docs(security): ratify AR-1 MFA for platform roles

AR-1 option B (BDR-083): aal2 for every platform role change. Contract 5.7 rule, path inventory, bootstrap, factors, recovery, sessions; F17-F19; M4 scope, tests, DEV preconditions; fresh security review integrated (19.4).
This commit is contained in:
bchanot committed 2026-10-28 12:20:11 +01:00
1 parent 41d3b09120
commit 9c13bae4bc
5 files changed
+447 -57

No files matched your search

+10
View File
@@ -603,3 +603,13 @@ Supersedes for official sheets the 2026-10-25 rule « no graphic editor until St
Consolidation also fixed: rights table moved to M4 (M4 used table created in M5); core step V1-2.2 now before worker step V1-2.3; progressive review via per-group acknowledgments with digests.
Alt: options not chosen listed in contract §17 and the 2026-10-28 review; O15 B and O4 C rejected outright.
Status: ratified; contract under owner final review; V1-2.1 not authorised.
## BDR-083 | 2026-10-28 | AR-1 ratified option B: aal2 for every platform role change
Decision (owner brief « V1-2.0 : résolution définitive d'AR-1 »): grant, revoke, replace, promote, demote of any platform role (owner, admin, moderator, support) need valid aal2 session, checked server-side in RPC at call time. No permanent exception (owner, first admin, old accounts, admin UI, direct RPC). aal2 necessary never sufficient: can_manage_platform_role, self-target refusal, last-owner guard stay. Ordinary users no MFA. Completes O4, no change to D1-D9, O1-O23.
Why: aal1 owner session could grant owner to attacker account → enrol own TOTP → aal2 → self-grant all catalogue capabilities. Read from `20261008100200` RPCs.
Shape (contract §5.7): aal2 check 2nd after auth, before role checks (no role oracle); private assign/withdraw re-check for non-null actor; service_role revoked on profile_roles/profiles + RPC EXECUTE (F17); owner-count race locked (F18); first owner: enrol 2 factors first, role assigned last by operator; DEV owner: enrol before M4 applied; recovery operator-only, ban account during enrol window; GoTrue v2.197.0 enrol/unenroll need aal2 once verified factor (source read, sandbox proves).
Alt: A (owner+admin only) rejected: leaves moderator/support weak path; C (accept residual) rejected.
Review (fresh challenger FATAL(4), all integrated, no ratified decision changed): grant needs target with verified factor; private fns: non-null actor = auth.uid() + aal2, null actor only without JWT; advisory lock both RPCs; DEV precondition before M4 = ENVIRONMENTS §3.6 hardening (public default JWT secret + demo service key + 54321 LAN → aal2 forgeable) + seed owner password changed + role holders listed; recovery reordered (ban, password set with owner, unban, enrol).
Reserves: staff read policies (is_platform_admin) stay aal1, owner decision; stolen aal2 session residual; refused attempts not in audit table.
Status: ratified; contract only, no code/migration/DEV change; V1-2.1 not authorised.
+5
View File
@@ -467,3 +467,8 @@ R5.4: `draft/{labels,changelog,review_report,drawing,overlay,app_review,svg_comp
- Contract rewritten: rights M4, catalogue access fail-closed, MFA + operator recovery, group acknowledgments, worker verdict states, species inactive, equine working copy, SQL mirror removal, steps §16 with 9 headings.
- 3 fresh challengers: business CONCERNS, security FATAL (AR-1: aal1 owner grants owner role → bypasses catalogue MFA; O4-D unratified → owner arbitration), simplicity CONCERNS. All other findings integrated (omitted status = inherited, F12 unconditional, F15/F16, reader row predicate, uploads to M5, PreviewSheet editor handle, removal groups).
- Docs only, committed + pushed on feature/v1-2-anatomical-sheets. STOP: verdict BLOCKED on AR-1.
## 2026-10-28 (V1-2.0 AR-1 B ratified, contract updated)
- Owner ratified AR-1 B (BDR-083). Contract §5.7 written: rule, chain, path inventory (2 RPCs, 2 private fns, service_role table grants F17, owner-count race F18), bootstrap, factor/password/recovery, session rules, deploy order (enrol before M4 on DEV), tests §15.
- GoTrue v2.197.0 source read: enrol needs aal2 once verified factor, unenroll verified needs aal2. auth.sessions check kept as to-verify.
- Docs only. Independent security review, then commit + push feature branch. STOP at checkpoint.
+2 -1
View File
@@ -535,7 +535,8 @@ Order (BDR-071): P0 docs + decisions → V1-1 clients and animals → V1-2 sheet
- [ ] V1-2.0 contract and decisions (docs only): read-only audit delivered; contract drafted; 3 fresh plan challengers (correctness, robustness, simplicity); doc fixes (V1-1 status in TODO + V1-1 contract header, CLAUDE.md M1-M3 line); challengers FATAL(2 blockers)/FATAL(1)/CONCERNS integrated (contract §19); STOP at checkpoint: owner ratification of §5-§16 and options O1-O20
- [x] owner review of O1-O20 delivered (2026-10-28, read-only); O21-O23 proposed (catalogue access, SQL mirror, failed verification)
- [x] owner ratified O1-O23 (brief « V1-2.0 : ratification des arbitrages O1 à O23 et consolidation du contrat », 2026-10-28, BDR-082)
- [ ] consolidation (docs only): contract rewritten (decisions §17, contradictions C1-C14 §19.2, steps §16 with 9 headings each); IMPLEMENTATION_PLAN §0 + Steps 13/17 notes, ANATOMICAL_SHEET_MODEL stale Zod/TS mentions + V1-2 pointer, OPEN_QUESTIONS conflicts 1-2, V1-1 §5.4 pointer (N1 amended), rule sheet-model, CLAUDE.md pointer; 3 fresh challengers (business CONCERNS 3 major, security FATAL 1 arbitration + 6 major, simplicity CONCERNS 5 major) integrated (contract §19.3); commit + push branch; STOP at checkpoint: owner final review + AR-1 (MFA on platform role grants, contract §17.2), no V1-2.1
- [ ] consolidation (docs only): contract rewritten (decisions §17, contradictions C1-C14 §19.2, steps §16 with 9 headings each); IMPLEMENTATION_PLAN §0 + Steps 13/17 notes, ANATOMICAL_SHEET_MODEL stale Zod/TS mentions + V1-2 pointer, OPEN_QUESTIONS conflicts 1-2, V1-1 §5.4 pointer (N1 amended), rule sheet-model, CLAUDE.md pointer; 3 fresh challengers (business CONCERNS 3 major, security FATAL 1 arbitration + 6 major, simplicity CONCERNS 5 major) integrated (contract §19.3); commit + push branch; STOP at checkpoint: owner final review, no V1-2.1
- [x] AR-1 ratified option B by the owner 2026-10-28 (BDR-083): `aal2` on every platform role change; contract §5.7 (rule, inventory, bootstrap, factors, recovery, sessions), F17-F19, M4 scope, tests §15, DEV order §16.11 (enrol before M4); docs only; independent security review then commit + push; no code, migration or DEV change
- [ ] V1-2.1 server access and workflow (M4, M5) · V1-2.2 core verdict, preview, change list, edit primitives, SVG cleaning · V1-2.3 drawings and worker verification (M6) · V1-2.4 wireframes (editor, review, admin entry) · V1-2.5 catalogue, admin space, rights, species, MFA · V1-2.6 editor level 1 · V1-2.7 preview, review, publication UI · V1-2.8 equine working copy and review preparation · V1-2.9 FR/EN acceptance and closure (owner checkpoints V1-2.3, V1-2.4, V1-2.6, V1-2.7, V1-2.8, V1-2.9 + each DEV operation of contract §16.11)
## Release readiness track (owner 2026-10-25, distinct from the functional roadmap; nothing here authorises a release or exposure)
+3
View File
@@ -90,6 +90,9 @@ enforced by an RLS helper `has_mfa()` on admin tables. Today:
- app: no enrolment UI (QR + code), no `aal2` step-up on sign-in;
- there is no admin table or screen yet to protect.
