Trust boundaries and policies

A Boundary names the information-flow decision made during packing. It does not mutate item labels, and it does not send data anywhere.

internal = cx.Boundary.internal("application")
external = cx.Boundary.external("openai")

Core policy matrix

For optional and preferred items, a violation creates a manifest exclusion. For required items, the same violation raises PolicyViolation because silently dropping the item would break its retention contract.

Item condition	Internal boundary	External boundary
Expired	Excluded / required raises	Excluded / required raises
quarantined	Excluded / required raises	Excluded / required raises
No sources	Allowed	Excluded / required raises
restricted	Allowed	Excluded / required raises
Matches a built-in secret pattern	Allowed	Excluded / required raises
untrusted with provenance	Allowed	Allowed
internal or confidential sensitivity	Allowed	Allowed by the core policy
Binary	Not packable in the 0.1 series	Not packable in the 0.1 series

The table describes the built-in core policy only. Applications often need stricter custom policies for confidential data, tenant separation, source allowlists, or untrusted content.

Untrusted is a label, not a block

External records default to untrusted. Provenance-bearing untrusted content remains eligible until an application policy rejects or quarantines it.

import cognoxium as cx

frame = cx.CognitionFrame.from_records([
    {"id": "web", "payload": "Public web result", "sources": ["https://example.com"],
     "created_at": "2026-01-01T00:00:00Z"}
])

allowed = frame.pack(budget=100, boundary=cx.Boundary.external("provider"))
blocked = frame.quarantine(lambda row: row["trust"] == cx.Trust.UNTRUSTED).pack(
    budget=100,
    boundary=cx.Boundary.external("provider"),
)

print([item.id for item in allowed.items])
print([(item["id"], item["reason"]) for item in blocked.manifest.excluded])

['web']
[('web', 'quarantined')]

Required violations

try:
    frame.pack(
        budget=100,
        boundary=cx.Boundary.external("provider"),
    )
except cx.PolicyViolation as error:
    print(error.code)         # CX_POLICY_001
    print(error.item_id)
    print(error.rule_id)
    print(error.boundary)
    print(error.reason_code)

Do not retry the same item through a weaker boundary without a separate application authorization decision.

Custom policies

Pass one or more Policy objects to pack(policies=[...]). A policy returns None to allow an item or a stable reason code to reject it. See Writing plugins for a complete implementation.