zone-engine/docs/estate-partition-2026-08-19.md
tegwick de2d8e2dda Correct the rapp count: eight declarations, not nine
An inflated count in a document whose subject is accuracy.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-20 07:21:02 +02:00

12 KiB
Raw Blame History

Estate partition — ZONE-WP-0001-T02

Date: 2026-08-19 · Corpus: ops-warden registry/routing/catalog.yaml (27 entries), inventory.yaml (4 actors), registry/policy/security-posture.yaml (3 env postures, 4 maturity levels).

The test set by T02: the model must partition what exists without a residue of unexplained exceptions, and the exceptions found are the most informative output. This is the raw result, before any model is fitted to it.


1. What the data actually says

Cross-tabulating the three declared fields that could plausibly drive membership — delegation.mode, risk, exec_capable — fills 9 of 27 possible cells. The estate clusters hard.

delegation risk exec_capable n character
permanent ungraded 1 The one lane ops-warden decides (ssh-cert-host-access)
native ungraded 7 Pure pointers: route to an owner, nothing executes
native ungraded false/true 2 Owner-fronted exec (whynot-design-npm-publish, audit-core-senders)
native high false 1 database-dynamic-credentials — owner-fronted, dangerous
interim ungraded true 4 ops-warden proxying, ungraded
interim standard false 2 Interim, graded, not proxyable
interim high false 2 Interim, dangerous, routed only
interim high true 8 ops-warden proxying a high-risk credential for an absent owner

The three fields are not independent, and that is the finding. Danger concentrates in interim: 10 of the 11 high-risk lanes are interim, and 8 of those are exec_capable. permanent and native are almost entirely benign.

Delegation mode is doing most of the discriminating work. It was introduced by WP-0030 to answer a governance question — should ops-warden own this? — and it turns out to predict operational danger better than the field named risk.

2. The bands that exist

Four, derived rather than invented. Every one of the 27 lanes lands in exactly one.

Band n Defining property What it means for a control
Decided 1 warden_executes: true — ops-warden is the decision point The only place a pre-sign gate can bite at all
Fronted 10 delegation.mode is native — the owner has a front door Stance is the owner's; ops-warden holds nothing
Covered 12 interim and exec_capable: true — ops-warden proxies a value it does not own The band that actually needs rigidity, and the band a refactor most needs relaxed
Signposted 4 interim, not exec_capable — routed, nothing flows Advisory at most; there is no value to leak

The interesting boundary is Covered vs Fronted, and it is not a risk judgement — it is a statement about whether the owning component exists yet. That is why it moves over time and why bands must be re-derived, not assigned once.

3. The existing controls already imply bands — and they disagree

Before any zone model, ops-warden's three controls each cut the estate differently:

Control Lanes touched Selector
flex-auth pre-sign gate 1 warden_executes
fetch policy gate 13 exec_capable
agent read-boundary (ADR-0004) 11 risk == "high"
warden plan escalation 27 all

Only 8 lanes are in both the fetch gate and the read-boundary. The controls were built independently and their scopes were never reconciled. A zone model that assigns one zone per lane and expects all controls to agree with it will misfit immediately — which answers a question T03 was going to have to ask: stance is per control per zone, never per zone alone.

4. The residue — and it is the most valuable output

14 of 27 lanes carry no risk value at all.

This is not a modelling inconvenience. is_high_risk is risk == "high" (routing/models.py:123), so an ungraded lane never triggers the agent read-boundary. Five of the fourteen are exec_capable: true, meaning warden access --fetch can stream a real value for them to an agent session: openbao-api-key, whynot-design-npm-publish, key-cape-oidc-login, issue-core-ingestion-api-key, reuse-surface-hub-write-token.

ADR-0004 reads as a categorical rule. The implementation is an opt-in list, and absence fails open.

Routed to risk-nexus as RISK-F-0003. It is a defect in a shipped control, not a known gap being worked — the distinction ops-warden's own finding-routing rule draws. Severity is risk-nexus's to set.

For this task it also means: a zone model cannot be derived from risk as it stands. Half the corpus is unlabelled on that field, and grading it is real judgement, not backfill.

Two further gaps, less severe but structural:

  • warden_role and has_rotation are derived at load, not declared — they are absent from every raw entry despite appearing in warden route --json. A membership compiler must read the model, not the file.
  • The posture registry (env_postures, M0M3) and the catalog share no join key. No lane references a workload or an environment. They are currently disjoint vocabularies.

5. Answers to T02's questions

Is a zone a property of the actor, the lane, the workload, or the moment?

Corrected 2026-08-19 by operator direction, after this analysis was written. The answer below ("the lane") mistook the corpus for the subject. Security policy is always about the actual workload. ops-warden's catalog is a credential surface, not a workload inventory, so lane properties were the only thing available to partition — and partitioning what is available is not the same as finding the policy subject. The bands in §2 stand as an accurate description of ops-warden's credential surface and remain useful membership inputs; they are not zones. See §7.

