diff --git a/AGENTS.md b/AGENTS.md
index 2b78f79..0c04829 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -26,7 +26,7 @@ Two tools scan the Vexanium blockchain voters table for accounts whose only vote
 - The node sometimes returns `staked` as a string instead of a number — normalize coerces to float (spec §V8).
 - The voters table has a sentinel first row (`owner="...........q"`, uint64 0) that must be skipped.
 - Authoritative vote weight is `last_vote_weight` (high-precision decimal string) — stored verbatim as TEXT, never float-converted (spec §V6).
-- Freshness filter (spec §V13/V40): `last_vote` (estimated last re-vote date) is derived at scan time from the Vexanium weight formula `last_vote_weight = staked_raw × 2^(weeks since 2000 / 52)` → `last_vote = 2000-01-01 + round(52 × log2(weight / (staked×10000))) weeks` — the chain (`stake2vote`) quantizes the exponent to whole weeks (`int64((now−2000-01-01)/(86400×7))/52.0`), so derived dates land on 7-day boundaries (B12; earlier 365.25-day divisor drifted ~33 days forward for 2026 votes). Since V40 `normalize()` stores **every** BP voter with a derivable `last_vote` — any age; only unverifiable votes (zero/empty weight) are dropped. Banding (BARU/VALID/KADALUARSA), display and payout are computed at query time (dashboard cuts 28/31 days on `last_vote` age; distribution pays `last_vote > now−28d`). **Maturity (3 days) is measured from `first_seen_at`, not `last_vote`** — `last_vote` is week-quantized so a new voter voting Mon–Thu would look 4–6 days old and bypass the wait (B15); V52 uses `voter_first_seen.first_seen_at` (scan-precise to the second) for the new-voter wait. The heuristic over-reads when a voter unstaked without re-voting (weight ÷ smaller stake ⇒ future date), so `normalize()` clamps any `last_vote` past scan time down to `now`. Since V41, `voter_first_seen.first_seen_at` separates PEMILIH BARU (new, first-seen within the 3-day maturity) from REVOTE (returning re-voters); since V42 REVOTE are VALID immediately (paid + shown in the main list, mint tag), only new voters wait out maturity; since V43 the mint tag lasts only `VEX_REVOTE_TAG_DAYS` (1) and VALID rows nearing the KADALUARSA cutoff (within `VEX_WARN_DAYS`, 3) get a `VOTE ULANG SEGERA` badge — display-only.
+- Freshness filter (spec §V13/V40): `last_vote` (estimated last re-vote date) is derived at scan time from the Vexanium weight formula `last_vote_weight = staked_raw × 2^(weeks since 2000 / 52)` → `last_vote = 2000-01-01 + round(52 × log2(weight / (staked×10000))) weeks` — the chain (`stake2vote`) quantizes the exponent to whole weeks (`int64((now−2000-01-01)/(86400×7))/52.0`), so derived dates land on 7-day boundaries (B12; earlier 365.25-day divisor drifted ~33 days forward for 2026 votes). Since V40 `normalize()` stores **every** BP voter with a derivable `last_vote` — any age; only unverifiable votes (zero/empty weight) are dropped. Banding (BARU/VALID/KADALUARSA), display and payout are computed at query time (dashboard cuts 28/31 days on `last_vote` age; distribution pays `last_vote > now−28d`). **Maturity (3 days) is measured from `first_seen_at`, not `last_vote`** — `last_vote` is week-quantized so a new voter voting Mon–Thu would look 4–6 days old and bypass the wait (B15); V52 uses `voter_first_seen.first_seen_at` (scan-precise to the second) for the new-voter wait. `first_seen_at`/`scanned_at` are stored UTC-naive by the scan (V62/B24) — all age cutoffs compute `datetime.now(timezone.utc)`, so the comparison is apples-to-apples; a local-TZ stamp would hold new voters in BARU ~7h longer (B24). The heuristic over-reads when a voter unstaked without re-voting (weight ÷ smaller stake ⇒ future date), so `normalize()` clamps any `last_vote` past scan time down to `now`. Since V41, `voter_first_seen.first_seen_at` separates PEMILIH BARU (new, first-seen within the 3-day maturity) from REVOTE (returning re-voters); since V42 REVOTE are VALID immediately (paid + shown in the main list, mint tag), only new voters wait out maturity; since V43 the mint tag lasts only `VEX_REVOTE_TAG_DAYS` (1) and VALID rows nearing the KADALUARSA cutoff (within `VEX_WARN_DAYS`, 3) get a `VOTE ULANG SEGERA` badge — display-only.
