emergent-security

Security guidance, delegation envelopes, and source-backed situational knowledge for coding agents in the NetKingdom ecosystem. Root stays human-controlled.

Agent knowledge

Start with the orientation card and the NetKingdom integration map. Install and use:

pip install -e '.[test]'
emergent-knowledge show start --audience codex
emergent-knowledge show access --audience claude --json
emergent-knowledge --repositories /path/to/checkouts check

The package serves reviewed estate-general summaries with source versions and freshness receipts. It supports Codex, Claude, Grok, and generic adapters, offline bundles, and changed/withdrawn-source withholding. It never retrieves credentials, private tenant content, or operational grants. Source checks verify documented bytes, not live health. See loading and maintenance and operational seams.

Release status is pilot. Provider proof and owner adoption are tracked in EMERGENT-WP-0002. Knowledge loading is separate from enforcement, identity isolation, and fleet rollout. The integration decision keeps the reference model distinct from the estate's sole PDP, access-engine.

Default envelope reference

The default delegation envelope of the Emergent Security canon v0.8 (A6) as policy data, the tenant envelopes (B10, B11) as examples, and a small reference decision function (A3) that the fixtures run against.

envelopes/
  scales.yaml                     ordinal risk scales shared by all envelopes
  default-envelope.yaml           A6: what an L0 agent decides alone in a HOT build estate
  examples/
    platform-tenant-acme.yaml     B10: platform-owned outer bound for one tenant
    tenant-governance-acme.yaml   B11: tenant-owned narrowing, self_governed mode
src/emergent_security/            one module per concern (see package docstring)
tests/fixtures/
  decisions/*.yaml                request → expected decision, data only
  tenants/governance-validation.yaml

Run

pip install -e '.[test]'
pytest

Semantics in one screen

• Modes: autonomous, autonomous_with_obligations, escalate, deny, root_authorization_required.
• Variants refine a row by request attributes; first match wins and may replace mode, rules, obligations and evidence.
• deny and root_authorization_required are constitutional: they apply in every scope, including tenants and COLD zones, and no delegation overrides them.
• Every other row applies only where applies_when matches (L0, HOT, build, no tenant).
• escalate means "not delegated here". A registry delegation can grant it (most specific scope wins, ties by id).
• Inside a tenant the default envelope never grants anything. Grants come from the platform tenant envelope or the registry, always bounded by the effective envelope: platform ∩ tenant governance.
• Risk fails safe. A dimension the envelope declares but the request leaves unclassified, or classifies with an unknown value, counts as exceeding.

Changing the policy, test first

1. Add or change a case in tests/fixtures/decisions/*.yaml describing the request and the expected decision. Run pytest; it must fail.
2. Change envelopes/default-envelope.yaml (or an envelope example) until it passes. Touch code only if the canon's semantics changed.
3. Adding an A6 row also means updating A6_CLASSES in tests/test_default_envelope.py and the canon table.

The suite also enforces determinism, idempotency (inputs never mutated, registry order irrelevant), that auto mode equals no tenant configuration, and the G7 tenant invariants.

Human decisions and guided procedures

Decision/runbook integration links informed-decision presentation and human binding to owner-held procedures using runbook-tutorials. Retrieve the decisions and runbooks cards. The native runbooks/security-orientation/pack.yaml prepares an evidence handoff; the unsigned adoption memo in docs/decisions/ is ready for review. Neither submits a human answer or executes an infrastructure change.