Pouya/adr-sml
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: rule Q62 by stating the truth; strike the /med-arb/ gloss; re-stamp R18

Browse Source 
Pouya's four rulings of 2026-09-01, applied 2026-09-02.

Q62 — RULED "state the truth", not "remove the access". /legal/privacy/
now says two people can read the intake table, names their role, and adds
the two stronger facts the false sentence had crowded out: the handler
role holds PutItem only, and adr-sml-deploy is implicitDeny on all seven
read and write actions. Wording is subject to Pouya's read-through —
Q63(a), with a TODO(pouya) beside the copy.

The ruling named one sentence; a vocabulary sweep found the falsehood in
three places, and the audit then found two more. Five paragraphs now
answer "who can see it" and change together.

The tripwire stays permanently, per ruling, and grew from two
alternatives to five. Every alternative is one string that reached dist/.
Proven both ways against the pre-correction page rebuilt from bd282aa:
exit 1 with 5 matches at dist/legal/privacy/index.html:54,67,67,68,72;
exit 0 on the corrected page, self-test 12 patterns / 36 approved
strings.

/med-arb/ — the gloss is struck with no replacement, per ruling. The
strike left "the section above" pointing at the ADRIC rule set and "the
agreement" with no antecedent; both fixed. The bare designations line
sitting under ADRIC's quoted competence requirement is also struck, which
goes beyond the ruling and is flagged for Pouya.

R18 — re-stamped, two-tier: (a)(c)(d) re-verified against a source,
(b)(e)(f)(g) held on a cadence judgement. All seven hold, no shipped
sentence changed. R18's trigger had NO cutover checklist item and had
stamped five extracts of seven; both fixed. Candidate limb (h) flagged.

R10 — fired and unsatisfied; left open on instruction.

The evidence behind the new privacy sentence was weaker than the
sentence. Re-measured: 33 of 33 roles simulated (23 of 26 carried inline
policies nobody had read; the two CDK lookup roles can read the table),
four trust policies, the CloudFormation escalation path for all five
users, 0 federated providers, root recorded. Every read path terminates
at the same two people.

