What a draft has to show
A draft is proven when all of these hold:
A “clean” trace is a normal trace the draft is tested against. Hitting one of them is an “other hit”: a normal trace the draft would also have stopped.
When a draft is not proven
The Prevent section shows the real counts and what to do next:
The “at most allowed” number depends on how many clean traces the proof had. A draft may also stop only a small share of normal traces: fewer than 0.5% of the clean traces, and never more than 5. With 100 to 200 clean traces that allows none and with 201 it allows 1. Because a proof reads at most 1,000 traces, the most it can allow is 4. Drafts that must never touch normal traffic, such as a tool allow-list or confirm-first, allow 0.
Recheck runs the proof again on current traffic. See Policies → Policies Zespan drafted.
Sending past traffic
If you are short of clean traces, send past traffic toPOST /v1/ingest with its original, past timestamps. A trace counts when its spans fall within the test window: the last 14 days, or the last 30 days for a confirm-first draft (see Policies → Confirm before destructive calls). Older traces do not count. A trace that was still being written in the last 5 minutes is left out until it is complete, and each proof reads at most 1,000 traces.
Back-dated traces help a draft reach its proof, but detection itself still checks only recent traces. A problem that sits only in old, back-dated traffic is not found as a finding. Re-scanning past traffic for findings is not available yet.
Multi-agent pipelines
Many apps run several agents in one trace: a coordinator hands off to intake, policy and payout agents. Prevention handles this the way you would expect:- Clean traces are counted across the whole pipeline. A trace counts for a draft when the draft’s agent appears anywhere in it, not only on its last step. A pipeline that ends on a different agent than the one that made the offending call is no longer left out of the count.
- The draft is scoped to the agent that made the offending tool call. Not the agent that ended the trace, and not the agent that started it. For a loop finding (repeated calls, no progress, retry after an error) the agent is the one that made the repeated calls. If several agents made them, Zespan uses the one with most of them, and the latest one when they are level.
- Replay judges each call by its own agent, with the whole trace as history, the same way the live check sees it.
Conditions and once-per-key drafts
Two kinds of draft use more of your own data to stay narrow:- A confirm-first draft for a destructive tool applies only to the calls that meet the tool’s Only when condition, if you set one in Settings → Detectors. Proof replays only those calls.
- A once per key draft limits a money-moving, deleting or destructive tool to one call per key, such as one payment per claim id. The key is up to 3 argument fields that stayed the same across the repeated calls, for example
amount_inrandclaim_id, so a retry that changes one of them is a different call and is not blocked. A field that is different on every normal call (a claim number without “id” in its name) is kept in the key first, never dropped. Fields such as request ids, idempotency keys, timestamps and trace or span ids are never part of the key. Arguments longer than 2,000 characters are cut when recorded, so a call with such arguments falls back to the full-argument key. See Policy file reference → loop_detection.
Next steps
- Issues → Prevent — where a draft’s proof shows
- Policies — review and apply a drafted policy
- Detectors — confirm your tools and set an Only when condition
- Policy file reference — the rules a draft is built from

