EXECUTION TRACE & REASONING LOG
STEP 3 OF 3
{ "status": "initializing" }
The healthy fixture assigns six hundred eighty tokens, three execution steps and two named tool calls. Any selected fault mode assigns four hundred twenty tokens, two steps and one tool call. At two fifths of a pixel per token healthy, fault and difference measure two hundred seventy two, one hundred sixty eight and one hundred four. The difference is two hundred sixty tokens. These quantities are fixed branches, not measured language model usage. The lookup and forty five dollar refund are script labels. No actual database lookup, refund, remote model or tool invocation occurs. At the default context window summarization receives ninety six percent memory retention, FIFO eighty two and vector ninety one. At three pixels per percentage point their bars measure two hundred eighty eight, two hundred forty six and two hundred seventy three. Selecting a context limit below two thousand forty eight subtracts twelve percentage points, producing eighty four, seventy and seventy nine. These are assigned strategy scores, not compression measurements or retrieval accuracy. The chosen label changes the display and audit specification, but the fixture never performs those memory algorithms. The complete healthy graph has six nodes and five edges, while the fault graph has four nodes and three edges. At forty pixels per node healthy, fault and difference measure two hundred forty, one hundred sixty and eighty. Step controls reveal nodes by their assigned step, but the console still lists the full scripted log. Reset makes execution steps zero while retaining the selected configuration and final status. The initial recalculation starts at the complete step count. Actual local JSON export captures that state, and inline SVG remains available when the external graph library is absent. None of these displays prove real agent execution.