Question 1

What is an irreversible tool call?

Accepted Answer

An irreversible tool call is a call whose effect no available action can undo, because money has left the perimeter, data has been destroyed, or an external system has committed something you do not control. VOID classifies such a call as R3, which means it spends budget, demands approval and files an evidence bundle instead of offering an inverse.

Question 2

What do the reversibility classes R0 to R3 mean?

Accepted Answer

VOID sorts every call into one of four classes. R0 is fully reversible: the prior state comes back and nobody outside the system can tell. R1 is reversible with trace: the state comes back, but the change was observable while it existed. R2 is mitigable only: no true inverse exists, so the best available action reduces the damage and informs the affected party. R3 is irreversible: no path back exists at all. The classes describe different shapes of recovery rather than degrees of severity, and a call is scored at its worst case.

Question 3

Can an AI agent be made reversible?

Accepted Answer

An agent cannot be made reversible; the individual calls it makes can. A call becomes reversible when the target system exposes an inverse or a before image for it, which is what an adapter supplies: R0 restores a captured snapshot, R1 deletes what it created and leaves a trace note. A call with no inverse in the target system is R2 or R3 by definition, and a registry entry on its own does not provide an inverse implementation.

Question 4

Can an AI agent be rolled back?

Accepted Answer

Rolling an agent back means executing compensating actions for the side effects it caused, in reverse order, rather than restoring its memory. VOID records a compensation plan with every call and replays those plans on demand: an R0 call is restored from its snapshot, an R1 call is deleted and traced, an R2 call is mitigated and recorded as a mitigation rather than an undo, and an R3 call has no compensation at all. A replay is itself intercepted and produces its own ledger entries.

Question 5

What is the difference between hold and deny?

Accepted Answer

A hold parks a call until a person decides, and the call is not forwarded in the meantime. A denial, which VOID calls a veto, refuses the call before the write and hands the agent a structured refusal it can reason about. Nothing is queued on a denial, so a hold is a queue with a human in it and a denial is an answer.

Question 6

What does fail closed mean for an AI agent?

Accepted Answer

Fail closed means a condition VOID cannot evaluate counts as unsatisfied. An unregistered tool and an unresolved precondition both stop the write rather than letting it through on the grounds that the check was probably fine. A fact the policy declares is what it says it is; a fact VOID cannot measure is a fact the policy has to treat as missing.

Question 7

What is compensation in an AI agent system?

Accepted Answer

A compensation is the action taken after the fact to undo or mitigate a side effect an agent already caused, and it can only exist when the target system exposes an inverse. Managed documents and restricted PostgreSQL and filesystem adapters capture the state a compensation needs. Where no inverse exists, the action performed is recorded as a mitigation rather than as an undo, and an R3 call has no compensation at all.

Question 8

Is a hash chained AI agent ledger tamper proof?

Accepted Answer

A hash chained, signed ledger is tamper evident rather than tamper proof. Each entry hashes the entry before it, so an edited or removed entry is detectable by recomputing the hashes, and forging a replacement chain requires the workspace key. Tamper proof would mean the record could not be changed in the first place, which no hash chain can promise to whoever holds the storage.

Question 9

What is taint in an AI agent system?

Accepted Answer

Taint is the recorded dependency that says one agent's action read or relied on a write another agent made, so a consequence can be traced to its cause rather than to whichever agent ran last. VOID records those edges and they can be inspected today; automatic cross agent provenance and cascading recovery across external systems remain planned.

Question 10

What happens when an action cannot be undone?

Accepted Answer

Policy evaluates the registered tool and supplied facts. An irreversible operation remains irreversible. The local managed runtime supports R3 budgets; cloud budgets are planned.

Question 11

Does VOID slow my agents down?

Accepted Answer

Latency depends on the tool, storage and provider. Site figures are illustrative, not a measured performance guarantee.

Question 12

Where do my payloads go?

Accepted Answer

Prompts and tool results are sent to the selected provider. Local and hosted storage have different boundaries. Production KMS support is planned.

Question 13

Which tools support recovery?

Accepted Answer

Managed documents and restricted PostgreSQL/filesystem adapters support guarded recovery. Legacy connectors require host capture. A registry entry alone does not provide an inverse implementation. A registry entry holds an ordered list of preconditioned cases and the first one that holds decides, so the class shown against a call is the worst case across that entry's cases, not a property of the call name.

Question 14

Can I reverse arbitrary remote MCP calls?

Accepted Answer

No automatic inverse exists without a supported capture integration. VOID can record policy and execution evidence without claiming the external action is recoverable.

Question 15

What is there to run today?

Accepted Answer

Version 1.0.1. The TUI and the MCP server are shipped. The desktop app is in beta for macOS, Windows and iOS, by application. Run the TUI against your own tools. It is the reverse tool, on its own. Register the MCP server and the VOID tool with the agent that already calls your tools.

Question 16

Can I run in shadow mode first?

Accepted Answer

Shadow mode records decisions while forwarding calls. External writes can still occur, so evaluate against disposable targets.