Two review rounds, 36 findings. 35 fixed, 1 declined. Five of round 2's
were defects in round 1's own fixes; stopped at two per D19.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Md3GndFqWPzK78xAoebsg5
This commit is contained in:
Pouya Lajevardi
2026-09-02 07:59:50 -04:00
co-authored by Claude Opus 5
parent bd282aa47d
commit 6aaf089b05
16 changed files with 1181 additions and 126 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/reference/intake-table-access-verification.md
+265 -6
View File
@@ -10,16 +10,56 @@ one claim on the site whose subject was entirely outside the repo.
Raised by `claims-auditor` in the D20 cutover audit, 2026-09-01, finding 8.
**Provenance.** Every figure below was read from AWS on **2026-09-01** with the
commands listed at the end, run read-only as `arn:aws:iam::327082975128:user/pouya`.
**Provenance.** Figures in the ORIGINAL sections were read from AWS on
**2026-09-01**; everything in the **2026-09-02 addendum** at the foot of this file
was read on 2026-09-02, and it supersedes the original role screen. Both were run
read-only as `arn:aws:iam::327082975128:user/pouya` with the commands listed at
the end.
No command in this file creates or changes anything. Re-run them rather than
trusting this file; it is dated for that reason.
---
## The claim being checked
## ✅ RULED AND APPLIED — 2026-09-02
`src/pages/legal/privacy.astro`, §Who can see it:
**Pouya ruled `state the truth`, not `remove the access`** (§9 Q62, ruled
2026-09-01, applied 2026-09-02). Option 2 below is the one taken; option 1 was
declined. `lars`'s membership of `admins` is **unchanged**.
**The shipped sentences**, so this file can be compared against the live page
rather than against a struck one:
> Two people can. The table sits in an Amazon Web Services account that also
> runs systems unrelated to this practice, and that account has two
> administrators — me, and one other person who administers it with me.
> Administrative access to the account carries the ability to read the table, so
> both of us can read what you send.
> No one outside those two people has been granted access to the table, and that
> is measured rather than assumed: every user and every role in the account was
> simulated against this table, and the only ones that come back able to read it
> lead to those same two people.
> The function that receives the form **can only add a record — it cannot read
> the table back.** And the credential that publishes this website has **no
> access to the table at all**, for reading or for writing.
**Wording is subject to Pouya's read-through approval** — §9 **Q63**, with a
`TODO(pouya)` beside the copy. The measurement is settled; the phrasing is not.
⚠️ **AND THE ENUMERATION BELOW WAS NOT ENOUGH TO SUPPORT THE SECOND OF THOSE
SENTENCES. See the addendum at the foot of this file**, which is what the page
actually rests on. Read it before citing the five-row table.
---
## The claim being checked — ⚠️ STRUCK 2026-09-01, QUOTED HERE AS THE DEFECT
This is **no longer on the page.** It is kept because it is the string
`check-claims.mjs`'s `sole-administrator-q62` pattern permanently bars, and a
tripwire whose target is not recorded anywhere becomes unmaintainable.
`src/pages/legal/privacy.astro`, §Who can see it, **as it stood at `bd282aa`**:
> The table is reachable by the function that writes to it and by one
> administrative account, which is mine — nobody else has access to the table.
@@ -75,9 +115,13 @@ access, and it belongs to a second administrator of a shared account.
`meshkini*`/`gitea*` users are evidence of that on the IAM surface, not just
in the S3 bucket listing §10 describes.
## What has to happen before `/legal/privacy/` goes public
## What had to happen before `/legal/privacy/` went public — ✅ RESOLVED BY OPTION 2
Tracked as `AGENTS.md` §9 **Q62**. It is one of two things and both are Pouya's:
Tracked as `AGENTS.md` §9 **Q62**, **closed 2026-09-02 on option 2.** Kept
unstruck because the reasoning is what makes the ruling re-readable, and because
option 1 remains live in one direction: if that access is ever actually removed,
the page and the tripwire both have to change, and `check-claims.mjs`'s `rule:`
line carries that instruction.
1. **Remove the access** — take `lars` out of `admins`, or replace that
membership with a policy that denies DynamoDB on this table — and then this
@@ -139,3 +183,218 @@ parameter-expansion bug (`$ACCT:user/` parses `:u` as a history modifier). The
loop above brace-quotes `${U}` for that reason, prints one line per principal so
a silently-skipped iteration is visible as a missing row, and suppresses nothing.
**Five rows, or the run did not happen.**
---
## ⚠️ ADDENDUM 2026-09-02 — THE ENUMERATION ABOVE WAS INCOMPLETE IN THREE WAYS, AND ITS CONCLUSION SURVIVES ANYWAY
**Why this addendum exists.** `/legal/privacy/` publishes a completeness claim
about who can read this table — the second of the three sentences quoted at the
top of this file, under **The shipped sentences**. Written as it stood, this file
did not support it.
*(This paragraph quoted a different sentence until 2026-09-02 — round 1's
**pre-fix** wording, *"every account and role in this infrastructure…"*, which
the audit then struck. So this file briefly quoted two different sentences as the
live one, in the two halves of the same document, which defeats the comparison
R14 exists for. Fixed by pointing at the block above rather than re-quoting it:
one copy of a fact, in one place. `adversarial-reviewer`, round 2.)* Under R14 the artefact is what a reviewer compares the
claim against, so a claim stronger than its artefact is unverifiable by