V1-2 decision (BDR-083): `aal2` is required for every platform role change,
see `V1-2_ANATOMICAL_SHEETS.md` §5.7 (rule, bootstrap, recovery).
The Security section shows a note instead. To close the gap (a later step):
migration adding `private.has_mfa()` and folding it into `is_platform_*` (or
into the policies of the first admin table) with pgTAP allowed/denied cases;
+427 -56
View File
@@ -1,14 +1,17 @@
# V1-2 Anatomical sheets: contract
Status: decisions RATIFIED, contract consolidated for the owner's final
review (V1-2.0, 2026-10-28), with ONE OPEN ARBITRATION REQUEST (AR-1, §17.2).
review (V1-2.0, 2026-10-28). The arbitration AR-1 is closed: the owner
ratified option B (`aal2` for every platform role change, §5.7, §17.2,
BDR-083).
The owner approved the orientations D1 to D8 and added the mandatory
requirement D9 (BDR-081, 2026-10-27), then ratified the decisions O1 to O23 in
the brief « V1-2.0 : ratification des arbitrages O1 à O23 et consolidation du
contrat » (2026-10-28, BDR-082). This revision writes those decisions into the
rules, resolves the contradictions C1 to C14 found at the review of the first
revision (§19.2), restates the step plan (§16) and integrates a second round
of fresh challenges (§19.3), one finding of which needs the owner (AR-1). It
of fresh challenges (§19.3), one finding of which needed the owner (AR-1,
decided) and a third review of the AR-1 B text (§19.4). It
authorises no implementation: V1-2.1 starts only on the owner's explicit go.
V1-2.0 changes no code, migration, policy, RPC, anatomical data or DEV state.
@@ -31,7 +34,7 @@ Parts: 1 facts, 2 scope, 3 nomenclature, 4 decisions D1 to D9, 5 access, roles
and permissions, 6 lifecycle, review, verification and publication, 7 species
and new sheets, 8 editor, 9 drawings, 10 loading, 11 screens and journeys,
12 SQL defects, SQL authority and migrations, 13 security, 14 equine sheet,
15 tests, 16 implementation steps, 17 ratified decisions and open arbitration,
15 tests, 16 implementation steps, 17 ratified decisions and arbitration,
18 documents touched by V1-2.0, 19 challenge record.
---
@@ -192,7 +195,7 @@ Each layer is checked on the server; none implies the next.
| Historical reads | the same predicate (O8) | published and deprecated versions by id, archived sheets included |
| Draft reads | a capability and `aal2` | drafts, rounds, verdicts, comments, history, drawings of drafts |
| Capabilities | `anatomical_catalog_grants` (§5.2) and `aal2` | edit, review, publish, per the matrix of §5.4 |
| Platform role | `profile_roles` | a precondition of `sheet_editor` and `sheet_publisher`; no catalogue right by itself |
| Platform role | `profile_roles`, changed only with `aal2` (§5.7) | a precondition of `sheet_editor` and `sheet_publisher`; no catalogue right by itself |
| Professional status | plan Step 9, not built | none; never inferred from a right or a role |
| Workspace membership | V1-1 tables | none: no catalogue rule reads workspace tables and no V1-1 rule reads catalogue tables |
@@ -374,14 +377,17 @@ A1 to A6):
- Enrolment (V1-2.5, Mon compte, Sécurité): add a TOTP factor (QR code, the
secret shown once and never stored or logged by the application), verify
it with a code (a factor counts only once verified), name, list and remove
factors. Supabase requires an `aal2` session to remove a verified factor.
factors. GoTrue v2.197.0 requires `aal2` to enrol a factor once one is
verified and to remove a verified factor (§5.7.3, from its source).
V1-2.5 proves that an `aal1` session of an account with a verified factor
cannot enrol another factor; if Supabase allowed it, the step stops for an
owner decision, since MFA would then be bypassable with the password alone.
can neither enrol another factor nor remove one; if the sandbox shows
otherwise, the step stops for an owner decision, since MFA would then be
bypassable with the password alone.
- Step-up: entering `/app/admin` with an `aal1` session opens the challenge;
success upgrades the session; no admin data is requested before.
- Sessions: `aal2` belongs to the session. A token refresh keeps it, sign-out
ends it, a new sign-in starts at `aal1` (Supabase behaviour, proven by a
ends the session (an access token already issued stays valid until
`jwt_expiry`), a new sign-in starts at `aal1` (Supabase behaviour, proven by a
V1-2.5 browser test). The access token lives `jwt_expiry` (3 600 s in the
local configuration; V1-2.1 records the values of the sandbox and DEV). No
session time box or inactivity limit is configured, and V1-2 adds no
@@ -408,14 +414,297 @@ A1 to A6):
The owner's own loss follows the same procedure, run by the owner at the
server. The runbook goes to `IDENTITY.md` and is rehearsed on the sandbox in
V1-2.9.
- Platform roles (O4-D): `grant_platform_role` and `revoke_platform_role`
require no `aal2`, and an owner may grant the owner role. The first revision
of this contract claimed that the catalogue stays closed without O4-D; the
security challenge of §19.3 showed it does not (an `aal1` owner session
grants owner to another account, which enrols its own factor and grants
itself every capability). O4-D was not ratified, so the contract does not
extend `aal2` to those RPCs: the choice is the open arbitration AR-1
(§17.2).
- Platform roles (AR-1 B, ratified 2026-10-28): `grant_platform_role` and
`revoke_platform_role` require `aal2` for every platform role; see §5.7.
The earlier option O4-D, left unratified by O4, is superseded by that
arbitration.
### 5.7 Platform role changes and MFA (AR-1 B)
Ratified by the owner on 2026-10-28 (BDR-083). It completes the security
contract of O4; it replaces none of O1 to O23.
