risk-nexus/findings/RISK-F-0008-audit-retention-legal-basis-assumed.md

---
id: RISK-F-0008
type: finding
title: "The legal basis for retaining audit facts against an erasure request has been assumed, never established"
status: open
reported_by: audit-core
reported_via: audit-core
routed_by: audit-core
date_reported: "2026-08-18"
date_filed: "2026-08-19"
system: audit-core
environment: production
fix_owner: risk-nexus
fix_tracking: unset
related: [RISK-F-0005]
supersedes: RISK-N-0002
# Graded by risk-nexus 2026-08-19 — docs/rulings/2026-08-19-third-grading.md
severity: medium
severity_at_production: high
impact: I3
likelihood: L2
fidelity_modifier: false
production_rescore: true
disclosure: public
publication: pending-handover
escalation: required
escalation_trigger: 2
escalation_status: partially-answered
escalation_answered: "2026-08-20"
escalation_answered_by: the-custodian
escalation_act: rule
decision: "identity in audit records: opaque subject ids preferred, agent identifiers where possible, operator credentials only where necessary, policy decisions tracked to the responsible party; zone-level privacy guarantees may raise the floor"
outstanding: "a defensible retention period per category (waits on audit-core's co-residency horizon), and the trigger list for buying an external answer"
determination: RISK-REG-0001
last_checked: "2026-08-20T10:02:41Z"
next_check: "2026-08-20T10:02:41Z"
cadence: instant
clean_streak: 0
waiting_on:
  - who: audit-core
    what: "does a keyed commitment restore erasability without breaking chain verification; what is the platform-pg co-residency horizon"
    since: "2026-08-20"
    would_change: "a working keyed commitment narrows RISK-REG-0001 to retained-by-obligation categories only"
    default: "encrypt-then-hash recorded as the only known route, and the retention period recorded as unstateable"
    default_at: "2026-11-17"
  - who: the-custodian
    what: "rule the trigger list for buying an external determination"
    since: "2026-08-19"
    would_change: "fixes when the estate stops running on an assumption"
    default: "the assumption is recorded in the register as an assumption"
    default_at: "2026-11-17"
graded_by: risk-nexus
ruling: RISK-RULING-2026-08-19-C
---

# RISK-F-0008 — the exemption nobody has established

## What is true

`audit-core` holds audit evidence across tenants, targets `R2` on the Tenancy
Posture retention ladder, and has declared `R4` — verified erasure —
unreachable by design. The technical reasoning is sound and documented
(`audit-core/docs/erasure-and-audit.md`, framework Decision 4.5.3):
crypto-shredding would destroy the evidence the service exists to hold, and
their integrity chain commits to a SHA-256 of the cleartext record, which
survives key destruction as a confirmation oracle against low-entropy audit
rows. Destroying a key does not erase content a surviving commitment can still
be tested against.

The consequence is that if an Article 17 request arrives naming a data subject
in the audit trail, `audit-core` has no mechanism. The answer would rest on
audit evidence being exempt — legal obligation, or legitimate interest in fraud
and security investigation.

**Those grounds are ordinary. Nobody in this estate has actually reached them.**
`audit-core` routed the question here on 2026-08-18 rather than absorbing it,
saying plainly that they are not competent to answer it and that they have been
assuming it. §19.11 of the framework says the same in its own words: the legal
basis for retaining audit facts remains a risk/legal question outside the
framework.

## Why this repo owns it

This is the first finding where `fix_owner` is `risk-nexus`.

`INTENT.md` moved regulatory intake here from `policy-nexus` on 2026-08-17,
precisely because deciding what a rule demands of us is a judgement about risk
rather than an act of publishing. `audit-core` routed it by both available
routes and asked for an owner rather than an opinion. Refusing it would be this
repo declining its own remit.

What this repo owns is the **record**: what the source says, when, and what
therefore is or is not established. It does not own legal advice — `INTENT.md`
is explicit — and it does not own the redesign. If the basis does not hold,
`audit-core` owns encrypt-then-hash at accept time, and that is not
retrofittable onto events already accepted.

## The three questions, as asked

1. On what basis does the estate retain personal data inside audit records
   against an erasure request, and does that basis hold for the categories
   `audit-core` stores?
2. Does it hold across the full 30-day recoverable window and beyond, given
   that at `P1` the real erasure horizon is the maximum across every
   co-resident on `platform-pg`, not the value `audit-core` declares?
3. If it does not hold, `R4` is urgent rather than theoretical, and the answer
   is a substantial redesign with a long lead time.

## Register ruling — 2026-08-19

`medium` today (`I3` × `L2`), `high` at production, `public`, **escalated on
trigger 2**.

