Trace-to-Gate (replay recorded traces)

Point the gate at the tool-call log an agent already produced and see what it would have decided — the onboarding hook, before wiring anything live.

from evidence_gate import TraceMapping, normalize, simulate

calls = normalize(trace_records, TraceMapping(tool="name", call_id="id",
                                              observed_at="ts", result="data.output")).calls
reports = simulate(calls, gate=gate, builder=builder,
                   action_mapping={"send_*": "marketing.send_sequence"}, now=now)
# -> [SimReport(request_id=..., effect=ALLOW/REVIEW/BLOCK, executed=..., reasons=[...])]

normalize maps arbitrary vendor exports (dotted field paths) into the ToolCall shape; a record missing a required field is skipped and surfaced, never guessed. For the common vendors this is a one-liner — ship-ready TraceMapping presets cover their per-observation shapes:

from evidence_gate import LANGSMITH, LANGFUSE, OPENAI, normalize

calls = normalize(runs, LANGSMITH).calls        # or LANGFUSE, or OPENAI

simulate scopes evidence per turn and runs every sensitive call through the untouched gate.check(). See examples/trace_replay.py.