 - The web list shows RANK / AKUN / STAKE (VEX) / TOTAL REWARD (VEX) / VOTE TERAKHIR columns (3 stats cells). TOTAL REWARD = all-time sum of `distribute_payments.amount` where `status='sent'` (0,0000 for never-paid), joined per page via a LEFT JOIN subquery. Vote weight stays in the DB and `/api/search` JSON but is not rendered as a column.
 - Token contract is `vex.token` (NOT `eosio.token` — that name doesn't exist on Vexanium), 4-decimal VEX, chain_id `f9f432b1851b5c179d2091a96f593aaed50ec7466b74f89301f957a83e56ce1f`. Distribution signs `vex.token::transfer` with the BP `active` key.
 - Distribution (spec §V19–V27): each run pays the whole liquid balance pro-rata by stored `staked`, shares floored at 4 decimals with dust left in the account, one transfer per voter. A txid that can't be confirmed (`GET {HYPERION_API}/v2/history/get_transaction?id=<txid>` returns no `executed`) is never re-sent the same run — that payment stays `failed` and rejoins the next run. No-op (exit 0, no writes) when the voters table is empty or balance < 0.0001. The dashboard's `/history` view renders `distribute_runs` + `distribute_payments` read-only.
@@ -37,7 +37,7 @@ Two tools scan the Vexanium blockchain voters table for accounts whose only vote
 ## Layout
 
 - `SPEC.md` — spec (goal/constraints/interfaces/invariants/tasks/bug log), in caveman encoding. Build/backprop flow through it.
-- `get_voters.py` — production fetcher: scan → filter (stake-min + BP target, semua umur vote disimpan V40; hanya last_vote tak terverifikasi dibuang) → `db.replace_snapshot` (each run replaces the table = daily snapshot, with `scanned_at` + derived `last_vote`; never appends history) → `db.record_first_seen` (V41: catat kemunculan pertama tiap owner ke `voter_first_seen`, idempoten).
+- `get_voters.py` — production fetcher: scan → filter (stake-min + BP target, semua umur vote disimpan V40; hanya last_vote tak terverifikasi dibuang) → `db.replace_snapshot` (each run replaces the table = daily snapshot, with `scanned_at` + derived `last_vote`; never appends history) → `db.record_first_seen` (V41: catat kemunculan pertama tiap owner ke `voter_first_seen`, idempoten). V62: `scanned_at`/`first_seen_at` disimpan UTC-naive (`datetime.now(timezone.utc)`) — konsisten dgn SEMUA cutoff umur (dashboard + distribusi) & `last_vote` (epoch UTC); jangan kembali ke `datetime.now()` lokal (B24: pemilih BARU tertahan +7 jam).
 - `db.py` — storage abstraction (sqlite default | mysql via PyMySQL); `connect/query/queryone/replace_snapshot`; `%s` → `?` for sqlite; `query` returns list. Juga tabel distribusi: `distribute_runs` + `distribute_payments` (append-only) + helper `ensure_distribute_schema/record_run/record_payment/update_payment_status/update_run_status/list_runs/list_payments`. Juga tabel klaim: `claim_runs` + helper `ensure_claim_schema/record_claim/update_claim_fee/pending_claim_fee`. Juga tabel V41 `voter_first_seen` (owner PK + `first_seen_at`, append-only) + helper `ensure_first_seen_schema/record_first_seen` — dipakai dashboard membedakan PEMILIH BARU vs REVOTE. Juga helper V57 `eligible_voters(owner=None)` — single-source SQL jendela reward (payout window, V40/V52): `last_vote > now−STALE AND first_seen_at IS NOT NULL AND first_seen_at ≤ now−MATURITY` (cutoff UTC-naive); tanpa `owner` → semua baris `(owner, staked)` utk distribusi; dgn `owner` → baris akun itu (`[]`|one row) utk endpoint cek.