`I3`: an unmet retention obligation in the audit store crosses from a technical
question to an obligation with an outside counterparty, and the remediation is
a non-retrofittable redesign rather than a patch. `L2`: no request has arrived
and the estate holds no real data subject's records yet, but the trigger is
somebody else's to pull and needs no foothold here.

`production_rescore: true`. The likelihood of an Article 17 request is a
function of having real users; that is exactly what production means.

**Escalation, trigger 2** — "creates or reveals an obligation with an outside
counterparty". It reveals one. The estate cannot decide unilaterally that this
obligation is small, and the operator is the only party who can commission an
answer that is more than an assumption. The ask is narrow: authorise someone to
establish the basis, or record that the estate knowingly runs on the assumption
and for how long.

**Disclosure `public`.** Nothing here shortens a path to a defect: it is a
question about a legal basis, published as a question. `audit-core`'s technical
reasoning is already written down and worth reading.

## How it got here

Ruled a note on 2026-08-19 (`RISK-N-0002`) on the reasoning that no obligation
exists yet. That ruling was made without reading `audit-core`'s message, which
had been in this repo's inbox since 2026-08-18 and asks specifically for an
owner. The note was wrong on the second floor test: recording this *does*
change a decision, because the redesign it might force cannot be retrofitted
and therefore has to be decided early or not at all.

`RISK-N-0002` is superseded by this record.

## Reviews

- **2026-08-19** — promoted from note, graded, escalated. Open at review:
  has the basis been established or the assumption recorded; has anything
  changed about what categories `audit-core` stores.

## Suggested disposition — 2026-08-20, proposed by risk-nexus

Offered because this repo owns the finding and the operator asked for a
direction. It is not legal advice, and this repo cannot make it into one: what
follows is a *route to an answer* and a hedge against the answer being no.

### The reframe: the expensive thing is not the legal question

`audit-core` asks whether the exemption holds. That question is cheap to
answer badly and expensive to answer properly, and the temptation is to
schedule the proper version and wait.

But the cost of a "no" is not fixed — it grows daily. The remedy they name,
encrypt-then-hash at accept time, cannot be retrofitted onto events already
accepted. **Every day the estate accepts events under the current scheme, the
un-erasable set grows by one day.** So the decision that actually needs taking
now is not "is it exempt" but "do we keep manufacturing records we could never
erase while we find out".

That splits the finding into two decisions with very different prices.

### 1. Establish the basis internally, now, for the cost of an afternoon

Not a legal opinion — a **written determination** that says which ground is
being relied on, for which category of data, and for how long. Today the
estate has no such document; that is the whole finding.

The shape it should take, per category of personal data in the audit trail:

| Category | Likely ground | The part that is actually arguable |
| --- | --- | --- |
| Operator and agent identifiers | Art 6(1)(f) legitimate interest in security, with Recital 49 squarely on point | little — this is the ordinary case |
| Counterparty or end-user identifiers in event payloads | Art 17(3)(e), defence of legal claims; Art 6(1)(f) | **duration**, not existence |
| Commercial records that happen to pass through audit | Art 17(3)(b) plus German §257 HGB / §147 AO retention | scope — retention duties cover books and invoices, not application logs generally |

Where such determinations usually fail is **not** the ground. It is the
retention period: a blanket "we keep audit forever under legitimate interest"
is much weaker than "we keep these fields for N months because X". That lands
precisely on `audit-core`'s question 2 — the real horizon being the maximum
across every co-resident on `platform-pg` rather than the declared value.

Recording the determination converts an assumption into a position that can be
argued with. That is what this register exists to produce, and it does not
require a lawyer to write down.

### 2. Stop the un-erasable set from growing — a cheaper hedge than the redesign

`audit-core`'s stated obstacle is precise and correct: their chain commits to
`SHA-256(cleartext)`, audit records are low-entropy, so the retained hash
survives key destruction as a confirmation oracle. Guess, hash, compare.

The oracle exists because the commitment is over cleartext with no secret in
it. A **keyed commitment** removes it: replace the digest with an HMAC (or a
hash over record plus a high-entropy per-subject salt) where the key or salt
lives outside the audit store and is destroyable per subject.

What that buys, and why it is cheaper than the redesign they costed:

- Destroying the per-subject key makes the commitment untestable — no guess
  can be confirmed. That is crypto-shredding restored, which their analysis
  correctly found unavailable under a plain hash.
- The integrity chain still verifies. It chains over commitment values, and
  those persist after key destruction; what is lost is the ability to
  re-derive a commitment from cleartext, which is exactly what erasure means.
- It is a change at accept time only. No re-processing of stored events, no
  new storage layer, no change to the read path.

