zone-engine/INTENT.md

# INTENT

> Why `zone-engine` exists and where it is going.
> What is true today is `SCOPE.md`. Checkable gates and the retirement
> condition are `GOAL.md`. This file is the argument, not the checklist.

---

## 1. The sentence this repo exists to make sayable

> *"This control is enforced **there**, advisory **here**, and relaxed **in this
> band until Friday** — and all three of those are written down, reviewable, and
> expire on their own."*

NetKingdom cannot say that today. It can say how exposed a workload is
(environment posture), how ready it is (`M0`–`M3`), and what state the
organization is in (`organization_posture`). Every one of those **describes**.
None of them **decides**. So every enforcement control in the estate has been a
boolean over a whole repo.

## 2. Where that bit

ops-warden built a pre-sign authorization gate, flex-auth enforced its side, and
the gate verified clean. Then it was deliberately not switched on.

`policy.enabled` is one boolean over the entire repo. With `fail_closed: true`
it makes flex-auth a hard dependency of *every* `warden sign` — including the
SSH certificates the ops-bridge tunnels depend on, one of which carries the
policy call itself. Turning it on would have been correct for a settled
production lane and wrong for an estate in the middle of decommissioning one
host and moving a service off it.

That is not a story about a flag. **Security that stops the work stops being
security and becomes an outage with good intentions** — and the reason it would
have was structural: there was nowhere to say *where* it applies. The decision
is ops-warden's `ADR-0006`; this repo is what it defers to.

## 3. What a zone has to be, to be worth having

A classification that changes nothing is a label. A zone earns its keep only if
it decides something, so the model must carry:

- **Membership** — what puts a lane, actor or workload in a zone, derived from
  posture already declared rather than a fourth hand-maintained list.
- **Stance per control** — enforced, advisory, or exempt.
- **Failure mode** — what happens when the control cannot run. This is where
  the bite is; `fail_closed` is what turns a dead tunnel into an outage.
- **Exceptions that expire without anyone remembering them.**

The last is the reason this is a repo rather than a document. A standing
classification needs no engine — canon plus a declaration file covers it, which
is exactly how tenancy posture works. **A relaxation with an expiry is state**,
and state wants an owner, an audit trail, and something other than good
intentions enforcing the deadline.

## 4. The line this repo must not cross

`flex-auth` is the policy decision point. It stays the only one.

zone-engine is authority over zone **identity and membership**. The **effect**
of a zone on a decision flex-auth renders belongs in a flex-auth policy package,
because flex-auth's decision envelope stamps `matched_policy_version` and
`policy_package` — and registry content appears nowhere in that provenance. Put
stance in the registry and two decisions with the same policy version and the
same request can differ, with nothing in the audit trail explaining why.

That boundary was drawn by flex-auth on review, against a first draft of ours
that got it wrong in an instructive way: the original invariant forbade sitting
*synchronously* in a decision path, which is a latency guarantee wearing an
authority guarantee's clothes. Compiled data that determines an outcome is still
deciding; it just decided earlier.

**Held to, in one line: membership is ours, stance is theirs, and the decision
point is neither.**

## 5. What success looks like from outside

Someone who has never read this repo should be able to:

- open a service's `tenancy.yaml`, read its `zones:` key, and know which
  controls bite it;
- ask "why was this relaxed?" and get a justification, an author and an expiry,
  not a shrug;
- watch a relaxation lapse **without a human remembering it**;
- and find the rule published in NetKingdom canon rather than in this repo's
  prose.

If a reader has to come here to understand their own posture, the model has
failed at being a standard and has succeeded only at being a service.

## 6. What would make this repo wrong

Stated up front, because a repo created on a recommendation should say what
would falsify it:

- **The exception lifecycle turns out not to need a runtime.** If reviewed
  declarations in git, with expiry evaluated at decision time, are sufficient,
  then the honest outcome is a canon standard and no engine — and this repo
  should be archived rather than kept for the sake of existing. `GOAL.md` gate 5
  and `ZONE-WP-0001-T04` both hold that door open deliberately.
- **The zones do not partition the real estate.** If the model needs a residue
  of special cases to cover today's 27 routing lanes and actor inventory, it is
  describing an aspiration, not a structure.
- **Nobody but the author declares one.** Two repos declaring and a third
  reading is the minimum evidence of a model rather than a preference.
- **It becomes a second decision point.** The failure would not announce itself;
  it would arrive as a small convenience — a stance compiled into membership
  "just for now".

## 7. What this repo deliberately will not become

Not a PDP. Not network segmentation — "zone" is overloaded and the routing sense
is someone else's word. Not a placement model: reefs answer *where a thing runs*
and are `repo-manager`'s, and the fact that `reef-railiance` caps availability
for everything bound to it is a canon composition problem (`NK-WP-0027`), not
ours. Not an owner of anyone's controls — the controls stay with the repos that
enforce them; this model only says where they bite.

## 8. Direction

Model first, from the real estate. Canon standard drafted here and published by
`net-kingdom`, in the family of `tenancy-posture_v0.1`. Membership declared in
the `zones:` key that canon has already reserved. Two consumers before it is
called adopted, the first being ops-warden retiring `policy.enabled`.

A runtime only if `ZONE-WP-0001-T04` proves the exception lifecycle needs one —
and the cheaper answer is a legitimate result, not a disappointment.