Assent to AUDIT-IN-0001 — approval evidence half, with conditions

gate-house proposed (security layer model v0.3 §9.4) that audit-core own
the tamper-evident record of approvals while approval-engine owns the
operative state. Assented, with two corrections and one condition.

Correction 1: INTENT principle 6 overstated the delivered guarantee.
docs/integrity.md is authoritative — an in-database chain does not
withstand a database owner without the external chain-head attestation,
and even then it is not WORM. Principle 6 now points at that bound.

Correction 2: the chain proves alteration and truncation, not omission
at source. A suppressed revocation leaves the chain intact.

Condition: approval-engine must guarantee emission atomicity, so an
approval cannot change state without the event being durably queued in
the same transaction. Not a task audit-core can discharge for it.

Boundary: no approval-validity query will be exposed — that would be
deciding early under §6.1.

Also declares audit-core's layer (Engine, not a decision point), which
§11 requires and INTENT.md was missing.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WpeL68AWHqtqPQZEXY5kFe

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 4040362@bnt-lap001
Assistant-Session: 4fd0fd24-2ee8-4413-bd67-43bd79ca73f1

INTENT.md

@@ -17,6 +17,18 @@ Audit Core must integrate cleanly with NetKingdom, Railiance, OpenBao,
 Kubernetes, identity providers, and application runtimes, but it must not depend
 on NetKingdom to exist or operate.
 
+## Layer
+
+Audit Core is **Engine layer** under the NetKingdom security layer model
+(`net-kingdom/canon/standards/security-layer-model_v0.3.md`): it exposes a
+deterministic API contract and holds operational custody of the records it
+accepts.
+
+It is explicitly **not a decision point**. Audit Core answers what was recorded;
+it never answers whether an action is permitted, and it exposes no surface that
+returns such a verdict. See `## Out Of Scope Initially` and
+`history/2026-08-28-approval-evidence-assent.md`.
+
 ## Problem
 
 Modern platforms generate audit-relevant events in many places:
@@ -111,9 +123,14 @@ The data plane owns:
 5. **Immutable archive and hot search are separate concerns.** Hot search can
    be short-lived and cost-tuned. Immutable archive is the evidence record.
 
-6. **Tamper evidence is explicit.** Audit Core should store signed or
-   hash-chained manifests for retained batches so operators can later prove
-   whether a record set was changed, omitted, or truncated.
+6. **Tamper evidence is explicit, and bounded.** Audit Core should store
+   signed or hash-chained manifests for retained batches so operators can later
+   prove whether a record set was changed or truncated. The delivered bound is
+   narrower than that aspiration and `docs/integrity.md` is authoritative on it:
+   an in-database chain does not withstand a database owner without the external
+   chain-head attestation, and the chain proves nothing about an event that was
+   never emitted. Completeness at the boundary is the emitting source's
+   property, not the archive's.
 
 7. **Least-privilege access.** Tenants may access their own audit records
    according to policy. Platform operators may access cross-tenant records only
@@ -225,6 +242,33 @@ NetKingdom must not be required for Audit Core's internal authorization model.
 Audit Core may accept NetKingdom OIDC claims when configured, but should also
 support a generic OIDC provider and local development auth mode.
 
+## Approval Evidence
+
+Audit Core carries the **evidence half** of approvals as a distinct source, per
+security layer model §9.4 and `AUDIT-IN-0001` (assented 2026-08-28).
+
+`approval-engine` owns the operative approval state — the durable object, atomic
+supersession, single consumption, and revocation. Audit Core owns the record of
+what happened to it: the four event classes **issuance, use, supersession, and
+revocation**.
+
+Terms Audit Core accepts this under:
+
+- `approval-engine` is registered as a distinct source, with its own sender
+  registration, tenancy mapping, retention profile, and `secret_policy`,
+  onboarded under principle 4 like any other source;
+- `approval-engine` guarantees **emission atomicity** — an approval cannot be
+  issued, consumed, superseded, or revoked without the corresponding event being
+  durably queued in the same transaction. Audit Core cannot detect a suppressed
+  emission and will not imply that what it received is everything that happened;
+- Audit Core exposes **no approval-validity query**. Records, yes; a verdict
+  answering whether an approval is still valid, never. That would be deciding
+  early under §6.1 and would couple authorization to the audit fabric, which is
+  what §9.4 exists to prevent. Callers needing current state ask
+  `approval-engine`.
+
+Full reasoning: `history/2026-08-28-approval-evidence-assent.md`.
+
 ## Tenant And Scope Model
 
