# Open points — pilot testing, 2026-08-07

Raised while running the first real βεβαιώσεις against `test.gsis.gr`. None of them block further
pilot testing; all of them should be settled before production.

**Updated 2026-08-07:** §1 and §2 are resolved in code. §3 is now a question to send, not
research to do. §4 is partly proven. **§5 is new and is the most serious item here** — a second
pilot session showed one ΧΚ carrying several βεβαιώσεις, which the close-day matching does not
handle correctly. §6 unchanged.

**Updated 2026-10-19:** §5 is fixed in code — the close-day matching now claims each
`tofeilisRecord` by `lineNo` and only falls back to ΑΦΜ when that is unambiguous. The only part
of §5 still open is how ΚΕΔ allocates γραμμές to συνυπόχρεοι.

**Updated 2026-10-08:** §7 is new — `kedeDeleteService` (M4) rejects a real Μείωση/Διαγραφή
attempt with an undocumented business error, and it looks structural rather than a request bug.

---

## 1. `doy` accepts a meaningless value — RESOLVED

**Observed:** a βεβαίωση with `doy = 1` was accepted by the pilot.

**Why it passed:** `doyFlag = 0`, and ΚΕΔ only consults `doy` when `doyFlag = 1` («βεβαίωση στη
ΔΟΥ ακινήτου», spec σελ. 5). So the field was carried and ignored.

**What the spec says.** `KEDE_Progress_Service_v1.03_MySQL.xlsx` defines it as
«ΔΟΥ καθ' υπέρβαση αρμοδιότητας (προαιρετικό)», **`CHAR(4)`**. So the four-digit rule is ΑΑΔΕ's
own, not our inference, and the migration already matched it (`char('doy', 4)`) — only the
validation was loose (`\d{1,4}`).

**Is a ΔΟΥ list needed?** No, and we could not build one anyway: **no document we hold contains a
list of ΔΟΥ codes.** It would have to come from ΑΑΔΕ. It only matters if `doyFlag = 1` is ever in
scope, and the ΔΟΥ of an ordinary βεβαίωση is not sent at all — ΚΕΔ derives it from the debtor's
ΑΦΜ and returns it as `result_doy` (`9776`).

**Decision taken (2026-08-07): `doyFlag = 1` is out of scope and refused outright.** The spec ties
it to Δ/νση Κρατικών Εγγυήσεων. Supporting it half-way — carrying a ΔΟΥ we cannot validate and
relying on a second field to neutralise it — is worse than not supporting it.

Implemented:

- `SubmitBeveosiRules`: `doy` must be exactly 4 digits; `doyFlag` is `in:0`, with a Greek message
  saying βεβαίωση σε ΔΟΥ ακινήτου is not supported.
- `BeveosiPayload::attributes()`: `doy` is forced to null unless `doy_flag === 1`, so a stale
  value cannot survive on the row waiting for someone to raise the flag. One line to re-enable.
- The ΔΟΥ input is gone from the submit form (`frontend/src/lib/fields.ts`); the read-only
  displays stay, because historical rows and `result_doy` still need showing.
- `KedeRequestBuilder` already omits null scalars, so `<doy>` is absent rather than empty —
  asserted in `KedeRequestBuilderTest`.

**Reopen if** Δημόκριτος says ΔΟΥ ακινήτου is in scope after all. Then we need the ΔΟΥ code list
from ΑΑΔΕ, `doyFlag` back to `in:0,1`, `doy` required when the flag is 1, and the form field back.

---

## 2. A new row per failed attempt — RESOLVED

**Question raised:** why does a failed submission leave a βεβαίωση row, and why should the table
fill up with them?

**Current behaviour**

| Outcome | Row? |
|---|---|
| Validation error (ours, before sending) | **no row** — nothing is created |
| Duplicate refused by the lock | **no row** — the draft we created is deleted again |
| ΚΕΔ rejects it (`failed`) | **row kept**, unlocked, resendable, keeps its `audit_protocol` |
| No answer (`needs_reconciliation`) | **row kept**, locked |

So only calls that actually reached ΚΕΔ leave a row. The reasoning:

- `needs_reconciliation` **must** persist — a debt may exist at the ΔΟΥ and the row is the only
  record that we have to go and check. This one is not negotiable (Παράρτημα Ζ).
- `failed` persists so the correction keeps the **same `audit_protocol`**. That id is what lets
  ΚΕΔ recognise a resend as the same βεβαίωση rather than a second one. It is also the audit
  trail: `ked_call_log` rows point at it.