 - `get_voters.js` — reference implementation only.
 - `distribute.py` — pembayaran harian: fetch liquid balance → baca pemilih **band reward V40/V52** (`db.eligible_voters()` — V57 single source: `last_vote > now−28d AND first_seen_at ≤ now−3d`, pemilih baru (`first_seen` baru/NULL) & basi ⊥ dibayar; akun yang sudah matang dibayar langsung, ⊖ peduli umur vote, V52) → `compute_shares` (Decimal floor 4-des, sisa di akun) → sign `vex.token::transfer` via pyntelope (`trx.link` ambil ABI+TAPOS dari node, `sign` dengan `VEX_BP_PRIVATE_KEY`, `.send()`) → catat `sent/failed` per voter; verifikasi txid via Hyperion sebelum kirim ulang; `--dry-run` = rencana ⊥ tanda tangan; notifikasi Telegram mulai/selesai/gagal via `telegram.py` (best-effort, ⊥ dry-run/no-op). **V58 failover node pool**: `_post_json` coba SEMUA `API_NODES` per putaran (semua gagal → RuntimeError); `build_signed_transfer` coba tiap node utk link+sign (net ter-bind → send ikut node sama); `fetch_balance_with_retry` backoff `[30,60,90,120,150,180]s` (~10 mnt) di real-run sebelum menyerah ke jadwal berikutnya, dry-run satu attempt. **V59 send aman**: `verify_txid` → None (Hyperion down) = TAHAN sungguhan (break + `failed` di attempt 1, ⊖ resend dgn tapos baru yang bisa ganda). **V61 send-path hardening**: `_send_rejected(resp)` WHITELIST — sukses HANYA dict ber-`transaction_id` + receipt `executed`/tanpa-status, body HTTP-500 apa pun (pyntelope ⊥ raise, pola V46/B11) → tolak → raise → jalur verify/retry, ⊖ pernah `sent` utk tx yg ⊥ mendarat; `_verify_settled(txid)` retry `verify_txid` 3× dgn delay utk lag indeks (None → False langsung, pool mati semua) — jalur exception send-loop hold (`failed`+break) pada apa pun selain True, ⊖ RESEND utk `executed:false` (B21). **V60 Hyperion failover**: `_get_json` (GET mirror `_post_json`, 3 putaran, 2s, coba semua `HYPERION_NODES` per putaran → semua gagal → None) dipakai `verify_txid` — None kini berarti SEMUA host Hyperion gagal, jadi hold V59 jarang palsu; semua pembaca `/v2/*` guard `isinstance(data, dict)` (non-dict 200 → skip/skip-node/None, ⊖ crash run, B23). `test_distribute.py` — oracle (mock chain+pyntelope, temp DB).