 Audit Core should distinguish:

history/2026-08-28-approval-evidence-assent.md

@@ -0,0 +1,135 @@
+# Approval evidence — audit-core's assent (AUDIT-IN-0001)
+
+**Date:** 2026-08-28
+**Intake:** `AUDIT-IN-0001`
+**Requested by:** gate-house
+**Standard:** `net-kingdom/canon/standards/security-layer-model_v0.3.md` §9.4 (proposed)
+**Outcome:** assent, with two corrections to the rationale and one condition on
+the emitting engine.
+
+---
+
+## What was asked
+
+gate-house proposed that audit-core own the **evidence half** of approvals —
+issuance, use, supersession, and revocation, emitted as audit events — while a
+new `approval-engine` owns the operative state: the durable object, atomic
+supersession, single consumption, and revocation without holder cooperation.
+
+The two reasons given for keeping the operative state out of audit-core are
+correct and audit-core does not contest them:
+
+- operational custody is append-only Postgres by design
+  (`docs/audit-backend-contract.md`), and an approval needs mutable
+  current-state semantics with compare-and-swap supersession; and
+- `access-engine` must read approval state in-path, so putting it behind the
+  audit fabric would make an audit outage an authorization outage.
+
+## The answer
+
+**Assent.** The split is the right one. Approval evidence is a record of what
+happened, not the authoritative answer to whether an approval is still valid,
+and that distinction is what keeps this inside `INTENT.md`'s "policy decision
+making" exclusion rather than a violation of it.
+
+Two corrections and one condition follow. None of them reject the split; they
+state what audit-core actually delivers so the standard does not rest on a
+stronger claim than the implementation supports.
+
+---
+
+## Correction 1 — principle 6 is an aspiration, and the delivered bound is weaker
+
+gate-house's rationale cites `INTENT.md` principle 6 — *signed or hash-chained
+manifests so operators can later prove whether a record set was changed,
+omitted, or truncated* — as already being what authenticated approval entries
+need forensically.
+
+`docs/integrity.md` is more careful than principle 6, and it is the contract
+that ships. In its own words: a chain *inside* the same database detects a
+rewritten `payload_hash` **only if the attacker does not also recompute the
+suffix** — and a database owner can. Tamper evidence against that class of
+attacker requires the **external chain-head attestation**, and even with it:
+
+> It still does not mean WORM, object lock, or ITC-CAP `data.archive`.
+
+`tamper_evidence=True` is conditional on two preconditions being live
+(`verify` failing on a rewritten row, and verify-against-attestation reporting
+a missing cited head as a break). It is not a property of the store at rest.
+
+**So: approval events get exactly the guarantee every other audit source gets —
+no more.** That guarantee is real and it is useful, but it is bounded by the
+attestation being taken and stored outside `platform-pg`. If §9.4 needs a
+stronger property for approvals specifically — WORM, object lock, or a
+transparency log — that is an engine gap to declare under §5.3, not a setting
+audit-core can turn on.
+
+## Correction 2 — the chain proves alteration, not omission at source
+
+This is the more consequential one, and it is the reason for the condition
+below.
+
+The hash chain proves that accepted records were not **altered or truncated**.
+It cannot prove anything about an approval event that was **never emitted**.
+For an approval object the omission case is not a corner: a suppressed
+*revocation* is precisely the event an attacker most wants missing, and it
+leaves the chain perfectly intact. Verify would report `intact`.
+
+Principle 6 claims audit-core can prove a record set was "omitted". Against a
+compromised or buggy *sender*, it cannot, and neither can any append-only
+archive. Completeness at the boundary is the emitter's property, not the
+archive's.
+
+## The condition — approval-engine owns emission atomicity
+
+audit-core accepts the evidence half **on the condition that `approval-engine`
+guarantees emission is atomic with the state change** — a transactional outbox
+or equivalent, such that an approval cannot be issued, consumed, superseded, or
+revoked without the corresponding event being durably queued in the same
+transaction.
+
+Without that, the evidence half is silently incomplete and audit-core has no
+mechanism to detect the gap. audit-core will report what it received; it will
+not imply that what it received is everything that happened.
+
+This is a requirement on `approval-engine`, not a task audit-core can discharge
+on its behalf. It should appear in `approval-engine`'s contract before the
+evidence half is treated as load-bearing.
+
+## Boundary audit-core states explicitly
+
+`approval-engine` will be registered as a **distinct source**, with its own
+sender registration, tenancy mapping, retention profile, and `secret_policy`
+under the existing onboarding rules (principle 4: a source is onboarded when
+ownership, retention, access, export, and evidence policy are declared and
+validated — not when events start arriving).
+
+**audit-core will not expose an approval-validity query.** No read surface will
+answer "is this approval still valid". Records, yes; a verdict, never. Such a
+surface would be deciding early under §6.1, and a consumer branching on an
+audit-core answer would be routing an authorization decision through the audit
+fabric — the exact coupling §9.4 exists to prevent. Callers who need current
+state ask `approval-engine`.
+
+The four event classes accepted as a distinct source: **issuance, use,
+supersession, revocation**, per §9.4.
+
+## Conformance note raised in passing
+
+§11 requires that every repository in §4 declare its layer in `INTENT.md`.
+audit-core declares none today. It is corrected in the same change: audit-core
+is **Engine layer** — it exposes a deterministic API contract and holds
+operational custody — and it is explicitly *not* a decision point, per the
+boundary stated above.
+
+---
+
+## What changed in this repository
+
+| File | Change |
+| --- | --- |
+| `INTENT.md` | layer declared; approval evidence named as a source; principle 6 cross-referenced to the `docs/integrity.md` bound |
+| `intakes/intakes.md` | `AUDIT-IN-0001` closed, outcome `assent-with-conditions` |
+| `history/` | this record |
+
+Reply sent to gate-house on the State Hub inbox thread.