**The legitimate complaint was that nothing enforced that.** A `failed` row holds no duplicate
guard, so five attempts made five rows and five `audit_protocol`s — defeating the point of
keeping the first one.

**Fixed (2026-08-07).** `SubmitBeveosiCommand::findRejected()` matches an existing `failed` row on
the natural key (ΑΦΜ, ΚΑΕ, έτος, αρ. απόφασης) and reuses it: same row, same `audit_protocol`,
repeated failures update rather than breed. Deliberately `failed` only — every other state that
reached ΚΕΔ still goes through `CertificationLockedException`, which is the anti-duplicate defence
of Παράρτημα Ζ and must not be bypassed. Drafts are excluded too; they are reached by `id` or
`auditProtocol`, and matching them here would let a submission quietly adopt a half-finished form
the caller never named.

**Found while fixing it, and also fixed:** `submit_beveosi` validated the merged params but then
submitted the row **as stored** — so a correction sent with the resend was checked and thrown
away, and the resend carried the very payload ΚΕΔ had rejected. It now writes the validated
payload to the row (parent attributes and the debtor/δόση/ΚΑΕ children) before sending, for any
unlocked row. `paramsFor()` also carries `discountPcg`/`discountDate` now, so resubmitting without
them no longer wipes them.

**Decided:** `failed` rows are **shown in the default list**. They are work in progress the user
has to correct; `ListBeveoseisCommand` applies no default status filter, and the status filter is
there for narrowing down.

---

## 3. `txtCode` / ΚΑΕ — a question to send, not research to do

**Settled:** the values are **assigned per φορέας by ΑΑΔΕ**, and the pilot is exercised with the
production values. Παράρτημα Ε, transcribed in `parartima-z-notes.md` §4:

> «Οι φορείς κάνουν δοκιμές στο test με τις τιμές Κωδικός Εξωτερικού Οργανισμού (extOrg), KAE και
> Είδος Φόρου (txtCode) **που τους έχουν δοθεί από την ΑΑΔΕ** για την παραγωγική λειτουργία.»

**Not discoverable, confirmed:** no `kedeProgressService` operation returns the granted set, and no
WSDL/XSD is committed. The spec redacts the value (`<txtCode>****</txtCode>`), the xlsx gives only
the type. We use `4102` because the Postman collection ΑΑΔΕ supplied uses it, and the pilot
accepted it. So this can only be answered by asking.

Already asked once and unanswered — `docs/source_docs/meetings/Simeioseis_Meeting_KED.docx`:
«Είδος φόρου … Ποιες τιμές ισχύουν; Διαφορά από τον ΚΑΕ;»

**Ask Δημόκριτος, precisely:** which (ΚΑΕ, `txtCode`) combinations has ΑΑΔΕ granted the φορέας, and
are they identical in production?

- **One pair** → the current design is right. `txtCode` stays invisible config, ΚΑΕ stays a
  per-βεβαίωση choice from a permitted set. Nothing to build.
- **Several** → seed `lookup_tax_types` and add a dropdown. ~1h: the table and the `get_lookups`
  entry already exist.

Related and still unexplained: `extOrg` and the ΚΑΕ are **both `82838`**. Nobody has said whether
that is a coincidence or a rule.

---

## 4. The `scheduler` container — starts and fires; the long-run behaviour is still unproven

`compose.yaml` gained a fourth service in phase 3 (`scheduler`, running `artisan schedule:work`,
which fires `ked:reconcile` every `KED_RECONCILE_INTERVAL` minutes).

**Exercised for the first time on 2026-08-07.** What is now proven:

- it starts with `sail up -d` and ticks once a minute;
- `schedule:list` shows `*/15 * * * * php artisan ked:reconcile`;
- it actually fired — `2026-08-07 09:30:00 Running ['artisan' ked:reconcile] in background … DONE`
  — and a hand-run sweep returns cleanly (`dates=0 matched=0 certified=0 unresolved=0`).

**Found:** the service overrides `command:`, which bypasses Sail's entrypoint, so the process runs
as **root** rather than `WWWUSER`. Harmless under the bind mount, where the host owns the files —
but it is exactly why a native deployment must run the scheduler as the app's user.