**Rule.** Every grant, revocation, replacement, promotion, demotion or
equivalent change of a platform role (`owner`, `admin`, `moderator`,
`support`) requires a valid `aal2` session, checked on the server when the
operation runs. A hidden button is not a control. There is no exception for
the owner, the first administrator, accounts already privileged, older
accounts, the administration screens or direct RPC calls; the only path
without a session is the audited operator path outside the application
(§5.7.5), used for a bootstrap or a recovery. `aal2` is necessary
and never sufficient: `can_manage_platform_role`, the self-target refusal and
the last-owner guard still apply. Ordinary accounts are not asked for MFA; the
requirement covers the operations of this contract (platform roles, catalogue
rights, catalogue administration, protected drafts, review, publication).
Reading the published catalogue follows O21.
**The flaw and the chain it closes.** `grant_platform_role` and
`revoke_platform_role` check the caller's role, not its assurance level:
1. an attacker learns the password of an owner;
2. they open a session, which is `aal1`;
3. they call `grant_platform_role` directly and give `owner` or `admin` to an
account they control (or revoke the real owner);
4. that account enrols its own TOTP factor (a first factor is enrolled at
`aal1`) and reaches `aal2`;
5. it grants itself `sheet_editor`, `sheet_reviewer`, `sheet_publisher`
(§5.2) and reviews and publishes official sheets.
With AR-1 B, step 3 is refused with no write, so steps 4 and 5 never start.
**Why B and not A.** A (owner and admin only) leaves `moderator` and `support`
changeable from a password-only session. They give no catalogue right today;
`support` opens staff reads (`profiles_select_staff`), and any future
role-based rule would inherit a weak path; revoking them is also a nuisance
attack. B
costs the same one predicate and removes the list of roles from the rule.
**Inventory of the paths that change platform privileges** (read from the
migrations and `seed.sql` on 2026-10-28; V1-2.1 re-runs the census in pgTAP
001):
| Path | Who may call it today | Controls today | Added by M4 | Audit |
|---|---|---|---|---|
| `public.grant_platform_role(target, role)` | `authenticated` (and `service_role`, which fails at `auth.uid()` null) | authenticated, not self, `can_manage_platform_role` (owner: any role; admin: moderator and support), not already active | `has_mfa()` as the second check; `EXECUTE` revoked from `service_role` | `platform_role.granted` |
| `public.revoke_platform_role(target, role, reason)` | same | same, plus the last-owner guard | same, plus the guard made race-free (F18) | `platform_role.revoked` |
| `private.assign_platform_role`, `private.withdraw_platform_role` | no API role (`EXECUTE` revoked from public, schema `private` not exposed); RPCs above and privileged SQL sessions | none of their own | a non-null actor must equal `auth.uid()` and pass `has_mfa()`; a null actor is accepted only when the request carries no JWT (`auth.role()` null), so neither an API session nor `service_role` reaches the operator path (§5.7.5) | same events |
| `private.can_manage_platform_role` and the `has_platform_role` family | predicates, no write | n/a | unchanged; `is_platform_*` do not read the assurance level (policies that use them for reads are out of AR-1, see reserves) | none |
| table `public.profile_roles` | `authenticated`: SELECT only, no write policy; `service_role`: default table privileges never revoked (F17) | RLS, privileges | `REVOKE ALL` from `service_role` on `profile_roles` and `profiles` (F17) | none |
| `profiles.active` (deactivation changes who counts as a role holder) | no user path: not in the column grant of UPDATE | column grants | F17 covers `service_role` | none |
| triggers and `handle_new_auth_user` | none writes `profile_roles` | n/a | census asserts it stays so | none |
| `seed.sql` bootstrap (`assign_platform_role(null, ...)`) | local stacks, privileged SQL | none | unchanged, local only (§5.7.5) | `platform_role.granted`, null actor |
| deletion of an auth user (cascade `on delete cascade` to `profile_roles`) | GoTrue admin API with the service key | none | operator procedure only (§5.7.5) | auth log |
| `grant_catalog_right`, `revoke_catalog_right` (M4, §5.2) | owner | owner, `aal2` | already `aal2` | per §5.2 |
No other path reachable from a user session was found: PostgREST exposes the
`public` schema only, and every write to `profile_roles` goes through the two
RPCs. A claim that this closes service-key abuse would be false: see §5.7.5.
**Order of checks in both RPCs** (M4), all before any write:
1. `auth.uid()` is not null;
2. `has_mfa()`; otherwise a refusal with the closed kind `mfa_required`;
3. the caller is not the target;
4. `can_manage_platform_role`;
5. the operation is valid: target exists and is active; for a grant, the
target holds a verified factor (the eligibility predicate of §5.2), so no
role lands on an account whose first factor a password or mailbox holder
could still enrol; role not already active, or active for a revocation;
last active owner kept;
6. the write and its audit row, in the same transaction.
The assurance check comes before the role checks so that a password-only
session gets the same answer whatever role it holds: no oracle on who is
owner or admin. A refusal raises, so it leaves no row (neither the role nor an
audit row); attempted escalations are therefore visible in the API logs, not
in the audit table. With `[analytics] enabled = false` a raised refusal
reaches the PostgreSQL log only (residual). Returning the refusal as a
closed kind after a committed `log_event` is an option V1-2.1 may propose,
not a decision taken here.
**Race on the last owner (F18).** `withdraw_platform_role` counts owner rows
without a lock and without `profiles.active`, so two simultaneous
revocations of two owners can both pass and leave none (or only a deactivated
one). M4 takes one transaction-scoped advisory lock at the start of both RPCs
and of the private functions, before checks 3 to 5, and counts owners with
the `has_platform_role` semantics (active profile). Two-session tests: mutual
revocation, revoke against grant.
#### 5.7.1 First enrolment and the bootstrap of the first owner
GoTrue v2.197.0 (the image pinned by the local stack; source read on
2026-10-28, behaviour to be proven on the sandbox by V1-2.5):
`EnrollFactor` refuses with `insufficient_aal` when the user already has a
verified factor and the session is not `aal2`. With no verified factor, a first
factor is enrolled from an `aal1` session. Enrolling, verifying and reaching
`aal2` are three steps: a factor counts once verified, and verification
upgrades the session and invalidates the user's sessions below `aal2`.
Ordinary journey: password sign-in (`aal1`) → Mon compte, Sécurité → enrol
TOTP → verify with a code → `aal2`. Enrolling a factor grants no role and no
capability; MFA is a condition, never a substitute for the checks of §5.2 and
§5.7.
New environment (no owner, no `aal2` session anywhere), in this order, never
the reverse:
1. the operator creates the account through the sign-up path or the admin
API, email confirmed, and hands over the initial secret out of band;
2. the future owner signs in, enrols two TOTP factors on two devices and
verifies both; no role exists yet, so nothing is at stake in the window;
3. the operator checks, read-only, that the account has verified factors and
that their names and creation times match what the owner did;
4. only then the operator assigns the owner role from a privileged SQL
session (`assign_platform_role(null, ...)`) and records, in the same
session, an audit row with the reason and the operator's name through
`private.log_event`;
5. the owner signs in again, passes the step-up, and proves an `aal2` call
works and an `aal1` call is refused.