diff --git a/SPEC.md b/SPEC.md
index 879b602..47d6b63 100644
--- a/SPEC.md
+++ b/SPEC.md
@@ -128,6 +128,7 @@ V58: hardening distribusi vs node RPC mati: `config` += `API_NODES` (CSV `VEX_AP
 V60: failover pool Hyperion utk SEMUA pembacaan `/v2/*`: `config` += `HYPERION_NODES` (CSV `VEX_HYPERION_NODES`, default = `HYPERION_API` + `API_NODES` dedup, utama di depan — kedua host mainnet terverifikasi layani `/v2`, probe 2026-08-13). `distribute._get_json(url, params, max_retries=3, timeout=10)` — GET mirror `_post_json` V58: coba SETIAP node per putaran, `max_retries` putaran penuh, jeda 2s; SEMUA gagal → None (⊥ raise — pemanggil butuh None utk jalur tak-terukur/hold). `verify_txid` pakai `_get_json` → None kini berarti SEMUA host gagal SEMUA putaran (bukan satu host hiccup) → hold V59 jarang palsu. `claim.reward_from_tx` pakai `dist._get_json` (tiap attempt settle V55 jadi pool-robust). `dashboard._get_liquid_vex` loop `HYPERION_NODES` sendiri (web ⊥ impor distribute, V31) — host mati → cadangan, SALDO LIQUID tetap live. Net-effect: blip Hyperion sesaat ⊖ lagi menahan payout / ⊖ gagal ukur reward / ⊖ kosongkan saldo dashboard. Oracle: test_distribute V60 verify_txid failover (host mati → cadangan `executed`; semua mati → None), test_claim V60 reward_from_tx failover, test_dashboard V60 saldo dari cadangan + config default HYPERION_NODES | V60,V22,V45,V16
 V59: kirim per-pemilih ⊖ pernah menyesatkan: (1) TAHAN yang sungguhan saat `verify_txid` → None (Hyperion tak terjangkau): break + tandai `failed` di ATTEMPT PERTAMA (bukan hanya di attempt terakhir — B19: kode lama print `ditahan` tapi lanjut retry → re-sign dgn tapos/expiration baru → txid BARU → re-broadcast berpotensi ganda walau tx pertama mendarat; oracle lama lolos karena set `DISTRIBUTE_MAX_ATTEMPTS=1`); (2) deteksi penolakan chain dari body HTTP-500: pyntelope `send()` ⊥ raise pada 500 (pola V46/B11), respons `{"code":500,"error":...}` datang sbg dict — `_send_rejected(resp)` mengenali bentuk error → raise RuntimeError → jalur verify/retry, ⊖ tandai `sent` utk tx yang ⊥ pernah mendarat (B19 juga: tanpa deteksi, penolakan duplicate/insufficient-balance dicatat `sent` dan ⊥ pernah dikejar). Oracle test_distribute: V59 hold-atau-retry di attempt 1 (3 attempt default, build dipanggil sekali per pemilih), V59 rejection-500 → `failed` + partial (⊖ `sent`) | V59,V22,V21
 V61: red-team send-path hardening (audit V58/V59/V60, B20–B23): (1) `_send_rejected(resp)` = WHITELIST, bukan blacklist: sukses HANYA bila resp dict ber-`transaction_id` DAN receipt `executed`/tanpa-status; body HTTP-500 apa pun (ber-`error` ATAU bentuk proxy lain) → dianggap DITOLAK → jalur verify/retry, ⊖ pernah `sent` utk tx yang ⊥ mendarat (F3/B22; B19-b hanya mengejar bentuk `{"code":500,"error":...}`, bentuk lain lolos sbg sukses); (2) `_verify_settled(txid, retries=3, delay=2)`: retry `verify_txid` beri Hyperion waktu mengejar indeks (pola B14), True → mendarat; None (pool mati SEMUA) → False langsung (⊥ tidur sia-sia); False konsisten → False — jalur exception send-loop sekarang hold (`failed`, break) pada True→sent / None|False→hold, ⊖ RESEND apa pun hasil verify selain True — tx yang `executed:false` (lag indeks) ⊖ boleh re-sign dgn tapos baru → txid baru → ganda (B21, perluasan V59-b: V59 hold hanya saat `verify_txid → None`, False FALLTHROUGH