intakes/intakes.md

@@ -7,7 +7,7 @@ id: AUDIT-IN-0001
 kind: intake
 title: 'Proposed: audit-core takes the approval evidence half (security layer model
   v0.3 §9.4)'
-status: open
+status: closed
 origin: cross-repo
 origin_ref: net-kingdom security-layer-model_v0.3 §9.4
 priority: medium
@@ -30,5 +30,26 @@ description: 'gate-house proposes that audit-core own the tamper-evident record
   approval events as a distinct source, say so and gate-house will record the evidence
   half as unowned rather than assume it.'
 created: '2026-08-28T20:35:09.148892Z'
-updated: '2026-08-28T20:35:09.148892Z'
+updated: '2026-08-28T21:10:00Z'
+outcome: assent-with-conditions
+closed: '2026-08-28T21:10:00Z'
+resolution: 'Assent. The split is right: approval-engine owns the operative state,
+  audit-core owns the tamper-evident record of issuance, use, supersession, and
+  revocation as a distinct source. Two corrections to the rationale and one
+  condition. (1) INTENT principle 6 overstates the delivered guarantee;
+  docs/integrity.md is authoritative — an in-database chain does not withstand a
+  database owner without the external chain-head attestation, and even then it is
+  not WORM. Approval events get exactly the guarantee every other source gets,
+  no more; anything stronger is an engine gap to declare under §5.3. (2) The chain
+  proves alteration and truncation, not omission at source. A suppressed
+  revocation leaves the chain intact and verify reports intact; completeness at
+  the boundary is the emitter''s property, not the archive''s. CONDITION:
+  approval-engine must guarantee emission atomicity (transactional outbox or
+  equivalent) so an approval cannot change state without the event being durably
+  queued in the same transaction — a requirement on approval-engine, not a task
+  audit-core can discharge for it. Boundary stated: no approval-validity query
+  will ever be exposed; a verdict surface would be deciding early under §6.1.
+  Also raised: audit-core declared no layer, contrary to §11 — now declared Engine
+  layer, explicitly not a decision point.'
+recorded_in: history/2026-08-28-approval-evidence-assent.md
 ```