**Still unproven, and it needs a pilot βεβαίωση to sit overnight:**

- whether `withoutOverlapping()` holds when a sweep runs long;
- whether a βεβαίωση left at `certified_pending_closeday` overnight is picked up **without**
  anyone opening the status page. That is the real test — with `KED_DRIVER=fake` the ΑΤΒ appears
  on the next poll, but a real ΑΤΒ only exists after the ΔΟΥ closes the day.

**Production note:** on-prem there is no compose. The scheduler must run under supervisor
(`schedule:work`) or as a one-line cron (`* * * * * php artisan schedule:run`). Without it,
βεβαιώσεις never leave `certified_pending_closeday`.

Commands: `docs/runbook.md` §5.

---

## 5. One ΧΚ carries several βεβαιώσεις — and the close-day matching does not handle it

**Observed 2026-08-07**, second pilot session. Two βεβαιώσεις for the same ΑΦΜ (`137569280`,
αποφάσεις 1/2026 and 2/2026) both landed in **ΧΚ 3** and both carry **ΑΤΒ 861690**. That much is
correct: the ΑΤΒ belongs to the ΧΚ, not to the βεβαίωση. What distinguishes them is the γραμμή.

`retrieveCloseDay` (callSequenceId 10071002450) returned the whole ΧΚ — `totalLines=3`,
`outBudget=1800`, three `tofeilisRecord`s all for ΑΦΜ `137569280`:

| lineNo | payId | belongs to |
|---|---|---|
| 1 | `…686169000001` | not in our database |
| 2 | `…686169000002` | cert #1, απόφαση 1/2026 |
| 3 | `…686169000003` | cert #4, απόφαση 2/2026 |

### This disproves a recorded assumption

`docs/data-model.md` «Open questions» §1 says, flagged as second-hand and unconfirmed:

> Δημόκριτος report that ΚΕΔ opens **one ΧΚ per ΑΦΜ**, i.e. no batching…

**Not true on the pilot.** One ΧΚ carried three γραμμές for one ΑΦΜ from at least two separate
βεβαιώσεις. Batching is the normal case, not a hypothetical. The `mc_no` + `line_no` + per-debtor
`pay_id` shape already represents it, so no redesign is needed — but the matching code does not
actually use it. Update that open question in `data-model.md` when this is fixed.

### The defect

`docs/data-model.md` specifies the sweep as:

> Match `tofeilisRecord` rows to `beveosi_debtors` by **`lineNo` and trimmed `afm`**

`ReconciliationService.php:204-213` matches on **`afm` alone**, and the `update()` runs inside the
loop, so **the last matching line wins**. The comment above it — «One tofeilisRecord per ΑΦΜ» — is
wrong: it is one record per **γραμμή**, and a ΧΚ can repeat an ΑΦΜ across γραμμές.

Both rows above hold the right `pay_id` **by accident**: each happened to be the last line in the
ΧΚ at the moment it reconciled. Cert #1 was transiently written line 1's `payId` — another
βεβαίωση's — before being overwritten by its own.

**How it goes wrong.** If cert #1 reconciled now, the loop would walk lines 1, 2, 3 (all ΑΦΜ
`137569280`) and leave it holding **line 3's** `payId`, which belongs to cert #4. A debtor would
then be told to pay against another βεβαίωση's Ταυτότητα Οφειλής. Reachable through a delayed
reconciliation, a re-run sweep, or a βεβαίωση whose συνυπόχρεοι repeat across ΧΚ γραμμές.

**Not fixed yet** — recorded here by decision on 2026-08-07.

**Fixed 2026-10-19.** `ReconciliationService` no longer matches by ΑΦΜ. What it does now, per
ΧΚ (one `retrieveCloseDay` per ΧΚ rather than per βεβαίωση — the old code also spent a call per
βεβαίωση on a ΧΚ shared by several):

1. Every βεβαίωση of ours on that ΧΚ participates — already-`certified` ones included, purely as
   claim-holders.
2. **Authoritative:** a `tofeilisRecord` goes to the debtor row whose stored `line_no` equals its
   `lineNo` — the primary's was handed out by `retrieveKedeAmount` at submit time, so a late
   reconciliation can never take a settled line's ΤΟ.
3. **Conservative fallback:** a leftover record goes to a debtor with neither `line_no` nor
   `pay_id` (a συνυπόχρεος, today) only when its ΑΦΜ appears **exactly once in the whole
   tofeilisList** and exactly one debtor is waiting for it. Anything ambiguous stays empty, with
   a `Log::warning` naming the ΧΚ and ΑΦΜ — the same philosophy as the sweep's «absence is never
   proof»: this service never guesses a real person's ΤΟ.