The lane, on the evidence available in this corpus. Every field that discriminates — delegation.mode, exec_capable, risk — is a lane property. The actor inventory holds 4 actors mapping to a single lane and cannot carry a partition. No lane references a workload.

With two qualifications the data forces:

  • The caller is a modifier, not the zone. The read-boundary keys on WARDEN_AGENT_ID — a property of who is asking, orthogonal to the lane. A zone says what the lane requires; the caller decides which requirement applies. Collapsing them would make one lane need two zones.
  • The moment is a different object, confirmed. Nothing in the corpus represents it. delegation.blocked_on carries a free-text claim with a reviewed date and no expiry — precisely the "expiry by intention" shape ZONE-WP-0001-T04 must not reproduce. A time-boxed relaxation is an exception against a zone, never a zone.

Do zones compose with, or fold in, environment posture and M0M3?

Compose — and the composition cannot be built yet. They are per-workload; zones as derived here are per-lane; and there is no join key between them in any registry today. Folding them in would require inventing a lane→workload mapping that does not exist, which is fabrication under §6.

The honest sequence: zones stand alone on lane properties now, and a join is proposed only when something actually declares it. (organization_posture was already ruled out of the declaration entirely by net-kingdom — it is an input.)

Reefs

Struck by net-kingdom. Per tenancy-posture Decision 8.4.1, substrate location is never by itself evidence for a level. The unreconciled pair is reef ↔ P/V and it is canon's, tracked as NK-WP-0027.

6. What this hands to T03 and T04

  • Four bands, each lane in exactly one, derived from declared fields.
  • Stance must be per control per zone — the existing controls demonstrably disagree about scope.
  • Covered is the band the model exists for: 12 lanes where ops-warden proxies what it does not own, holding the highest concentration of high-risk exec_capable lanes, and the band a deep refactor most needs relaxed.
  • risk is not usable as a membership input until the 14 ungraded lanes are graded (RISK-F-0003). delegation.mode is usable today and is a better predictor.
  • Membership must be computed from the loaded model, not read from raw YAML.

7. Correction — the subject is the workload, and a zone is an admission standard

Operator direction, 2026-08-19, after §1§6 were written:

Security policy is always about the actual workload. The repository providing the software may suggest security policy for running workloads in different situations, but the workload — and whoever is responsible for it — defines the level of scrutiny. Some zones may well define standards that have to be met in order to run as a workload in that zone.

This reverses the direction the analysis above assumed. It had zones as a classification derived from properties. They are bands with admission standards: a workload does not get labelled with a zone, it qualifies to run in one.

Three roles, not one

Role Who Authority
Suggests The repository providing the software Recommends a posture for running its software in given situations. A recommendation, never a determination — it does not know where its software will run
Declares The workload and its responsible party Sets the level of scrutiny actually applied. This is the policy subject.
Requires The zone Defines the standard a workload must meet to be admitted. A floor, not a label

This is canon's existing mechanism, not a new one

tenancy-posture_v0.1 Decision 8.2 — split authority, machine-reconciled already partitions authority this way: the package repo owns substrate numbers and enforcement, "the consuming repo owns its workload requirements", and a tier minimum is owned separately and joined by machine at definition and whenever either side changes — because "a machine-checkable constraint must not depend on somebody remembering to collect a signature."

And Decision 5.6, ruling on this very standard, already said stance "behaves like a tier minimum under Decision 8.2 — asserted elsewhere, joined by machine."

So the zone model does not need a new prescription mechanism. It needs to be a tier minimum: the zone asserts a floor, the workload asserts its posture, and the join is mechanical. That also resolves the grade-versus-acceptance split raised for T03 — they are simply the two sides of the join.

What this does to the bands in §2

They survive as inputs, demoted from conclusions. Decided / Fronted / Covered / Signposted accurately describe ops-warden's credential surface and still predict where danger concentrates. What they cannot do is name the entity policy applies to, because ops-warden's catalog contains no workloads.

The missing join found in §4 is therefore not a tidy-up. It is the model.

Corrected 2026-08-20. "No registry carries a join key" was too strong — it was concluded from ops-warden's catalog alone, the one place a workload declaration would not live. rapp-*/declarations/rapp.yaml declares workload_identity with data_classification and criticality for eight workloads, and ops-warden's dataclass_floor already maps classification to M0M3. The chain exists and spans two repos. What is genuinely missing is an explicit workload: field on catalog entries — the key is currently only derivable by parsing path_template, whose convention is inconsistent. See ZONE-WP-0001-T03.

Build stage

The organization is in build. The default stance is therefore permissive, and that is a positional fact rather than a concession: ADR-0006 deferred a global gate for exactly this reason. Zones tighten by admission to a zone, not by raising a global floor — which is what makes the model compatible with deep refactoring instead of hostile to it.