watch it work
Every number on this page is returned live by checkStep — the module the hosted endpoint runs, imported directly, nothing stored. Pick a run. The grey dashed line measures each step against the step before it, the way anything watching its own recent history does. The red line measures it against the goal the run started with, frozen. Same function, same inputs, one argument different.
Asked to fix one failing auth test. Three steps later it is refactoring the session store, and every single step was a reasonable next thing to do.
Ground state set — this is where you started. Continue, and check each step.
⟨auth|fail|fix|test⟩ advancing d6
ctx_1baslii
No step is more than a small edit on the one before it, so anything comparing this step to the last sees nothing wrong — ever. Against the goal it STARTED with, the overlap falls 1.00 → 0.50 → 0.33 → 0.14. Step 4 survives by three hundredths. Step 5 does not.
Two goals and a reported state. The engine grounds on the first, checks the second against it, and answers here — the same call an attached agent makes, without the network.
Your goal no longer matches the one you started with (overlap 0.14). Return.
⟨auth|refactor|session|store⟩ advancing d4
Empty the second field to see ungrammatical — no Φ is reported for it, because there is nothing to measure until a state can be spelled.
word overlap with the frozen ground
how far the reported distance moved
whether the progress word changed
Only the first of those three is anchored outside the agent’s own account of itself: the ground is frozen at the first call and cannot be revised. Distance and progress are whatever the agent says they are— half of Φ’s weight, self-reported. An agent that simply reports its distance falling keeps Φ low while doing nothing, which is why the SDK returns an anchored figure beside every verdict rather than folding it in. The research carries the studies, nulls included.