Regression tests in `CertificationServiceTest` reproduce this section's exact shape (two
βεβαιώσεις, same ΑΦΜ, one ΧΚ, fake-driver sentinel ΑΦΜ `888888888`) and fail against the old
code — verified: the old matcher left the first βεβαίωση holding the second's ΤΟ
(`…000000002`). The 2026-09-18 evidence run was checked and is clean: ten distinct ΑΦΜ; the only
repeated ΑΦΜ (066111650) sits in a different ΧΚ, and matching is per-ΧΚ. Rows written before the
fix hold correct data — the 08-07 write landed right by ordering luck, and nothing since has
repeated the shape.

**Live-validated 2026-10-26 (pilot).** Two βεβαιώσεις, same ΑΦΜ `137569280` (αποφάσεις 811/2026
and 812/2026, certs 27/28), deliberately submitted three minutes apart — and the ΔΟΥ batched
them into **ΧΚ 21** on its own, the 08-07 shape reproduced on demand. The scheduler's 11:30
sweep made **one** `retrieveCloseDay` (`ked_call_log` 596, under the lead's audit protocol
402/26-10-2026 — the old code would have spent two calls) and certified both with the same ΑΤΒ
`862011`, each keeping its own line's ΤΟ: cert 27 γραμμή 1 → `…686201100001`, cert 28 γραμμή 2
→ `…6568620110002`. The old matcher would have left cert 27 holding cert 28's ΤΟ.

**Still open — how ΚΕΔ allocates γραμμές to συνυπόχρεοι.** `retrieveKedeAmount` returns a single
`lineNo` for the βεβαίωση, and nothing we hold says whether the συνυπόχρεοι occupy the following
consecutive γραμμές or are allocated some other way. Only position 0 ever gets a `line_no` at
submit. Two ways to settle it: a live pilot βεβαίωση with a συνυπόχρεο, read `tofeilisList` at
close-day (costs quota, and the record stays — see §7 on deletion), or ask ΑΑΔΕ/Δημόκριτος.
Until then the fallback above handles only the unambiguous case, and an ambiguous συνυπόχρεος
shows an empty ΤΟ in the reply — visible, honest, and correctable the moment the rule is known.

---

## 6. Smaller items

- **Draft validation vs storage mismatch on `sunolo`** — `SubmitBeveosiRules` marks `sunolo`
  `sometimes` for drafts, but `BeveosiPayload::debtors()` stores it as the primary debtor's
  `amount`, which the schema requires — so a draft without `sunolo` surfaces as a generic
  status-3 «Απροσδόκητο σφάλμα» instead of a status-4 `extraData.fields` map (observed
  2026-09-03 while typing the contract). The rules should just require `sunolo` for drafts
  too; `message-contract.md` documents today's behaviour (required in practice) until then.
- **ΚΕΔ daily call quota** — `GET_AUTHORIZATION_MAX_OPEN_CALLS_EXCEEDED` exists, so a limit
  exists, but nobody has told us the number. It sets the reconciliation sweep interval
  (`KED_RECONCILE_INTERVAL`, currently a 15-minute guess).
- **How slow does ΚΕΔ get under load?** `KED_TIMEOUT` is 60s on single-call evidence. Too low
  manufactures locked `needs_reconciliation` rows out of calls that would have succeeded.
- **Is (ΑΦΜ, ΚΑΕ, έτος, αρ. απόφασης) genuinely unique?** It is the duplicate guard, and since §2
  it is also what makes a resend find its rejected row. If one απόφαση can legitimately produce
  several βεβαιώσεις for the same debtor and ΚΑΕ, the key needs a fifth component — a cheap change
  now, a painful one after go-live.
- **`comments` content** — the spec requires the έτος and αριθμός of the απόφαση. ΚΕΔ did not
  object to «Τίποτα» in testing, so it is not enforced there; it is what the ΔΟΥ reads to identify
  the debt, so we should probably enforce it ourselves.