No public RPC creates an owner, none disables MFA, and no permanent bypass
exists. The `seed.sql` bootstrap (role assigned before any factor, known
password) is a local-fixture exception to this order, acceptable only on a
stack no untrusted party can reach (§5.7.2). Assigning the role last is what closes trust on first use for the
owner account.
#### 5.7.2 DEV: preconditions of the future transition
The DEV owner is the seed account, it has no factor, and nothing changes on
DEV during V1-2.0. Two facts of DEV defeat AR-1 B as long as they hold:
- the stack signs tokens with the Supabase CLI's public default JWT secret,
its `service_role` key is the public demo key, and port 54321 is open to
the LAN (`ENVIRONMENTS.md` §3.6, risk accepted by the owner on 2026-10-24
« before any untrusted device or user »). Anyone on the network can mint an
`aal2` token for the owner, or act as `service_role`; `has_mfa()` proves
nothing there;
- the owner account's password is the seed's, committed in Git.
So, before the owner applies M4 on DEV or makes the first grants (§16.11):
0. the §3.6 hardening is done (own JWT secret or signing keys, or 54321
narrowed to the test devices; the owner's choice), and the seed owner's
password is changed, or a separate owner is bootstrapped per §5.7.1 and
the seed owner revoked; every active `profile_roles` row of DEV is listed,
and each holder enrols and verifies a factor before M4 or loses the role;
1. read-only check of every factor of each role holder, verified or not (an
unknown one is deleted, and an unknown verified one means a reset first,
§5.7.7);
2. the owner enrols two factors and verifies both from the branch's Mon
compte; this needs no migration, and until M4 is applied every existing
right is unchanged;
3. on the sandbox rehearsal of M4 from the DEV state, `aal2` is shown to
work and `aal1` to be refused for the same account;
4. the recovery of §5.7.7 has been rehearsed on the sandbox;
5. a backup exists.
Any `db:reset`, or `db:start` on a fresh volume, of DEV after M4 reruns the
seed and reopens these preconditions.
A failed transition is never a lockout: an owner without a verified factor
still signs in and uses the application at `aal1`, only role and rights
management waits, and the first factor can always be enrolled at `aal1`.
#### 5.7.3 Additional, removed and replaced factors
Native behaviour of GoTrue v2.197.0 (source read, V1-2.5 proves it):
- a second factor: `EnrollFactor` needs `aal2` once a verified factor exists
(403 `insufficient_aal`). A stolen password therefore cannot add a factor;
- removing a verified factor: `UnenrollFactor` needs `aal2` (422
`insufficient_aal`); an unverified factor can be removed from `aal1`. An
unverified factor gives no `aal2` until verified, so a dormant one left by
an earlier password holder is a path: factor checks list unverified
factors too, and V1-2.5 tests a dormant factor verified later (GoTrue
appears to delete unverified factors of the same type on verification;
not assumed). Assumption: TOTP is the only enabled factor type;
- removing the last factor: allowed from `aal2`. The account then returns to
the first-factor case, where an `aal1` session can enrol; a role holder in
that state keeps the role (residual, §13). A stolen `aal2`
session can therefore replace factors; this is the stolen-session residual
(§5.7.8), not a bypass from the password alone;
- GoTrue downgrades the sessions of a removed factor to `aal1`
(`DowngradeSessionsToAAL1`); its exact scope and the lifetime of
already-issued access tokens are recorded by V1-2.5, not assumed;
- freshness: tokens carry `amr` entries (method and time), so a recent-TOTP
check on the rare role RPCs is technically possible; adopting it is an
owner choice under O4, not part of this contract (§5.7.9).
If the sandbox shows `aal1` can enrol a second factor or remove a verified
one, V1-2.5 stops for an owner decision (as §5.6 already says).
#### 5.7.4 Password change and reset
A reset or change of password yields an `aal1` session. For an account with a
verified factor it cannot add or remove a factor (§5.7.3) and cannot change a
platform role (rule above). For an account without a verified factor, the
reset lets whoever holds the mailbox enrol a first factor: this is the
accepted first-factor case, harmless while no role or capability is held, and
the reason the owner role is assigned after enrolment (§5.7.1) and a
capability only to an account with a verified factor (§5.2). Email change and
identity linking stay as configured (`enable_manual_linking = false`); V1-2.5
adds a test that none of them raises a session to `aal2`.
#### 5.7.5 Operator path
The technical paths are the service key (RLS bypass and GoTrue admin) and a
privileged SQL session on the database container. They are not exceptions
usable from the application: the app never holds the service key; the worker
holds it and can call worker RPCs only; the role RPCs refuse `service_role`
(no `auth.uid()`, and `EXECUTE` revoked in M4). Each use is an owner-authorised
operation (§16.11), written down with its reason, and leaves an audit row
(`private.log_event`, null actor, operator in the detail) in the same
session.
#### 5.7.6 Actors
| Actor | Can change platform roles | Note |
|---|---|---|
| Ordinary account | no | no MFA requirement for ordinary use |
| Owner or admin at `aal1` | no | refused at check 2 whatever the role |
| Owner or admin at `aal2` | within `can_manage_platform_role` | admin: moderator and support only |
| Moderator or support at `aal2` | no | `aal2` is not sufficient |
| Worker | no | worker RPCs only (M6) |
| `service_role` key holder | not through the RPCs; yes through GoTrue admin and SQL | outside MFA, §5.7.5 and §13 |
| Operator SQL session | yes | audited by procedure, not by the database |
#### 5.7.7 Recovery
Complements the procedure of §5.6, which stands (operator only, out-of-band
identity, reason recorded, no RPC, no screen).
- Lost phone, account with other factors or another admin: the person uses
the second factor. A person whose only factor is lost follows §5.6 steps 1
to 4.
- Loss of every factor of the last owner: no application path exists, by
design. On a single-owner platform the operator is the owner acting at the
server, and the real control is server access, not a second person. Order:
1. the operator signs out every session, deletes every factor (verified or
not), bans the account and writes the audit row `auth.mfa_factor_reset`;
2. with the owner present, the operator sets a new password with the owner
and lifts the ban;
3. the owner signs in at once, enrols and verifies two factors;
4. the operator lists every factor of the account; any it does not
recognise sends the procedure back to step 1.
The owner role and the owner's capabilities stay in place: inert at
`aal1`, since `caller_has` and the role RPCs need `aal2`. Nothing is
re-granted. (For another account, §5.6 step 2 applies: the owner revokes
its capabilities first.)
- Return to a safe state: `aal2` proven and `aal1` refused again on that
account; the audit rows exist; no temporary bypass was left.
