tegwick
a835c11ae7
Assistant: codex Assistant-Model: gpt-6-astra Assistant-Session: 01a101af-e240-7882-a1e3-e37720cbdbea
86 lines
4.5 KiB
Markdown
86 lines
4.5 KiB
Markdown
---
|
|
layer: Taxonomy
|
|
role: null
|
|
---
|
|
|
|
# Emergent Security — Intent
|
|
|
|
This repository declares Taxonomy: its products are a security model, reference
|
|
policy/conformance tooling, and source-backed agent knowledge. The reference
|
|
evaluator is not an operational PDP. NetKingdom integration follows
|
|
`docs/decisions/EMERGENT-DEC-2026-001-integration-profile.md`; cross-owner adoption
|
|
is tracked explicitly and is not implied by this declaration.
|
|
|
|
Coordination domain: `netkingdom`.
|
|
Statehub placement: domain `communication`, topic `netkingdom`.
|
|
|
|
Enable maximum local autonomy consistent with the guarantees actually owed.
|
|
Delegate understood decisions into explicit, testable risk envelopes; escalate
|
|
unresolved decisions without expanding privilege; keep platform-root under human
|
|
control. As evidence and decision maturity grow, routine work should need fewer
|
|
manual approvals without weakening security or tenant guarantees.
|
|
|
|
## Purpose and audience
|
|
|
|
Provide the security governance canon and executable policy reference for coding
|
|
agents, repository and domain maintainers, and platform and tenant governance
|
|
owners. Make it clear who may decide which effects, within what scope and risk,
|
|
with which evidence, and where an unresolved decision must go.
|
|
|
|
## Repository scope
|
|
|
|
- Maintain the versioned architectural canon in `specs/emergent-security-0.8.md`
|
|
and its self-contained agent extract in `emergent-security-agent.md`.
|
|
- Express the default delegation envelope, ordinal risk scales, and illustrative
|
|
platform and tenant governance envelopes as reviewable YAML policy data.
|
|
- Provide a small, pure Python reference decision function that evaluates scope,
|
|
prohibitions, root requirements, delegations, tenant bounds, negative memory,
|
|
evidence, risk, and reserved human decisions in a defined order.
|
|
- Preserve executable fixtures for autonomous decisions, obligations, denials,
|
|
escalations, root authorization requirements, and tenant governance validation.
|
|
|
|
The current implementation is a reference library and policy fixture suite. The
|
|
canon also describes future integration seams for intent storage, delegation
|
|
registries, credential brokering, enforcement, observation, and audit; this
|
|
repository does not currently implement those operational services.
|
|
|
|
## Constraints to preserve
|
|
|
|
- Identity, available credentials, readable context, and successful precedent do
|
|
not confer authority. Agents cannot authorize their own privilege expansion.
|
|
- Deterministic systems decide and enforce effects. Models may reason and
|
|
propose, but are not security boundaries. The reference decision function stays
|
|
deterministic, free of side effects, and independent of model calls.
|
|
- Applicable prohibitions prevail over grants. Platform-root produces
|
|
`ROOT_AUTHORIZATION_REQUIRED`; it is never autonomously granted or activated.
|
|
- The default autonomous envelope applies only to L0 agents on explicit human
|
|
tasks in a HOT build estate outside tenants. Other scopes require explicit
|
|
delegation; constitutional restrictions continue to apply everywhere.
|
|
- Tenant governance may only narrow the platform tenant envelope. Exceeding the
|
|
platform envelope routes to the platform; tenant-reserved human decisions
|
|
cannot be silently approved by automation or timeout.
|
|
- Missing scope, delegation, evidence, or acceptable risk produces an explained
|
|
escalation. Unclassified or unknown risk values fail safe. Applicable negative
|
|
memory blocks routine autonomy until resolved.
|
|
- Privilege is scoped and leased, decisions remain attributable, and evidence of
|
|
authority use must survive that authority. Learning proposes policy changes;
|
|
it does not grant permissions.
|
|
|
|
## Evolution and success
|
|
|
|
Keep policy, canon, agent guidance, and fixtures aligned. Start policy changes
|
|
with a failing decision fixture, change policy data where possible, and change
|
|
the evaluator when the intended semantics require it. Preserve deterministic
|
|
results, unmodified inputs, registry-order independence, and tenant narrowing
|
|
invariants.
|
|
|
|
Harden a zone when real consumers, data, contracts, or recovery costs require
|
|
guarantees. Revisit structure through explicit, bounded governance when it
|
|
creates needless friction; temperature changes never relax the constitution or
|
|
parent guarantees.
|
|
|
|
Success means more routine decisions resolved safely at the lowest sufficient
|
|
scope, fewer repeated escalations and root interventions, less standing
|
|
privilege, and clear evidence for every decision. This intent describes the
|
|
repository's purpose; it does not itself grant operational authority or replace
|
|
the canon and applicable delegation envelopes.
|