ke resend); (3) `_send_fee` (claim fee) pakai `_send_rejected(resp)` — penolakan chain fee (duplicate/insufficient-balance, V46/B11) → `failed` + retry siklus berikutnya, ⊖ `sent` untuk fee yang ⊖ pernah mendarat (F1/B20); (4) SEMUA respons `/v2/*` di-guard `isinstance(data, dict)` — `_get_json`, `verify_txid`, `reward_from_tx`, `dashboard._get_liquid_vex` → bentuk non-dict (200-an list/string dari proxy) → skip/skip-node/None, ⊖ `AttributeError` di dalam `except` yang membatalkan run (F5/B23). Oracle: test_distribute V61 (verify-False → hold attempt-1 dgn MAX_ATTEMPTS=3 default, `_send_rejected` whitelist 5 bentuk, `_get_json`/`verify_txid` non-dict → None), test_claim V61 (fee rejection-500 → `failed` ⊖ `sent`), test_dashboard V60 (override `VEX_HYPERION_NODES`). F4 (`_get_json` first-200-wins: primary basi menaungi cadangan — hari ini kedua host = backend vexascan.com sama, efek nol) & F7 (hold latency ~64 s/pembayaran saat SEMUA Hyperion mati; jalur claim chain node tunggal) — DITERIMA, ⊖ diubah | V61,V60,V59,V56,V46
+V62: stempel scan UTC-naive (B24): `get_voters.persist` menulis `scanned_at`/`first_seen_at` dgn `datetime.now(timezone.utc).replace(tzinfo=None)` — sama basis waktu dgn SEMUA cutoff umur (dashboard `_freshness_cutoffs`/`_revote_cut`/countdown + `db.eligible_voters` + `distribute` semuanya `datetime.now(timezone.utc)`) dan dgn `last_vote` (diturunkan dari epoch 2000 UTC). Dulu `datetime.now()` (waktu LOKAL container `TZ=Asia/Jakarta`, UTC+7) → `first_seen_at` tersimpan WIB tapi dibandingkan sbg string UTC → semua ambang maturity BARU/REVOTE bergeser +7 jam: pemilih baru tertahan di band BARU ~7 jam lebih lama dari seharusnya (B24). Migrasi data existing: UPDATE `voter_first_seen`/`voters` kurangi 7 jam (WIB→UTC), backup tabel `*_bak_wib` di prod. Oracle test_get_voters V62: stempel ≈ now-UTC-naive (selisih <2 mnt) DAN ≠ now−7h (guard anti-WIB) | V62,V52,V40,V41
 V35: kill-switch distribusi `DISTRIBUTE_ENABLED` default false → run nyata (bukan dry-run) no-op: exit 0, ⊥ baca saldo, ⊥ tanda tangan, ⊥ tulis DB, ⊥ notif; `--dry-run` tetap menampilkan rencana (read-only); nilai `true`/`1`/`yes` → normal
 V36: notifikasi klaim reward (sukses `notify_claim` & gagal `notify_claim_failure`) → kanal INTERNAL saja, ⊥ pernah ke KOMUNITAS (V30); fee gagal ⊥ spam — satu notif per klaim (saat transisi ke `failed`)
 V37: jadwal distribusi MINGGUAN (⊥ harian): `distribute_loop` menunggu `DISTRIBUTE_WEEKDAY` — SATU hari (`sat`) atau CSV beberapa hari (`wed,sat`), nama `mon`..`sun` atau 0-6 (V44: multi-hari, event terdekat di antara hari-hari terdaftar) — jam `DISTRIBUTE_HOUR` (lokal TZ); plus satu one-off `DISTRIBUTE_FIRST_RUN` (`YYYY-MM-DD`; default baked `2026-08-17`) yang diambil bila lebih dekat dari mingguan; setelah lewat (atau ⊥ diset) → hanya mingguan; batas dihitung ulang tiap iterasi loop (setelah run → jadwal berikutnya); payout di-tunda bila `DISTRIBUTE_ENABLED=false` (V35) — jadwal tetap maju, run jadi no-op; `distribute_loop.main` baca jadwal via `config` (⊥ `os.getenv` sendiri) agar konsisten dgn dashboard V38
@@ -190,6 +191,7 @@ T51|x|hardening distribusi vs node RPC mati V58: `config.py` += `API_NODES` (CSV
