Agent Gate Core is the public evidence layer for AI-agent incident replay runtimes such as Agent Gate Incident Replay.
It consumes existing runtime and audit evidence, normalizes it into a stable
agent.gate.result.v0 shape, and emits deterministic PASS/FAIL verdicts. It is
not a policy engine, agent runtime, credential broker, or product lifecycle
manager.
Core owns the evidence schema and verdict reducer. Replay runtimes own the browser, VM, incident module, savestate, and transcript surfaces that produce the evidence consumed by Core.
agent-gate-policy-result policy-events.jsonl
agent-gate-run ...
agent-gate-decide *.json
agent-gate-policy-result normalizes existing runtime, policy, or ledger event
JSONL into policy-result.json.
agent-gate-run runs one agent or tool command through the existing audit path
and emits one normalized agent.gate.result.v0 gate result.
agent-gate-decide reduces one or more normalized result files into one
deterministic verdict.
runtime / policy / ledger events
-> agent-gate-policy-result
-> policy-result.json
-> agent-gate-run
-> gate-result.json
-> agent-gate-decide
-> VERDICT: PASS / VERDICT: FAIL
Gate v0 fails closed when evidence is missing, malformed, has an invalid schema, has invalid audit evidence, or contains a CRITICAL policy violation. Otherwise it emits PASS.
From this repository root:
python3 -m unittest discover -s tests -p 'agent_gate_v0_test.py'Run the smallest replay fixture:
scripts/agent-gate-policy-result examples/agent-gate-replay/events-fail.jsonlRun the smallest end-to-end PASS/FAIL replay:
cat examples/agent-gate-replay/README.mdThat fixture shows how policy events become policy-result.json, how
agent-gate-run emits a normalized gate result, and how agent-gate-decide
returns the final verdict.
Downstream replay runtimes may call these CLI scripts directly or package them into a VM image. They must treat this repository as the source of truth for:
agent.gate.result.v0- policy evidence normalization
- deterministic PASS/FAIL reducer behavior
If a replay runtime or incident module needs new evidence fields or verdict semantics, update Agent Gate Core first instead of redefining reducer behavior inside the replay runtime.
The current signature field and verdict output line are reserved placeholders
for a later provenance layer. They must not be represented as production
cryptographic identity proof in v0.
Agent Gate Core owns evidence schemas, normalization, deterministic verdict rules, CLI fixture tests, and examples.
Agent Gate Core does not own browser replay, v86, VM savestates, incident catalog publication, AI Workforce OS internals, credential broker internals, or third-party certification claims.