This is a suggestion to `audit-core`, not an instruction, and they own whether
it is sound — they know their chain and this repo does not. The claim worth
testing with them is narrow: **does a keyed commitment restore erasability
without breaking chain verification?** If yes, the expensive redesign becomes a
contingency rather than a plan, and the daily accrual stops.

### 3. Buy the real answer only when something triggers it

An external determination costs money and needs a real question. Propose three
triggers, any of which fires it:

- the estate first holds a real person's data;
- a counterparty contract requires a stated erasure position;
- an actual Art 17 request arrives.

Until one fires, the internal determination plus the hedge is a proportionate
posture, and `severity_at_production: high` plus `production_rescore: true`
already guarantee this is re-read before production completes.

### What this repo would record if the operator agrees

`status: accepted` with the determination attached, `escalation` answered as
`rule`, and the review kept at 90 days. The finding stays open and visible
until the determination exists — an accepted risk with no written basis is the
same assumption it started as, wearing a different word.

### Also worth saying, because it is the cheapest fix of all

Every field of personal data that never enters the audit trail is a field with
no erasure question. Where an opaque subject identifier would carry the same
evidentiary weight as a name or an address, the identifier is strictly better,
and that is a `audit-core` design choice available today at no legal cost.

## Operator decision — 2026-08-20: minimise the identity, keep the accountability

The custodian ruled on what goes into an audit record, which is the half of
this finding that shrinks the question rather than answering it:

1. **Opaque subject identifiers are preferred.** Where an opaque id carries the
   same evidentiary weight as a name or an address, it is the id that goes in.
2. **Agent identifiers where possible.** Agents act; attribute to the acting
   agent identity rather than to a person behind it.
3. **Operator credentials only where necessary.** Not as a convenience, not as
   a default — where the record genuinely requires the operator.
4. **Policy decisions are tracked to the responsible party.** Accountability is
   preserved by linking a decision to who is answerable for it, not by
   retaining personal data in the record itself.
5. **Zone guarantees may raise the floor.** If a zone establishes additional
   privacy, pseudonymity or anonymity guarantees, those apply — the current
   level is not a permanent ceiling. That work is `zone-engine`'s
   (`ZONE-WP-0001`), and this finding should be re-read when a zone lands one.

**Why this is more than a preference.** Personal data that never enters the
audit trail has no erasure question, no exemption to establish, and nothing to
argue about with a regulator. Points 1-3 shrink the population the legal basis
has to cover; point 4 is what stops that shrinking from costing accountability,
which is the usual objection to minimising an audit log.

It also changes the shape of the accrual problem. The un-erasable set still
grows daily, but each day's records now carry less that would need erasing —
so the cost of a "no" answer falls with every event accepted under the new
rule rather than rising.

**What is still outstanding**, and stays escalated:

- The **written determination** of the retention basis — which ground, for
  which category, for how long. `risk-nexus` owns writing it; it needs no
  further authorisation and is scheduled into the next workplan.
- The **trigger list** for buying an external answer (first real person's data,
  first counterparty contract requiring a stated position, first Art 17
  request). Proposed, not yet ruled on.

The escalation is therefore `partially-answered`, not closed. `make check` will
keep listing it.

**Routed to `audit-core` on 2026-08-20**, together with the keyed-commitment
question — which remains theirs to judge, because they know their chain.

## The determination exists — 2026-08-20

`docs/regulatory/RISK-REG-0001` (`audit-retention-basis.md`). The estate now
has a written position rather than an assumption, which was this finding's
substance.

What it says, in short: Art 6(1)(f) with Art 32 for operator and agent audit
records; Art 17(3)(e) for records evidencing a counterparty transaction;
Art 17(3)(b) only where a commercial or tax retention duty independently
applies, and not extended to application logs generally.

**The weak part is duration, not existence**, and the record says so rather
than sounding confident. A position of the form "we keep audit forever because
it is audit" is the one that fails; a period per category is what holds. The
estate does not have one yet, and the reason is `audit-core`'s own question 2 —
at `P1` the real horizon is the maximum across every co-resident on
`platform-pg`, not the declared value. **That infrastructure fact is the most
likely point of failure in the whole position.**

The operator's minimisation ruling improves this materially: it shrinks the
category whose retention is hardest to justify, leaving mostly the row where
the ground is strong. A weak argument avoided by holding less data beats a
strong one relied upon.

The finding stays open. What remains is a retention period per category, which
waits on the co-residency horizon, and the trigger list for buying an external
determination. The record is reviewed every 90 days with this finding, or
immediately on any trigger.
- **2026-08-20** — not clean: The determination now exists: RISK-REG-0001 states the grounds per category and names duration as the weak point. Cadence instant → instant; checked again immediately.