- Rehearsed on the sandbox in V1-2.9; the runbook goes to `IDENTITY.md`.
Secrets (QR codes, TOTP seeds) never go in Git, documents, logs or tickets.
Prevention: two factors on two independent devices for every owner.
#### 5.7.8 Sessions, tokens and stale rights
- Roles and capabilities are read from the tables at every call; the project
adds no claim to the JWT (no access-token hook), so a revoked right or role
is refused at the next call whatever the token says.
- The `aal` claim lives as long as the access token (`jwt_expiry`, 3 600 s
locally). For the role and rights RPCs only (rare, administrative), V1-2.1
adds a check that the session named by the token's `session_id` still
exists in `auth.sessions` at `aal2`. The columns and the claim are verified
on the pinned image first; if one is missing, V1-2.1 stops for an owner
decision (claim alone with the residual of up to one token lifetime after
a downgrade, or another check). Nothing here is assumed present.
- A stolen `aal2` session is a distinct risk this design does not remove:
it keeps working until it ends. Reducing it (freshness window, session
limit) is not part of V1-2.
- Concurrent role changes: unique active row per (user, role), lock on the
owner count (F18), one transaction per write.
#### 5.7.9 Reserves recorded, not decided
- `is_platform_admin` and the staff read policies (`profiles_select_staff`,
`audit_logs_select_admin`) do not read the assurance level. AR-1 B covers
changes of roles, not those reads; IDENTITY.md §4 foresaw folding `has_mfa`
into them. Left for an owner decision; V1-2.1 does not change them.
- Refused escalation attempts are not in the audit table (§5.7 order of
checks).
- A recent-TOTP freshness check on the role RPCs through `amr` (§5.7.3).
- The DEV hardening of `ENVIRONMENTS.md` §3.6 becomes a precondition of M4
on DEV (§5.7.2 step 0); the owner chooses its form.
---
@@ -1194,6 +1483,9 @@ test that fails on the current code before fixing it.
| F14 | `can_edit_sheet` gives every owner and admin full rights on official sheets, with no capability and no MFA | l.494-510 | confirmed by reading | M4 |
| F15 | the sheet policy reads `can_use_sheet`, whose `public` branch opens rows to every authenticated account, outside any catalogue right | l.513-526, l.1183-1190 | confirmed by reading | M4 (reader row predicate, §5.3) |
| F16 | asset rows have no immutability trigger: a privileged UPDATE of `sha256` changes the drawing of a published version, which its fingerprint does not cover | asset table; `fingerprint.rs` | confirmed by reading | M4 (asset-row trigger), M5 (publication compares sha256) |
| F17 | `service_role` keeps the default table privileges on `profile_roles` and `profiles` (the revokes name `public, anon, authenticated`), so a service-key session can write platform roles directly; `grant_platform_role` and `revoke_platform_role` are also granted to `service_role` without use | `20261008100200` l.247-248 and the revoke on `profile_roles`; `20261008100100` l.100 | confirmed by reading; effective grants checked on the sandbox in V1-2.1 | M4 (`REVOKE ALL` from `service_role` on both tables, `EXECUTE` revoked on the two RPCs) |
| F18 | `withdraw_platform_role` counts active owners without a lock: two concurrent revocations of two owners can both pass the guard and leave none | `20261008100200` l.143-181 (count l.165-169) | confirmed by reading | M4 (advisory lock in both RPCs and the private functions, count on active profiles, two-session tests) |
| F19 | `grant_platform_role` and `revoke_platform_role` do not check the assurance level (AR-1); a grant does not check that the target holds a verified factor; the private write functions trust a null actor from any session | `20261008100200` l.108-181, l.189-238 | confirmed by reading | M4 (`has_mfa()` as the second check, §5.7) |
### 12.2 Rust authority and SQL guards (O22)
@@ -1243,7 +1535,7 @@ Inventory of the current SQL checks and their fate:
| Id | Step | Content | Data on DEV |
|---|---|---|---|
| M4 `sheet_access_hardening` | V1-2.1 | `anatomical_catalog_grants` and its RPCs; `has_mfa`, `grant_active`, `caller_has`; `can_read_official_catalog` and the reader row predicate; `can_edit_sheet` redefined (F14); F1, F2, F3 (fork closed), F5, F6, F8, F9, F12 (every catalogue table), F13, F15; asset-row immutability trigger (F16); historical-sheet protection trigger (§14.2); reader column grants narrowed; personal creation and visibility changes refused | no row changes |
| M4 `sheet_access_hardening` | V1-2.1 | `anatomical_catalog_grants` and its RPCs; `has_mfa`, `grant_active`, `caller_has`; `can_read_official_catalog` and the reader row predicate; `can_edit_sheet` redefined (F14); AR-1 B on the platform role RPCs and the private write functions (F19, §5.7), F17, F18; F1, F2, F3 (fork closed), F5, F6, F8, F9, F12 (every catalogue table), F13, F15; asset-row immutability trigger (F16); historical-sheet protection trigger (§14.2); reader column grants narrowed; personal creation and visibility changes refused | no row changes |
| M5 `sheet_review_workflow` | V1-2.1 | version columns `revision`, `authors`, `source_version_id`, `published_review_id`, `document_fingerprint`; sheet column `working_copy_source_version_id`; revision, author and freeze triggers, the historical-sheet write rule; extended immutability guard (§6.8); rounds, comments, acknowledgments; `create_official_sheet`, `create_sheet_draft` adapted, `save_sheet_draft`, submit, withdraw, comment, acknowledge, decide, publish; `create_sheet_working_copy`; species RPCs; catalogue and version read RPCs; upload: `register_sheet_asset` extended, `confirm_sheet_asset_upload` with the storage check (F7), INSERT policy, bucket limit 2 MiB; mirror removal and status check (§12.2); superseded signatures dropped | new columns take their defaults on the equine v1 row (`revision` 0, empty author set); its document, status, number and assets unchanged; bucket limit lowered; no round, no grant |
| M6 `sheet_verification_jobs` | V1-2.3 | worker RPCs (claim, input, complete with the verdict structure check, fail), service role only | none |
@@ -1261,7 +1553,7 @@ outside the document.
| Threat | Control |
|---|---|
| Privilege escalation through the catalogue | default deny; rights in their own table, written by the owner only with `aal2`; `can_edit_sheet` redefined; no table privilege for API roles or the service role; security definer functions with `search_path = ''`, EXECUTE revoked from public and anon; superseded signatures dropped |
| Escalation through platform roles | open: AR-1 (§17.2) |
| Escalation through platform roles | AR-1 B: `aal2` on every grant and revocation of a platform role, checked in the RPCs before any role check and in the private write functions for a non-null actor; `service_role` has no table privilege and no `EXECUTE` on them; owner count race closed; first factor enrolled before the owner role is assigned (§5.7) |
| Takeover of an admin or reviewer account | `aal2` in RPCs and in the capability branches of policies; a verified factor required before a capability is granted; deactivation, expiry and revocation cut access at the next request; two factors for the owner |
| Recovery as a back door | no application path resets a factor; operator procedure with out-of-band identity check, capabilities revoked and sessions signed out first, an audit row (§5.6) |
| Self-approval | author set maintained by trigger and frozen at submission; no author acknowledges or approves; empty saves refused; content writes without an actor refused |
@@ -1285,11 +1577,20 @@ outside the document.