 T52|x|send-per-pemilih ⊖ menyesatkan V59: `distribute.py` — (1) `verify_txid` None → break + `failed` di attempt 1 (hold sungguhan, ⊖ resend; B19-a), (2) `_send_rejected(resp)` deteksi body HTTP-500 (duplicate/insufficient-balance) → raise → jalur verify/retry, ⊖ `sent` palsu (B19-b); docs §V/V59 + §B/B19 + T52 + AGENTS; oracle test_distribute (V59 hold attempt-1 dgn 3 attempt default + rejection-500 → failed/partial)|V59,V22,V21
 T53|x|failover pool Hyperion V60: `config.py` += `HYPERION_NODES` (CSV `VEX_HYPERION_NODES`, default `HYPERION_API` + `API_NODES` dedup); `distribute.py` `_get_json(url, params, max_retries=3, timeout=10)` GET mirror `_post_json` — coba semua `HYPERION_NODES` per putaran, semua gagal → None; `verify_txid` + `claim.reward_from_tx` (via `dist._get_json`) + `dashboard._get_liquid_vex` (loop lokal, web ⊥ impor distribute V31) pindah ke pool; docs §I/§V/V60 + T53 + AGENTS; oracle test_distribute (failover verify_txid + semua-mati None), test_claim (reward_from_tx failover), test_dashboard (saldo cadangan + config default)|V60,V22,V45,V16
 T54|x|red-team send-path hardening V61 (B20–B23): `distribute.py` — `_send_rejected` jadi WHITELIST (sukses = dict + transaction_id + receipt executed/tanpa-status), `_verify_settled(txid)` retry verify utk lag indeks, jalur exception send-loop hold pada None|False (⊖ resend, perluasan V59-b), guard `isinstance(data, dict)` di `_get_json`+`verify_txid`; `claim.py` `_send_fee` + `_send_rejected` (rejection-500 → failed ⊖ sent) + guard non-dict di `reward_from_tx`; `dashboard._get_liquid_vex` guard non-dict; docs §V/V61 + §B/B20-B23 + T54 + AGENTS; oracle test_distribute (verify-False hold attempt-1 MAX_ATTEMPTS=3, whitelist 5 bentuk, non-dict → None), test_claim (fee rejection → failed), test_dashboard (override VEX_HYPERION_NODES); F4/F7 diterima (dokumentasi)|V61,V60,V59,V56,V46
+T55|x|stempel scan UTC-naive V62 (B24): `get_voters.persist` ganti `datetime.now()` → `datetime.now(timezone.utc).replace(tzinfo=None)` utk `scanned_at`/`first_seen_at` (konsisten dgn semua cutoff umur UTC + `last_vote`); migrasi data existing DB prod WIB→UTC (−7 jam, backup `*_bak_wib`); docs §V/V62 + §B/B24 + T55 + AGENTS; oracle test_get_voters V62 (stempel ≈ now-UTC-naive, ⊖ WIB)|V62,V52,V41,V7
 
 ## §B — Bug log
 id|date|cause|fix
@@ -216,3 +218,4 @@ B20|2026-08-13|fee klaim (claim.py `_send_fee`) ⊖ punya guard `_send_rejected`
 B21|2026-08-13|V59 hanya menahan saat `verify_txid → None` (Hyperion tak terjangkau) — `False` (`executed:false`, lag indeks V50/B14) FALLTHROUGH ke jalur retry → re-sign tapos BARU → txid BARU → broadcast ulang → GANDA kalau tx pertama sebenarnya mendarat; risiko nyata setiap Hyperion tertinggal indeks sesaat (pola yang sudah dibuktikan di klaim B14)|jalur exception send-loop: `_verify_settled(txid)` (retry 3× dgn delay utk lag indeks), True → sent+break; None (pool mati semua) ATAU False konsisten → hold `failed` di attempt 1, ⊖ RESEND (V61, F2); oracle test_distribute V61 verify-False → hold attempt-1 dgn MAX_ATTEMPTS=3 default
 B22|2026-08-13|`_send_rejected` BLACKLIST (deteksi `error` di body) — body 500 bentuk lain (proxy tanpa `error`, `{"code":500,"message":...}`) atau 200 `soft_fail`/`delayed` receipt dianggap SUKSES → dicatat `sent` walau tx gagal/ditolak|whitelist: sukses HANYA dict ber-`transaction_id` + receipt `executed`/tanpa-status; sisanya → tolak (V61, F3); oracle `_send_rejected` 5 bentuk (sukses nodeos, 500 ber-error, 500 tanpa error, soft_fail, non-dict)
