch will be accepted
- Whether the postcondition will be observable
Conflicts
Speed preference conflicts with the firm proof-before-completion boundary.
Authority
local read
No authority is inferred from capability discovery.
Budget
- Wall time
- 2000 ms
- Tokens
- 3000
- External calls
- 2
- Human interruptions
- 1
Available moves
1/ 3
Moves available now under the current authority; the others need a host grant or authority this demo does not have.
One conflict stays open.
The preference for speed still clashes with the rule that nothing counts as done until it is proven. The view records the clash rather than settling it, and the agent can raise an objection before it plans, acts, proves, or learns.
Capabilities become situated choices
Ceiling: reversible writeChoose an action to inspect
Proof stays plural
What has actually been established?
Each kind of assurance keeps its own axis. A schema digest says nothing about who published the schema, and neither settles whether the result can be trusted.
Integrity
The schema digest and local candidate lineage are declared. Nothing in this demo checks them.
Authenticity
No publisher signature is loaded.
Authority
Local reads are present. Writes remain gated.
Execution
No effect has crossed the simulated effect boundary.
Completion
Nothing yet proves the release channel was actually set.
Trust
No one has judged whether to trust the result, and integrity and authority alone cannot settle it.
Complete simulated reference loop
Phase: orientedPlan, authorize, act, recover, prove, and learn
Pick a failure to inject, or none, then use the buttons below to walk one change through the loop. Every effect is simulated, and no scenario grants real authority.
Artifact chain
- situation viewcurrent
vcp:artifact:situation:situation.demo.releaseBounded orientation root
Event log
1 situation.compiled
Cursor: cursor.demo.1
external calls
external calls
calls avoided
call remaining
Spent so far counts both runs. The baseline is what those two runs would cost without a saved procedure, so the difference appears once the second run is done. These numbers come from a fixed teaching trace. They say nothing about latency or model quality in production.
Bounded reproduction
It records this run so it can be replayed, and leaves out live grants, credentials, and raw private context.
Portable, inspectable, replayable
The exact SituationView artifact
The JSON validates against the draft Agent Runtime schema and names every artifact it depends on.
sha256:920b308c2438d47ac234240fc35e088afa4dd6621e3e1884da993db9dbce1e01This is an educational, source-level reference simulation. It carries no live grant and makes no deployment, publication, ratification, or production assurance claim.