Residual risks:
- AR-1 until the owner decides (§17.2);
- the token's `aal` claim outlives a factor removal for up to one token
lifetime unless the session check of §5.7.8 is confirmed on the pinned
image; a stolen `aal2` session can also change platform roles and replace
factors (same residual as below);
- a role holder who removes the last factor from `aal2` keeps the role in
the first-factor case (§5.7.3);
- refused escalation attempts leave no audit row (the transaction raises);
- staff reads gated by `is_platform_admin` stay at `aal1` (§5.7.9);
- the service key is also a GoTrue admin credential: whoever holds it can
delete factors and sign in as anyone, so MFA does not protect against a
service-key holder; the key stays on the server, in the worker and in the
operator's environment;
operator's environment, except on DEV today, where it is the public demo
key and the JWT secret is the CLI default (`ENVIRONMENTS.md` §3.6), so
AR-1 B is forgeable on DEV until that hardening (§5.7.2 step 0);
- a capability does not prove a professional qualification: the owner's
verification note is a declaration until Step 9;
- the owner can create and habilitate a second account of its own; traced,
@@ -1475,6 +1776,45 @@ The final approval needs all 30 groups acknowledged, no `to_confirm` and no
- deprecated version deletion refused; D8 reads after archive and
deprecation; superseded signatures and mirror functions absent; upgrade
from the DEV state.
- AR-1 B (SQL, per §5.7; PGlite sets the `aal` claim through the JWT
settings, the `auth.sessions` check needs the real image):
- matrix: no role, owner, admin, moderator, support, each at `aal1` and
`aal2`, plus a deactivated account, on grant and on revoke, for each of
the four roles, on self and on another account; expected: `aal1` always
refused with `mfa_required`, `aal2` allowed only where
`can_manage_platform_role` allows (owner any role, admin moderator and
support), a moderator or support at `aal2` refused;
- regression of the chain: an owner at `aal1` calls the grant RPC directly
for an attacker-controlled account; refused, no `profile_roles` row, no
audit row, no capability grantable to it; same for revoking the real
owner and for a demotion;
- the private functions refused for a non-null actor at `aal1`, the null
actor path still working for the seed; `service_role` refused on the RPCs
and on both tables (F17); the last-owner guard under two concurrent
revocations (F18);
- no partial write on refusal; audit row on success with no secret in it;
- a role revoked then a call with the old token: refused at the next call;
- an account with no role, enrolling a factor, gains nothing; a first owner
initialised on a fresh stack in the order of §5.7.1;
- grant to a target without a verified factor refused; a null actor from
an API session and from `service_role` refused in the private functions;
`aal` claim absent, null or wrongly cased refused; mutual owner
revocation and revoke against grant in two sessions;
- fixtures that call the role RPCs (pgTAP 004 and others, the sandbox
`workspaces.sandbox.ts` seed owner at `aal1`) moved to `aal2` claims or to
the operator path; the PGlite shim gains stand-ins for `auth.sessions` and
`auth.mfa_factors` with columns checked against the pinned image;
- V1-1 regression: workspaces, clients, animals, relations, sign-in and a
practitioner without role unchanged; pgTAP 001 to 014 green with the
existing identity fixtures (adapted only where they assert the old
behaviour of the role RPCs).
- MFA on the sandbox stack (real GoTrue): first enrolment at `aal1`, verify,
`aal2`; second factor from `aal1` refused; removal of a verified factor
from `aal1` refused; removal from `aal2` and the session consequences;
password reset does not reach `aal2`; refresh keeps `aal2`; revoked session
and old token; recovery rehearsal of §5.7.7. Expectations are what the
pinned image really does; any gap with §5.7.3 is written up and decided by
the owner.
- Concurrency on the real stack, two sessions: two saves, save against
submit, publish against withdraw, two publishers, deprecate against publish,
completion against withdrawal.
@@ -1505,7 +1845,8 @@ The final approval needs all 30 groups acknowledged, no `to_confirm` and no
- Browser (`e2e:shell` on the sandbox, never DEV): journeys A0 to A8, S1, W1,
R1, P1 with two accounts (an editor and publisher, a reviewer) plus reader
assertions, TOTP codes generated by a test helper (Node `crypto` HMAC, no
dependency, V1-2.5); axe; 320 to 430 px for the catalogue, 375 px for admin
dependency, built in V1-2.1 for the real-GoTrue acceptance of §16.2 and
reused in V1-2.5); axe; 320 to 430 px for the catalogue, 375 px for admin
screens; FR and EN.
- Each step: `pnpm check:all`, fresh verifier against this contract, fresh
security audit.
@@ -1529,7 +1870,7 @@ V1-2.1; 5 core loading and validation → V1-2.2; the server verdict → V1-2.3;
categories, references → V1-2.6; 10 preview and validation → V1-2.7;
11 equine → V1-2.8; 12 acceptance and closure → V1-2.9.
Owner checkpoints: V1-2.0 (final review of this contract, AR-1), V1-2.3
Owner checkpoints: V1-2.0 (final review of this contract), V1-2.3
(server workflow and verdict), V1-2.4 (wireframes), V1-2.6, V1-2.7, V1-2.8
(with the business reviewer), V1-2.9 (closure, merge), and each DEV operation
of §16.11.
@@ -1546,8 +1887,8 @@ V1-2.7 on.
- Security prerequisites: none (documents only).
- Application changes: none. SQL changes: none.
- Tests: documentary checks only (cross-references, ids, tables).
- Acceptance: the owner's final review finds no contradiction; AR-1 decided;
ratification recorded.
- Acceptance: the owner's final review finds no contradiction; AR-1 decided
(B, BDR-083); ratification recorded.
- Rollback: revert the documentary commits.
- Checkpoint: owner, final review; V1-2.1 starts only on the owner's go.
@@ -1560,7 +1901,7 @@ V1-2.7 on.
the historical-sheet protection; the upload RPCs and policy; removal of the
SQL mirror and the status check; `packages/supabase/src/sheets.ts` with
closed failure kinds; the client gate deleted.
- Dependencies: V1-2.0 go, AR-1 decided.
- Dependencies: V1-2.0 go; AR-1 B decided (BDR-083).
- Security prerequisites: effective grants of the service role recorded on
the sandbox before F12's fix; census of private functions updated (pgTAP
001); every RPC requires `aal2` except the reader RPCs.
@@ -1578,6 +1919,12 @@ V1-2.7 on.
working copy created from the seeded historical sheet with its drawing; an
inactive species invisible to the animal path; types check clean; `pnpm
check:all` green.
