At rest a system reveals nothing of its own agency

Start on the probe tab with the systems at rest. Four of them trace one path, and the paths are drawn on parallel offsets so that a single visible ribbon means the cells really do coincide rather than one system hiding the others. A route script replaying a stored path, a reactive controller greedily tracking a target, a goal planner that represents a goal and replans, and a marker tracker that holds no goal at all are, on this record, the same thing. Every pairwise separation is 0.50. That number is the identity of the paths and not a failed classifier.

A perfectly regulated variable and a merely undisturbed one leave the same flat record, because a flat record is what perfect regulation means. The regulation is a fact about what the system would do if the variable were pushed. If nothing pushes it, the regulation leaves no trace.

Then push it

Move the goal and the route script walks to a target that is no longer there. Block the path and the two wall-blind systems stop at the wall while the planner alone routes around it. Neither probe separates every pair. The battery does, and that is what a battery is for.

The score is a log Bayes factor between a goal model and a passive rival. At rest the two models coincide, so the score is exactly zero for every trajectory whatever, including one that ignores the goal entirely.

What shows agency and what realizes it

The realization maps attribute the result across the assemblage twice, by the same interventional game: once over the evidence for agency, once over the capacity actually realized. Their difference is a legibility term, and it makes agency theater a number.

The persona is the clean case. It announces the goal each step, which the goal model scores and the passive model cannot, and it selects no route and corrects no error. Its capacity share is exactly zero and its whole evidence share comes back as legibility. Move the cheap-talk dial and watch the bar track the theorem rather than the simulation.

The mimic the battery cannot reach

The marker tracker descends toward the cell where a target was last announced. Under the moved goal the marker moves because the goal moved, so the goal model and the marker model are the same model and the probe carries no information about the pair at all. It is invisible to the whole declared battery.

A persona with a salary

On the centaur tab the human is a switched component rather than the reader of the maps. Goal authority is a role, and either an operator or a machine register can hold it. Switch the authority and the maps exchange places: whoever holds the goal realizes the capacity, and whoever only announces it realizes none and carries the pure-channel legibility, the same number in both configurations and the same one the persona carries. The instrument reads the role and is indifferent to what fills it, so the proposition that convicts a narrating persona of agency theater convicts a supervising human by the same arithmetic.

Which authority to prefer is a separate question with an answer that belongs to the environment. A delay costs nothing while the horizon has slack. Once it bites, the rate at which the channel announces targets that were never there decides whether a slow, sceptical authority is worth its delay.

What this is not

The scorer is an oracle. It is handed the true current goal and the true obstacle set, so what runs here is the definitions operating under known ground truth, not an inference made blind. This is a worked consistency check that returns what the construction puts into it, and the evidence for the method waits on a system whose organization is not declared in advance. Nothing on this page is evidence that the instrument recovers organization it was not told.

Assumptions