---

## 7. `deleteKedeAmount` rejects a real Μείωση — a question to send, not research to do

Observed live, 2026-10-08, `ked_call_log` ids 227/228 (`callSequenceId` 10089502235 /
10089496314), against a real certified debt (ΑΦΜ `137569280`, ΑΤΒ `861859`, ΧΚ `7`/γραμμή `17`,
ΚΑΕ `82838`, amount `100.00`):

> **errorCode** `WS_PROCEDURE_ERROR`
> **errorDescr** «Δεν συμπεριλαμβάνεται ΚΑΕ Εντός Προϋπολ/μού και ο Τύπος Εκκαθάρισης θα πρέπει
> να είναι Διαγραφής Υπολοίπου» — *"No 'within-budget' ΚΑΕ is included, and the settlement type
> should be 'write-off of the remaining balance'."*

**Ruled out first:** this is not a request-shape bug. The envelope sent matches the spec's
documented `deleteKedeAmountInputRecord` field-for-field (confirmed by re-reading
`Kede_Delete_Amount_Service_K___v_0.99.pdf` §2.1 end to end) — every field present, correctly
named, correctly ordered. The error text itself does not appear anywhere in the spec's own
Πίνακας 2 (deleteKedeAmount error table) — it is undocumented, same category as the `doyFlag`
and `approvalDate` landmines, except those were at least nameable from a field we controlled.
This one names two concepts — «ΚΑΕ Εντός Προϋπολογισμού» and «Τύπος Εκκαθάρισης» — that **do not
exist as fields anywhere in this spec's request or response.**

**The likely cause, cross-referencing §5 above:** that section's live `retrieveCloseDay` sample
shows this exact kind of certification coming back `outBudget=1800`, `inBudget` unmentioned —
i.e. **εκτός προϋπολογισμού** (out-of-budget). `CertificationService::buildInput()` hardcodes
`inMctCod=1`/`inSctCod=98` (Είδος Χρηματικού Καταλόγου / Τύπος Πηγής) on every certification per
the progress-service spec — a value we do not choose per-βεβαίωση. If that fixed combination is
what puts every Δημόκριτος certification on the "εκτός προϋπολογισμού" side at the ΔΟΥ, then
`deleteKedeAmount`'s apparent requirement for an "εντός προϋπολογισμού" ΚΑΕ may mean **Μείωση/
Διαγραφή via this method cannot succeed for anything we certify, structurally** — not a bug in
our request, a property of the Φορέας' own ΚΑΕ/Πηγή setup. This is a hypothesis, not a finding;
it is not confirmed anywhere, and nothing in this spec documents an "εντός/εκτός προϋπολογισμού"
field or a "Τύπος Εκκαθάρισης" field we could even test the hypothesis against.

**Not discoverable by us, confirmed:** no field exists to set either concept; no operation
returns the granted ΚΑΕ-to-budget-category mapping. Same category as §3's `txtCode`/ΚΑΕ
question — this can only be answered by asking ΑΑΔΕ/ΓΓΠΣΔΔ (or whoever holds the ΑΑΔΕ pilot
relationship) directly.

**Ask, precisely:** does `deleteKedeAmount` require the targeted ΚΑΕ to be classified «εντός
προϋπολογισμού» at the ΔΟΥ, and if Δημόκριτος' certifications are always «εκτός προϋπολογισμού»
(per the fixed `inMctCod`/`inSctCod`), is Μείωση/Διαγραφή via this service available to us at
all — and if not, what IS the correct mechanism for correcting an over-certified debt of ours?

**Until answered:** do not keep retrying `deleteKedeAmount` against the live pilot by varying
fields — the rejection is server-side business logic, not something any request-field
combination we control has been shown to fix, and each attempt spends real ΚΕΔ quota. Use
`KED_DRIVER=fake` for continued UI/workflow testing of the M4 code path in the meantime — the
fake driver is unaffected and already covers the full submit → finalize flow.

### §7 update — 2026-11-03: `deleteKedeProgress` called live, works end to end

First live call of the delete service's read operation (`ked_call_log` id 791, our protocol
`549/03-11-2026`, **ΚΕΔ `callSequenceId` 10091516832**, window 2026-06-01 → 2026-11-03 via
`get_reduction_progress`): the pilot answered within seconds with an empty
`<afekProgressList/>` plus `errorRecord` `WS_PROCEDURE_ERROR` «Δεν βρέθηκαν δεδομένα που
ικανοποιούν τα κριτήρια» — the documented empty-window response of Πίνακας 4, not a
communication failure. This converts §7's inference into evidence: **the rejected
`deleteKedeAmount` of 2026-10-08 minted no ΑΦΕΚ** (none exists anywhere in a 5-month window).
Envelope, WSSE signature, TLS and the `KedeDeleteInterface` namespace all verified live. The
spec landmine `totalAmnt` vs `totalAmount` remained unsettled then (settled 2026-11-10: `totalAmnt`). The
follow-up audit log went to ΚΕΔ with the 35430 reply as promised; `deleteKedeComplete` stays
blocked until a `deleteKedeAmount` succeeds, exactly as stated to ΚΕΔ.

