Skip to main content
Policies separate detection configuration from detection logic. Instead of changing code to raise a block threshold or demote a noisy signature to monitor-only, you edit a YAML file. InferenceWall auto-discovers policy files at startup, so changes take effect on the next restart without any package modification.
1

Copy the default policy

mkdir -p ~/.inferwall/policies
cp $(python -c "import inferwall; print(inferwall.__path__[0])")/policies/default.yaml \
   ~/.inferwall/policies/my-policy.yaml
This gives you the exact policy that ships with InferenceWall as a starting point, with all five thresholds and an empty signatures map.
2

Edit thresholds, mode, and per-signature overrides

Open ~/.inferwall/policies/my-policy.yaml and modify the values you want to change:
name: my-policy
version: "1.0.0"
mode: monitor                   # Start in monitor mode

thresholds:
  inbound_flag: 6.0             # Raise flag threshold to reduce noise
  inbound_block: 12.0           # Raise block threshold
  outbound_flag: 4.0
  outbound_block: 8.0
  early_exit: 15.0

signatures:
  INJ-D-001:
    action: enforce             # Force-enforce this signature even in monitor mode
    anomaly_points: 12          # Increase weight
  CS-T-003:
    action: monitor             # Demote to monitor-only
3

Point InferenceWall at your policy

Drop the file into ~/.inferwall/policies/ and InferenceWall will auto-discover it on the next startup. To explicitly select a specific file, set IW_POLICY_PATH:
export IW_POLICY_PATH=~/.inferwall/policies/my-policy.yaml
IW_POLICY_PATH is useful in CI/CD pipelines or container deployments where you want deterministic policy selection rather than auto-discovery.

​
Policy YAML fields

​
Top-level fields

FieldTypeDescription
namestringHuman-readable policy name for logging and audit events
versionstringSemVer string; increment when you make changes
modemonitor | enforceGlobal enforcement mode (see below)
thresholdsobjectFive scoring thresholds (see below)
signaturesmapPer-signature overrides keyed by signature ID

​
Thresholds

ThresholdDescriptionDefaultStrict
inbound_flagScore to flag an incoming request4.02.5
inbound_blockScore to block an incoming request10.07.0
outbound_flagScore to flag an outgoing response3.02.0
outbound_blockScore to block an outgoing response7.05.0
early_exitScore at which downstream engines are skipped13.010.0

​
Enforcement modes

ModeBehavior
monitorAll signatures run and log matches, but no blocking or flagging actions are taken
enforceSignatures contribute scores to threshold comparisons and trigger flag or block decisions

​
Per-signature overrides

Override individual signatures within the signatures map: 
signatures:
  INJ-D-001:
    action: monitor        # Override to monitor even in global enforce mode
    anomaly_points: 3      # Lower scoring weight for this signature
  INJ-D-008:
    action: enforce        # Force enforce even in global monitor mode
FieldValuesDescription
actionenforce, monitorOverrides the global mode for this specific signature
anomaly_points1–15Overrides the signature’s default scoring weight

​
Override precedence

When multiple sources could set a signature’s action, InferenceWall resolves them in this order (highest wins):
  1. Per-signature override — signatures.<ID>.action in the policy file
  2. Global policy mode — the top-level mode field
  3. Signature default action — tuning.default_action in the signature YAML

​
Environment variable

export IW_POLICY_PATH=~/.inferwall/policies/my-policy.yaml
When set, InferenceWall loads exactly this file and skips auto-discovery. When unset, it discovers all .yaml files in ~/.inferwall/policies/.

​
Auto-discovery

Drop any .yaml file into ~/.inferwall/policies/ and InferenceWall will pick it up on the next startup. Use this when managing multiple policy profiles (e.g., strict.yaml, permissive.yaml) and switching between them by setting IW_POLICY_PATH.
Recommended rollout workflowRoll out new policies gradually to avoid unexpected blocking in production:
1

Deploy in monitor mode

Set mode: monitor. All signatures run and log matches, but no requests are blocked.
2

Observe logged matches

Watch the scan logs for 1–2 weeks. Identify which signatures fire most often and which generate false positives.
3

Allowlist false positives

For signatures that fire on known-benign traffic, set action: monitor in the per-signature overrides.
4

Enforce high-confidence signatures individually

Flip signatures you trust to action: enforce one at a time, monitoring the effect of each change.
5

Switch global mode to enforce

Once you’re confident in the policy, set mode: enforce to activate blocking globally.

​
Further reading

Policy concepts

How InferenceWall resolves policies, applies overrides, and manages the policy lifecycle.

Scoring concepts

The anomaly scoring model: confidence weighting, diminishing corroboration, and threshold math.