 B23|2026-08-13|respons `/v2/*` bentuk non-dict (200-an `[...]`/`"..."` dari proxy salah) → `AttributeError: 'list' object has no attribute 'get'` DII DALAM `except RequestException` → run distribusi BEBENTI total; sama di `verify_txid`/`reward_from_tx`/`dashboard._get_liquid_vex`|guard `isinstance(data, dict)` di semua pembaca `/v2/*` → non-dict → skip/skip-node/None (V61, F5); oracle `_get_json`/`verify_txid` non-dict → None
+B24|2026-08-17|`first_seen_at`/`scanned_at` tersimpan WIB (UTC+7) tapi dibandingkan sbg string UTC di SEMUA cutoff umur (dashboard + `db.eligible_voters` + distribusi): `get_voters.persist` memakai `datetime.now()` (waktu lokal container `TZ=Asia/Jakarta`) sedangkan cutoff pakai `datetime.now(timezone.utc)` — offset +7 jam membuat pemilih BARU tampak belum matang ~7 jam lebih lama (kasus nyata: idrsaku33351/m4dhanchocil/danuprasetya first_seen 13 Agt 17:39 WIB, seharusnya matang 16 Agt 17:39 WIB, kode melepas dari BARU baru 17 Agt 00:39 WIB)|`persist` stempel UTC-naive (`datetime.now(timezone.utc).replace(tzinfo=None)`); migrasi data existing WIB→UTC (−7 jam) dgn backup `*_bak_wib` (V62); oracle V62 guard stempel ≈ now-UTC ⊖ WIB
diff --git a/get_voters.py b/get_voters.py
index 1093107..2f48381 100644
--- a/get_voters.py
+++ b/get_voters.py
@@ -154,7 +154,8 @@ def persist(rows):
     Lalu `db.record_first_seen` (V41): catat kemunculan pertama tiap owner
     (idempoten) utk membedakan pemilih BARU vs REVOTE di dashboard.
     """
-    scanned_at = datetime.now().isoformat(timespec='seconds')
+    scanned_at = datetime.now(timezone.utc).replace(tzinfo=None).isoformat(
+        timespec='seconds')
     db.replace_snapshot(rows, scanned_at)
     db.record_first_seen(rows, scanned_at)
     return scanned_at
diff --git a/test_get_voters.py b/test_get_voters.py
index 685aafb..6637e45 100644
--- a/test_get_voters.py
+++ b/test_get_voters.py
@@ -10,7 +10,7 @@ membuktikan ia lolos ke snapshot, bukan dibuang.
 import os
 import sqlite3
 import tempfile
-from datetime import datetime, timedelta
+from datetime import datetime, timedelta, timezone
 
 import requests
 
@@ -229,6 +229,20 @@ def main():
     check('V7 owner PRIMARY KEY', 'owner' in cols and pk == ['owner'])
     check('V7 kolom last_vote ada', 'last_vote' in cols)
 
+    # V62: stempel scan UTC-naive (konsisten dgn cutoff umur dashboard/
+    # distribusi yang pakai `datetime.now(timezone.utc)`). Dulu `datetime.now()`
+    # (waktu lokal container, TZ=Asia/Jakarta UTC+7) → `first_seen_at` WIB
+    # dibandingkan sbg string UTC → maturity pemilih BARU bergeser +7 jam
+    # (B24). Guard: stamp ≈ now-UTC-naive, ⊖ WIB (offset ±2 mnt toleransi).
+    utc_now = datetime.now(timezone.utc).replace(tzinfo=None)
+    for s in (scanned_a, scanned_b):
+        stamp = datetime.fromisoformat(s)
+        delta = abs((utc_now - stamp).total_seconds())
+        check(f'V62 stempel scan UTC-naive ({s})',
+              delta < 120 and stamp.tzinfo is None)
+        wib_delta = abs((utc_now - timedelta(hours=7) - stamp).total_seconds())
+        check(f'V62 stempel scan ⊖ WIB ({s})', wib_delta > 60)
+
     # V41: first_seen — kemunculan pertama tiap owner, idempoten (⊥ overwrite)
     db.record_first_seen([{'owner': 'aaa1'}], '2000-01-01T00:00:00')
     conn = sqlite3.connect(db.DB_PATH)