- Acceptance, AR-1 B: the tests of §15 for §5.7 pass on the sandbox; a
stack upgraded from the DEV state with an owner holding two verified
factors keeps working at `aal2` and is refused at `aal1`; the same upgrade
with an owner holding none stays usable at `aal1` for everything but role
and rights management; the census lists no other write path to
`profile_roles`.
- Rollback: branch revert; sandbox rebuilt.
- Checkpoint: gates only (owner at V1-2.3).
@@ -1647,12 +1994,17 @@ V1-2.7 on.
- Security prerequisites: server rules of V1-2.1 in place; the UI is never
the boundary.
- Application changes: routes and screens of §11.1 except the editor and the
review screen; Mon compte, Sécurité; the TOTP test helper. SQL changes:
none.
review screen; Mon compte, Sécurité; the TOTP test helper (from V1-2.1). SQL changes:
none. Scope split: M4 (V1-2.1) holds every server control of AR-1 B; this
step holds the frontend journey only (enrolment, step-up, factor list, the
`mfa_required` message in FR and EN).
- Tests: unit; `e2e:shell` A0, A1, S1, P1; access per right; MFA required;
refresh keeps `aal2`; no `aal1` enrolment of a second factor; axe; 320 to
430 px for the catalogue, 375 px for admin screens; FR/EN.
- Acceptance: the owner with MFA grants rights and creates a sheet; an editor
- Acceptance: the owner enrols two factors from Mon compte, reaches `aal2`,
and only then manages platform roles and rights (an `aal1` attempt is
refused in the UI and in SQL); the owner with MFA grants rights and creates
a sheet; an editor
creates an inactive species that the animal forms do not show until the
owner activates it; a sheet published on the sandbox through SQL fixtures
is browsable by a catalogue reader and refused to a confirmed account
@@ -1748,8 +2100,12 @@ V1-2.7 on.
Each is an owner action, never run by Claude, after its own checkpoint:
1. applying M4 to M6, after a backup and a sandbox rehearsal from the DEV
state;
2. enrolling the owner's two TOTP factors and the first rights grants;
state; M4 is not applied before the preconditions of §5.7.2 (DEV hardening
of `ENVIRONMENTS.md` §3.6, seed owner password, role holders listed,
item 2's factors enrolled and verified), nor before the recovery of §5.7.7 is rehearsed; M4's checks take
effect when it is applied, with no flag that could be left open;
2. enrolling the owner's two TOTP factors (before item 1, §5.7.2) and the
first rights grants (after it);
3. starting the verification worker on DEV: a separate operating checkpoint
that first inventories the queues (the avatar upload waiting since
AVATAR-01, the due purge rows, quarantine leftovers), states the expected
@@ -1757,7 +2113,7 @@ Each is an owner action, never run by Claude, after its own checkpoint:
4. creating, activating or renaming a species;
5. creating the equine working copy and applying the O18 corrections;
6. any submission, approval, publication, deprecation or archive;
7. any factor reset (§5.6) and any `object_conflict` deletion (§9.3);
7. any factor reset (§5.6, §5.7.7) and any `object_conflict` deletion (§9.3);
8. rerunning the equine seed.
### 16.12 Deferred (not V1-2)
@@ -1777,7 +2133,7 @@ registry in any build that can be rolled back to; V1-2 does not resolve it.
---
## 17. Ratified decisions and open arbitration
## 17. Ratified decisions and arbitration
### 17.1 Ratified decisions O1 to O23
@@ -1789,7 +2145,7 @@ which would have accepted a browser verdict or deferred MFA, are rejected.
| O1 | A + D: three explicit capabilities (`sheet_editor`, `sheet_reviewer`, `sheet_publisher`), granted and revoked by the owner; editor and publisher need a platform role owner or admin; reviewers may be confirmed non-admin accounts; every grant audited, with a required reason, a verification note, a date and an optional expiry; e-mail confirmation is no proof of competence | §5.1, §5.2 |
| O2 | A: no author acknowledges or approves; author set frozen at submission and checked on the server; two people suffice; the publisher may be the approver; no silent override for a sole owner | §5.5 |
| O3 | B: no `to_confirm` and no `inherited` (absent statuses included) at approval or publication; the historical import is never a base; new elements start `to_confirm`; changes re-open coverage | §6.6, §6.7, §14 |
| O4 | A: `aal2` for every catalogue operation except reading the published catalogue; enrolment, step-up, sessions, loss and operator recovery documented; O4-D evaluated, not ratified, now the subject of AR-1 | §5.6, §17.2 |
| O4 | A: `aal2` for every catalogue operation except reading the published catalogue; enrolment, step-up, sessions, loss and operator recovery documented; O4-D evaluated, not ratified at the time; the arbitration AR-1 then extended `aal2` to every platform role change (§17.2) | §5.6, §5.7 |
| O5 | A: one approval closes a round; trace kept for a future quorum | §6.3 |
| O6 | A: scope disclaimer required, present in FR and EN and non-blank after trimming; equine texts untouched | §6.7 |
| O7 | A: new elements start `to_confirm`; nothing assigns `reviewed` or `inherited` automatically; draft state, element marking and business approval distinguished | §6.6 |
@@ -1810,33 +2166,30 @@ which would have accepted a browser verdict or deferred MFA, are rejected.
| O22 | Rust core the only authority on document rules; SQL keeps access, relations, states, revisions, locks, atomicity and the verdict's identity; the business mirror is removed in M5 without an unprotected window; one declared status check kept | §12.2 |
| O23 | Fail-closed verification: explicit states and outcomes, closed codes per category, idempotent retries under a limit, withdrawal always possible, audit, no partial publication | §6.5 |
### 17.2 Open arbitration AR-1: MFA on platform role grants (O4-D)
### 17.2 Arbitration AR-1: MFA on platform role changes (ratified, option B)
Finding (security challenge, §19.3, confirmed by reading
`20261008100200_platform_roles.sql` l.83-95 and l.189-238): an owner may grant
any platform role, owner included, and `grant_platform_role` does not check the
assurance level. From a stolen owner session at `aal1` (password only), an
attacker grants `owner` to an account of their own; that account enrols its
own TOTP factor, reaches `aal2`, grants itself every capability (§5.2) and a
reviewer capability to a third account, then may revoke the real owner. Every
`aal2` control of the catalogue is bypassed. O4-D (extending `aal2` to these
RPCs) was explicitly left unratified, so this contract cannot close the gap on
its own.
`20261008100200_platform_roles.sql`, the two RPCs): an owner could grant any
platform role, owner included, with no check of the assurance level. From a
stolen password-only owner session, an attacker grants a role to an account
of their own, which enrols its own factor, reaches `aal2` and grants itself
every catalogue capability; every `aal2` control of the catalogue was
bypassed. The chain is written out in §5.7.