construction even when it happens to be true.
**What was missing.**
1. **`aws iam list-role-policies` was never run.** The command block above lists
only `list-attached-role-policies`, which returns *managed* policies. **23 of
the 26 non-service-linked roles carry inline policies**, and none of them had
been read. The screen for "broad policies" could not have seen an inline grant.
2. **No role was ever simulated against the table.** Access was inferred from
policy *names* (`AdministratorAccess`) rather than measured as a decision.
3. **The five users were simulated for reads only** — `GetItem`, `Query`, `Scan`.
So the page's *"the credential that publishes this website has no access to the
table at all"* covered three read actions and said "at all".
**What the measurement found, and it changes the role count.** Simulating all 26
non-service-linked roles across **seven** actions — `GetItem`, `Query`, `Scan`,
`BatchGetItem`, `PutItem`, `UpdateItem`, `DeleteItem`:
| role | decision on the table | trust |
|---|---|---|
| `cdk-hnb659fds-cfn-exec-role-…-ca-central-1` | **allowed on all 7** | `cloudformation.amazonaws.com` only |
| `cdk-hnb659fds-cfn-exec-role-…-us-east-1` | **allowed on all 7** | `cloudformation.amazonaws.com` only |
| `cdk-hnb659fds-lookup-role-…-ca-central-1` | **allowed on the 4 READS**, denied on writes | `arn:aws:iam::327082975128:root` |
| `cdk-hnb659fds-lookup-role-…-us-east-1` | **allowed on the 4 READS**, denied on writes | `arn:aws:iam::327082975128:root` |
| `adr-intake-lambda-role` | `PutItem` **only**; implicitDeny on the other six | `lambda.amazonaws.com` |
| the other 21 | implicitDeny on all 7 | — |
**So FOUR roles can read the table, not the two this file recorded.** The two
`lookup` roles were missed by exactly the gap above: their grant is the inline
`LookupRolePolicy`, and `list-attached-role-policies` returns nothing for them.
**Why the published sentence is nevertheless correct.** The question is not how
many roles exist but which *people* they lead back to.
- The two **cfn-exec** roles trust `cloudformation.amazonaws.com` and nothing
else. No human can assume them; they are reachable only by deploying a
CloudFormation/CDK stack, which requires a principal who can deploy one.
- The two **lookup** roles trust the account root, which delegates the decision
to the caller's own identity policy. Simulated for `sts:AssumeRole` against
both role ARNs, for all five users, reading `ResourceSpecificResults` (ten
per-resource decisions, not five — `EvaluationResults` is one entry per
**action**, and an earlier pass here asserted the wrong count):
| principal | assume `lookup-…-ca-central-1` | assume `lookup-…-us-east-1` |
|---|---|---|
| `user/pouya` | **allowed** | **allowed** |
| `user/lars` | **allowed** | **allowed** |
| `user/adr-sml-deploy` | implicitDeny | implicitDeny |
| `user/gitea-deploy-meshkinilaw` | implicitDeny | implicitDeny |
| `user/meshkini-backend-deploy` | implicitDeny | implicitDeny |
- The **CloudFormation escalation path this file named and left untested** —
*"a real path to the table for anyone who can deploy a CDK stack"* — is now
measured. Simulated for `cloudformation:CreateStack`, `UpdateStack`,
`CreateChangeSet`, `ExecuteChangeSet`, `iam:PassRole` and `sts:AssumeRole`,
five users × six actions = **30 decisions**, count asserted:
| principal | the CDK / CloudFormation path |
|---|---|
| `user/pouya` | **allowed** on all six |
| `user/lars` | **allowed** on all six |
| `user/adr-sml-deploy` | implicitDeny on all six |
| `user/gitea-deploy-meshkinilaw` | implicitDeny on all six |
| `user/meshkini-backend-deploy` | implicitDeny on all six |
This is the finding that mattered most, because `meshkini-backend-deploy` is by
its name another project's backend-deploy credential and
`gitea-deploy-meshkinilaw` is held on a **jointly administered** Gitea instance
(Q23). Either one, had it been able to drive CloudFormation, would have read
the table without appearing in the five-row table above — and *"Two people
can"* would have been wrong. Neither can.
- **And the seven roles every sweep here had excluded BY CONSTRUCTION are now
measured too.** Every role loop in this file filters `grep -v
'^AWSServiceRole'`, and the assertion was written as *"twenty-six role rows"* —
so a service-linked role was outside the claim rather than inside it, which
matters because a service-linked role for a backup or migration service can
read table contents. All **7** (`APIGateway`, `CloudFrontLogger`,
`InternetMonitor`, `RDS`, `ResourceExplorer`, `Support`, `TrustedAdvisor`) are
**implicitDeny on all seven actions** — 49 decisions, count asserted
`[verified 2026-09-02]`. **So the enumeration is 33 of 33 roles, not 26 of
33**, and the page's *"every user and every role in the account"* is now
literally true. Raised by `adversarial-reviewer`, round 2.
- **There is no federated identity surface at all**: `list-saml-providers` **0**,
`list-open-id-connect-providers` **0**, `sso-admin list-instances` **0**
`[verified 2026-09-02]`. So "every user and every role" is not leaving out a
federated principal, because there is none to leave out.
Every read path therefore terminates at `pouya` or `lars`. **The count of people
is two, and it is now the result of an enumeration rather than of a policy name.**
**The account root user, recorded because an enumeration that quietly omits it is
not an enumeration.** Root is not an IAM user and does not appear in
`list-users`, so it cannot be simulated and no policy constrains it — root can
always read the table. Two facts bound it: `get-account-summary` reports