### §7 RESOLVED — 2026-11-03/04: ΚΕΔ answered; the missing knob was `afekType`

ΚΕΔ's reply to our question: *«Σχετικά με το error στη deleteKedeAmount στις περιπτώσεις που
η βεβαίωση έχει μόνο 5ψήφιους ΚΑΕ (χωρίς ύπαρξη κάποιου ΑΛΕ), τότε επιλέγουμε τον τύπο
afekType = 2 (χωρίς επιστροφή).»* All four rejections (08/10 ids 222/227/228/229, 27/10 id 629)
were submitted with `afekType=1`; our only granted ΚΑΕ (`82838`) is 5-digit — so the fix is the
request value `2`, already accepted by our validation (`DeleteKedeAmountRules.php:52`,
`in:1,2`), no code change needed for the API surface. Changed: UI default `1→2`
(`ReductionPage.tsx`), smoke tool now takes `--afek-type` with default 2, message-contract
documents the rule for Δημόκριτος. Retest: fresh end-to-end (certify → reduce with
`afekType=2` → finalize) — result recorded here once run. The two error concepts turned out
to exist in the pilot's business logic but not as request fields — they map to the
reversible/irreversible ΑΦΕΚ type, not to any «budget category» field; the §7
`inMctCod`/`inSctCod` hypothesis is therefore moot for this error.

**§7 CONFIRMED LIVE — 2026-11-04, first successful Μείωση end to end.** Fresh certification
(protocol `564/04-11-2026`, ΑΤΒ 862050, ΧΚ 26, Τ.Ο.
`137569280910104214686205000001`), then `submit_reduction` with `afekType=2`
(`ked_call_log` id 811, `callSequenceId` **10091590754**) → **ΑΦΕΚ no. 1 minted**,
`awaiting_finalization`; inline finalize via `get_reduction_status` ran `deleteKedeComplete`
(id 812, `callSequenceId` **10091590768**) → `finalized`, assessDate 2026-11-04, οριστικό
ΑΒΚΤ **404946** — under six minutes certification→finalization. The `afekType=2` rule is
therefore proven, not just documented. Docs' hold-off caveats lifted the same day
(message-contract, bff-getting-started, README, bff-call-examples).

### §7 evidence campaign: 2026-11-10, 10 × full-balance Διαγραφή, all finalized

The audit log promised to ΚΕΔ on 03/11. Ten debts already certified on the pilot on 27/10
(600.00 each, debtors 33/35/40/41/42/44/45/47/48/49) were each run through `submit_reduction`
(afekType=2, the full balance), `get_reduction_status` (inline deleteKedeComplete) and
`get_reduction_progress` (window = today). The scheduler was stopped for the run. That gave
**30 calls, all SUCCESS**: ΑΦΕΚ nos 2-11, οριστικά ΑΒΚΤ 405008-405016 plus 400182 (ΔΟΥ 8110),
protocols 698-717/10-11-2026, decisions 1101-1110/2026. Full-balance deletion with afekType=2 is
therefore proven at volume, under ~10 s per ΑΦΕΚ from submit to finalized.

- **`totalAmnt` vs `totalAmount` is settled:** the wire says **`totalAmnt`** (value `600`, no
  decimals, as in the spec sample). The parser already accepts both spellings.
- Debtor 43's Τ.Ο. `…240686201500006` was deliberately skipped, because it is the one ΑΑΔΕ was
  asked to run a test είσπραξη on. Debtor 46 was never certified (dead ΑΦΜ).
- **Ledger contamination found:** protocols 557/558 of 04-11 (audit user `smoke`,
  callSequenceId 2007…, afek_records id 6 on debtor 50) are fake-driver rows in
  `ked_audit_records`. They are excluded with `ked:audit-report --exclude-protocol` and never go
  in a pack.
- Evidence: `storage/app/ked-pilot-delete-evidence-20261110.csv` (send version: 32 SUCCESS
  rows, 11/11/10 per method) plus `…-with-errors.csv` (internal, 38 rows). Payloads and replies
  are in `storage/app/pilot-delete-20261110/`. Email draft:
  `docs/internal/ked-followup-35430-delete-audit-log.md`.