Options for the owner:
Options put to the owner: A (`aal2` for `owner` and `admin` only), B (`aal2`
for every platform role), C (leave the RPCs and accept the residual).
- A (recommended): ratify O4-D at least for granting and revoking `owner` and
`admin`, the roles that make an account eligible for `sheet_editor` and
`sheet_publisher`: `has_mfa()` in `grant_platform_role` and
`revoke_platform_role` for those roles, in M4, with pgTAP cases. Cost: one
predicate; it changes a V1 identity contract and requires the owner's
session at `aal2` to manage those roles.
- B: ratify O4-D for every platform role. Same cost, broader.
- C: keep the RPCs unchanged and accept the residual explicitly. Then the
catalogue's MFA protects against a stolen editor or reviewer password, not
against a stolen owner password.
Decision (owner, 2026-10-28, BDR-083): **B**. Every grant, revocation or
change of a platform role requires a valid `aal2` session checked on the
server, with no permanent exception. It complements O4 and leaves D1 to D9
and O1 to O23 unchanged (O1, O2 and O4 are applied, not amended: a role gives
no capability, `aal2` replaces no permission, §5.2, §5.5, §5.7). Where it
lands: rule, inventory, bootstrap, recovery and session rules in §5.7; M4 in
§12.3 (F17, F18, F19); tests in §15; acceptance in §16.2, §16.6; the order of
DEV operations in §16.11.
Until the owner chooses, V1-2.1 does not start (§16.2).
Reserves left to the owner, not decided here: `aal2` on the staff read
policies (§5.7.9).
---
@@ -1855,7 +2208,7 @@ First revision (2026-10-27):
Consolidation (2026-10-28):
- this file: decisions ratified, contradictions resolved, steps restated,
second challenge round integrated, AR-1 open;
second challenge round integrated, AR-1 decided (B) and integrated in §5.7;
- `docs/development/IMPLEMENTATION_PLAN.md`: §0 roadmap row of V1-2, Steps 13
and 17 notes (editor scope moved into V1-2 for official sheets), Step 13's
stale `content_jsonb`;
@@ -1873,6 +2226,15 @@ Consolidation (2026-10-28):
- `.claude/tasks/TODO.md`, `.claude/memory/decisions.md` (BDR-082),
`.claude/memory/journal.md`.
AR-1 B integration (2026-10-28):
- this file: §5.7 (rule, scenario, inventory, bootstrap, factors, password,
recovery, sessions), F17 to F19, M4, tests, steps and DEV order, §17.2
closed;
- `.claude/memory/decisions.md` (BDR-083), `.claude/tasks/TODO.md`,
`.claude/memory/journal.md`;
- `docs/development/IDENTITY.md` §4: one pointer to §5.7, no other change.
Not touched, with stale `content_jsonb` mentions left for V1-2.9:
`BACKEND_ARCHITECTURE.md` l.110, `DOMAIN_MAP.md` l.119.
@@ -1926,3 +2288,12 @@ repository before being integrated; none changed a ratified decision.
| Business coherence | CONCERNS (3 major, no blocker, no arbitration) | Acknowledgments voided by the later marking, so every equine group acknowledged twice (protocol « editor marks, then reviewer acknowledges », acknowledgment refused on groups with unread elements, §6.3, §14.6); removed zone or view had no group to cover (removal groups); working copy slot blocked by an archived copy (only `active` copies count); anatomical decisions carried out by editors from reviewer comments (§14.5); grant expiry and renewal semantics (§5.2); owner missing from the species list (§5.4); round columns per state (§6.3); base from deprecated or pre-V1-2 versions (base only from a V1-2-published version, §6.2); deprecating the last published version (§6.9); DEV copy as the single source of truth, sandbox copies as rehearsals (§14.4); cleaning table misdescribed the reader (§9.2); `published_by` already exists (M5); O6 « structure » defined; active species names (owner renaming) |
| Security and data | FATAL (1 blocker as arbitration request, 6 major) | AR-1: an `aal1` owner session grants the owner role and bypasses every catalogue MFA control (§17.2, open); the defensive status check could not see omitted statuses (structural walk, absent = `inherited`, §12.2); service-role privileges and direct publication (F12 unconditional, guard requires an approved round with a passed verdict, asset-row immutability F16, sha256 compared at publication, GoTrue residual in §13); readers could list draft-only sheets through the `public` branch (reader row predicate, F15); historical sheet deletable or archivable (RPC refusals and trigger, §14.2); poisoned content-addressed key without exit (`object_conflict` runbook with an audited scoped deletion, earlier detection by `PreviewSheet`, §9.3); first-factor trust on first use (verified factor required before a capability, recovery revokes and signs out first, `aal1` enrolment test); minor: serialised size, worker body cap, base from legacy publications, verdict structure checked in SQL, predicates split by caller and stored actor, signed URLs (residual), cleaning fixpoint |
| Simplicity and feasibility | CONCERNS (5 major, no arbitration) | Editor step waited for the worker step (uploads moved to M5, editor independent of V1-2.3, verification passing moved to V1-2.7); worker changes understated (separate verification mode and container, per-command limits, avatar lease reused without backoff, sandbox worker a deliverable); no editor render path (`PreviewSheet` as the editor handle with document replacement and `view.canvas`); overlapping coverage rules (one rule on group digests, change list for display only, no earlier digests sent to the worker); status check blind to omitted statuses (fixed as above); minor: core step ordered verdict first; level 1 trimmed (no move to another parent, no dragged vertices, numeric crop); two cleaning classes; rules epoch replaced by release migrations withdrawing rounds; two failure outcomes with categorised codes; comments with a group key only; time budgets and named fixtures; TOTP test helper; `register_sheet_asset` extended instead of a new RPC; purge guard by bucket. Not taken: moving the working copy RPC to V1-2.8 (it stays in M5 so DEV receives one set of migrations) |
### 19.4 AR-1 B review (2026-10-28)
One fresh challenger (robustness and security lens) read the AR-1 B text
against the migrations, the seed, the PGlite shim and `ENVIRONMENTS.md`.
| Lens | Verdict | Findings and what changed |
|---|---|---|
| Security | FATAL (1 blocker, 3 major, 6 minor) | DEV signs tokens with the public default secret and exposes 54321 to the LAN, so `aal2` is forgeable there (precondition §5.7.2 step 0, §16.11, §13 corrected); a grant to an account without a factor reopened the chain (target factor required, check 5); seed owner with a committed password and re-seed after M4 (password change or new owner, reset reopens the preconditions, seed declared a local-fixture exception); recovery runbook unrunnable under a ban and silent on the owner's capabilities (reordered, capabilities inert at `aal1`); minor: unverified factors, null actor keyed on the argument, owner race beyond the count and on inactive profiles, `amr` freshness overclaim, session-check fallback made a stop, sign-out wording, API-log claim, harness gaps (shim, fixtures, TOTP helper moved to V1-2.1), cross-references. None changed a ratified decision |