`AccountAccessKeysPresent: 0`, so **there is no programmatic root credential**,
and `AccountMFAEnabled: 1`. Root access therefore requires the root password and
its MFA device. **The page does not mention root and should not**: an account
owner's own credential is inherent to every cloud account and is not a third
party who has been granted access — which is why the shipped sentence is scoped
to *"every user and every role"* and to who has been *granted* access, rather
than to a bare "nobody else can". Who holds the root credentials is not
established in this repository; it is not a page-blocking fact, and it is noted
here rather than guessed at.
**Two page sentences are now supported that were not before.**
- *"The function that receives the form can only add a record and cannot read the
table back"* — `adr-intake-lambda-role` returns `allowed` for `PutItem` and
`implicitDeny` for `GetItem`, `Query`, `Scan`, `BatchGetItem`, `UpdateItem` and
`DeleteItem`. Previously this rested on reading the policy document; it is now
the simulator's decision.
- *"the credential that publishes this website has no access to the table at
all"* — `adr-sml-deploy` is `implicitDeny` on all **seven**, so "at all" now
covers writes and deletes as well as reads.
**And it corroborates §10 from the IAM surface.** Of the 26 non-service-linked
roles, **9 belong to CDK bootstrap** and **14 to four unrelated production
systems** in the same account — which is what `/legal/privacy/` now tells a
reader in as many words. *(Their role names were listed here until 2026-09-02 and
are not any more: this is a committed file, they are another project's IAM
surface, and the count carries the whole of the argument. `adversarial-reviewer`,
round 2.)*
⚠️ **THE INSTRUMENT FAILED FIRST, UNIFORMLY, AND IN THE DIRECTION THAT READS AS
CLEAN.** The role sweep was first run as `--action-names $ACTS` with the seven
actions in a shell variable. **zsh does not word-split parameter expansions**, so
`simulate-principal-policy` received **one** action name — the whole string — and
answered it: `implicitDeny` for 22 roles, and `allowed` for the four with a `*`
grant, because `*` matches a bogus action too. Twenty-two clean rows and a
plausible four. The tell was the shape of the output, not the verdict: one
decision per role where there should have been seven. **The fix is the assertion,
not the memory** — the loop now counts `EvaluationResults` per call and refuses a
row that does not carry exactly seven, and the users' assume check counts
`ResourceSpecificResults` and refuses a row that does not carry exactly two.
`CLAUDE.md` records this class five times over; this is the sixth, and it is the
"uniformly good" half.
### Commands — the ones this addendum rests on
Read-only, run as `user/pouya` in `ca-central-1`. No stderr suppressed, exit
status read on every call, and note the **literal** action lists: they are not in
a variable, which is the whole point above.
```bash
# 1. Inline policies — the command the original block never ran.
aws iam list-roles --query 'Roles[].RoleName' --output text \
| tr '\t' '\n' | grep -v '^AWSServiceRole' \
| while IFS= read -r R; do
aws iam list-role-policies --role-name "$R" --query 'PolicyNames' --output text
done
# 2. Every non-service-linked role, seven actions, decision asserted per row.
# The count check is what makes a broken call loud instead of clean.
aws iam simulate-principal-policy \
--policy-source-arn "arn:aws:iam::327082975128:role/<ROLE>" \
--action-names dynamodb:GetItem dynamodb:Query dynamodb:Scan \
dynamodb:BatchGetItem dynamodb:PutItem dynamodb:UpdateItem \
dynamodb:DeleteItem \
--resource-arns "arn:aws:dynamodb:ca-central-1:327082975128:table/adr-intake-submissions" \
--query 'length(EvaluationResults)' --output text # must print 7
# 3. Trust policies of the four roles that can read.
aws iam get-role --role-name <ROLE> --query 'Role.AssumeRolePolicyDocument'
# 4. Who can assume the two lookup roles — per RESOURCE, not per action.
aws iam simulate-principal-policy \
--policy-source-arn "arn:aws:iam::327082975128:user/<USER>" \
--action-names sts:AssumeRole \
--resource-arns "arn:aws:iam::327082975128:role/cdk-hnb659fds-lookup-role-327082975128-ca-central-1" \
"arn:aws:iam::327082975128:role/cdk-hnb659fds-lookup-role-327082975128-us-east-1" \
--query 'EvaluationResults[].ResourceSpecificResults[].{R:EvalResourceName,D:EvalResourceDecision}' \
--output text # must print 2 rows
```
```bash
# 5. The CloudFormation / CDK escalation path, per user. Six actions, literal.
aws iam simulate-principal-policy \
--policy-source-arn "arn:aws:iam::327082975128:user/<USER>" \
--action-names cloudformation:CreateStack cloudformation:UpdateStack \
cloudformation:CreateChangeSet cloudformation:ExecuteChangeSet \
iam:PassRole sts:AssumeRole \
--query 'EvaluationResults[].{A:EvalActionName,D:EvalDecision}' --output text
# 6. Federated identity surfaces, and root.
aws iam list-saml-providers --query 'length(SAMLProviderList)'
aws iam list-open-id-connect-providers --query 'length(OpenIDConnectProviderList)'
aws sso-admin list-instances --query 'length(Instances)'
aws iam get-account-summary \
--query 'SummaryMap.{AccessKeysPresentRoot:AccountAccessKeysPresent,MFA:AccountMFAEnabled}'
```
**Five user rows and twenty-six role rows, or the run did not happen.** And for
every simulation: **seven decisions per role call, six per CDK-path call, two
per-resource decisions per assume call** — the counts are the assertion, because
a call that silently received one bogus action name answers `implicitDeny` and
reads exactly like a clean row.