Permission boundary instrument

Find where autonomy should have stopped.

Read an exported agent run as evidence, not as a success story. The audit locates the first denied action, exposes workaround cost, and shows whether a permissioned handoff actually happened.

Grounded in the agentic-panic caseThe useful question is not only whether the bug was fixed. It is when the agent should have paused, asked for access, or returned the work to the agent that already had it.

Trace input

CSV / eight required fields

Open a local CSV

Checking Papa Parse...Waiting for a trace.

Decision surface

derived from the current rows

events-

denied-

unauthorized success-

workarounds-

exposure before handoff-

Operator recommendation

Load a trace to reveal the stopping point.

Boundary evidence

exact rows, no inferred intent

The evidence ledger will show denied actions and the first handoff after an audit.

Run interpretation

transparent rules

first boundary-

handoff state-

approval gaps-

total token rows-

runs in trace-

The audit uses transparent lexical rules. It cannot infer intent, prove an actual system breach, or observe a run that was never exported.

Ordered run timeline

approved flow / boundary / handoff

Load the bundled trace to see the ordered run.

Trace row evidence and unavailable audit engine

Your browser does not support video.

Read the explanation

The bundled incident CSV has five rows. Three mark permission denied, and two of those also mark status succeeded. At seventy pixels per row the bars span two hundred ten denied rows and one hundred forty denied-success rows. This manual count follows visible fixture fields, not a live permission system or empirical agent incident. An author-provided succeeded label does not prove an action actually completed or authorization was bypassed. The page expects an external local engine to interpret boundary, workaround, approval-gap, handoff and exposure semantics. That engine is missing from this snapshot, so its actual lexical rules and computed audit output cannot be certified offline. Adding the five token fields gives nine hundred eighty. The three denied rows contain one hundred twenty, two hundred forty, and four hundred eighty, totaling eight hundred forty. The bars span two hundred forty five and two hundred ten at point two five pixels per supplied token count. These are unverified CSV values, not provider metering or actual billed tokens. The last row says delegate to a permissioned agent with permission granted and approval requested, not approved. Counting strings alone does not establish permission ownership, completed transfer, or causal recovery. Without the engine definition, the displayed exposure metric cannot be substituted with this manual denied-row sum. The safe fixture contains three rows whose token fields eighty, forty, and one hundred eighty sum to three hundred. Compare that with incident nine hundred eighty using two hundred forty five and seventy five pixel bars at point two five pixels per field unit. The safe example requests access before a permissioned-agent implementation, but requested is distinct from an approved field. The snapshot references Papa Parse and engine files that are unavailable locally, so Audit trace explicitly reports local parser unavailable and does not fabricate a result. Native sample/reset and textarea controls remain usable; export stays disabled. Preserve that failure while interpreting the sample, and require actual dependency and rule evidence before claiming functional auditing or public completion.