A grid token is scoped. With a grid operator token, writing requires the owner's consent (FlexMgmt GetSettings.WriteAccess); without it, commands answer 403 and are journalled as rejected, while releases (NORMAL, an inactive RestrictPower) are always accepted. Withdrawing the consent ends every command in force or pending at the next evaluation cycle. With a simulation token, the token is the consent (WriteAccess true): each command lasts one hour at most (the declared parameters in the evidence status say so), is marked simulated, answers 409 while a real grid command is in force, and ends when one arrives.

OpModeLoadCmd returns the last command; OpLoadState returns the mode actually enforced. They differ when MaximumLockTime has elapsed, when a restriction is deferred by MinimumRunTime, after 24 h without a new command (fall-back to NORMAL), and while the engine does not apply commands to real devices (engine offline or in simulation, observe-only): the command is then recorded, journalled as not_enforceable, and takes effect if the engine applies again within its lock window. The evidence status says which (apply_enabled, not_applying_reason).

RestrictPower bounds the import at the connection point (MaximumPowerKw), never below the declared MinimumLoad; a restriction down at MinimumLoad is a lock and lasts at most MaximumLockTime. The feed-in bound (MinimumPowerKw) needs PV control, declared in GetSettings.PVControl. MAX is not in force (OpLoadState NORMAL, journalled not_enforceable) while a higher-priority signal, such as an active restriction, outranks it.

Writes carry the value as a query parameter (OpModeLoadCmd=..., RestrictPower=<JSON>): sgr-commhandler up to 0.5.2 never sends the requestBody of a REST write call. A JSON body with the same field is accepted too.

Evidence: GET /api/sgr/evidence/status and /api/sgr/evidence/events?after_seq=N (same credentials) return declared parameters and the journal of commands, decisions and device commands, correlated by id.