Awas— watch out Open the app

17 Okt 2026 rekaman

Awas answers the family chat with pages, not scores.

A job offer arrives — customer service in Cambodia, 2,000 USD — and someone must answer tonight. Awas answers with the official rule, evidence code found word for word, and a stamp. Out of sample: 9 of 25 reported scam offers got STOP and 15 more VERIFY FIRST; 0 of 10 licensed vacancies got STOP.

Paste an offer Pitch deck Source on GitHub

No login · 5 recorded runs on /judge · try to fool the gate

Awasberkas bukti · evidence file
rekaman · recorded 2026-10-17
the offer, as KP2MI described it (2022) — not the ad

PMI ([nama]) asal Medan mengaku mendapat tawaran bekerja di Kamboja dari Facebook. [nama] dijanjikan bekerja sebagai Costumer Service sebuah perusahaan dengan gaji sebesar 2000 USD.

Negara tujuan bukan negara penempatan resmi pekerja migran · The destination is not an official placement country NEVER rule → STOP · kp2mi.go.id · 2025-10-28
“Perlu saya tegaskan, Kamboja itu bukan negara penempatan.”
liks.suara.com · context, not a verdict inputkuning (yellow) = found word for word “Para WNI direkrut melalui media sosial dengan janji pekerjaan legal. Kenyataannya, setibanya di Kamboja, mereka justru terperangkap.” These Indonesian citizens were recruited via social media with promises of legal work. In reality, upon arriving in Cambodia, they instead became trapped. open the sentence
“Korban scam Usaha untuk kabur tidak tiba-tiba terjadi, cerita , salah satu korban.” name masked bbc.com · not found in the page — never shown as evidence
13.6 s · 3 Nemotron calls · 1 Tavily credit · 4 more lines found open this case
A recorded evidence file: STOP.

17 Okt 2026 tranche 2

Thirty-five cases. One graded run.

25 reported scam offers and 10 licensed vacancies, never tuned on, graded in one run: a Nebius Serverless Job sent them through the live app, its src/lib frozen at code-freeze.

9 / 25

reported scam offers got STOP — 15 more VERIFY FIRST, 1 missed

0 / 10

licensed agencies’ vacancies got STOP

25 publicly reported scam offers

  1. VERIFY FIRSTvery high pay, light work
  2. STOPnot a placement country
  3. VERIFY FIRSTno company named
  4. VERIFY FIRSTa fee to get the job
  5. VERIFY FIRSTa very fast departure
  6. STOPtourist visa for work
  7. VERIFY FIRSTa fee to get the job
  8. VERIFY FIRSTa fee to get the job
  9. VERIFY FIRSTa fee to get the job
  10. STOPnot a placement country
  11. VERIFY FIRSTa fee to get the job
  12. VERIFY FIRSTvery high pay, light work
  13. STOPnot a placement country
  14. NO RED FLAGS FOUND
  15. STOPnot a placement country
  16. STOPnot a placement country
  17. VERIFY FIRSTno company named
  18. VERIFY FIRSTa fee to get the job
  19. STOPanother outlet’s report names the company
  20. VERIFY FIRSTvery high pay, light work
  21. VERIFY FIRSTin an official warning
  22. VERIFY FIRSTno company named
  23. STOPnot a placement country
  24. VERIFY FIRSTno company named
  25. STOPtourist visa for work

10 licensed agencies’ vacancies

  1. NO RED FLAGS FOUND
  2. NO RED FLAGS FOUND
  3. NO RED FLAGS FOUND
  4. NO RED FLAGS FOUND
  5. VERIFY FIRSTa page names a similar company
  6. NO RED FLAGS FOUND
  7. NO RED FLAGS FOUND
  8. NO RED FLAGS FOUND
  9. NO RED FLAGS FOUND
  10. NO RED FLAGS FOUND
48/48
quotes shown, all found word for word6 of 54 dropped · plain-GET re-check: 33 found, 14 not checkable by script, 1 changed
13.7 s
median to a verdictp95 31.7 s
$0.0078
median Token Factory cost per checkat list prices · + 1–3 Tavily credits
19/20
same verdict on the rerun set in bench.py before the run (10 licensed)20/20 on a later rerun of the first 10 reported cases

The ninth STOP is the one report naming its company, matched in another outlet’s report: one case, not a rate. An earlier tranche, run before the freeze, made three misses visible: two were fixed first, one (domestic work in Qatar) is a documented gap. every row · aijob-e03nhcfcwc4vnee33m

README limits

What failed, said plainly.

A newspaper’s corrections box is the ralat. Four of the README’s six honest limits, verbatim:

RALAT
  1. “Our day-0 search-coverage gate failed (8 of 20 known cases had a relevant item; we had set 12). The web evidence is therefore context; the verdict rests on the cited rules.”
    — README · Honest limits, 1

    Done: the verdict moved onto cited rules, before tranche 1 ran.

  2. “The benchmark's reported cases are reporters' descriptions of offers, not the ads.”
    — README · Honest limits, 2

    Done: each case says what it is: the file at the top is “as KP2MI described it”.

  3. “Indonesian is not a listed language for the Nemotron models used (per their model cards)”
    — README · Honest limits, 3

    Measured instead: field accuracy on 25 dev texts, against labels our agent drafted, not yet reviewed.

  4. “The phone path is unbenchmarked. A phone number with no company name returns VERIFY FIRST, not a verdict.”
    — README · Honest limits, 4

    Why: benchmark rows carry no phone numbers.

Also: 53 collected rows were discarded before the pipeline read any of them — their publishers forbid AI use (protocol amendment 2).

in your browser

Try to fool the gate.

The rules and the quote gate run here, in your browser — the app’s own src/lib.

The quote gate

6+ words, verbatim (case, quotes, dashes, spacing aside).

the page · humas.polri.go.id · 2026-02-04
Identitas perusahaan tidak jelas, alamat kantor tidak pasti dan komunikasi hanya dilakukan melalui media sosial atau WhatsApp tanpa pertemuan langsung di kantor resmi. Dokumen bermasalah, menggunakan visa turis atau visa kunjungan untuk bekerja, bukan visa kerja (work permit) yang sah.
“Dokumen bermasalah, menggunakan visa turis atau visa kunjungan untuk bekerja, bukan visa kerja (work permit) yang sah.”found verbatim in one chunk of humas.polri.go.id · 17 words · open it on the page

The rules

Edit the offer; the rules re-run as you type.

STOP — rule: tourist or visit visa offered for work + rule: the destination is not an official placement country for Indonesian workers. 6 rules fired; a sourced NEVER rule decides it.

17 Okt 2026 06:58 UTC

13.6 seconds, start to stamp.

The file at the top is this run, recorded 17 October 2026 on the frozen pipeline — open on /judge, its receipt one command away.

Awas — the recorded evidence file in the live app: the offer underlined, the STOP stamp, the rules with their official sources
the live app’s file: the stamp and its rules
Awas — a context card: the quote highlighted yellow because the gate found it word for word on liks.suara.com, with its translation and link
a line the gate found word for word
Awas — the message to forward to the family, in Bahasa, with copy and share buttons
the answer for the family chat, ready to forward

Durations from the stored receipt, laid end to end.

curl -sL https://awas.edycu.dev/api/case/qm5h5lnvunjgjlyqbw6kkabtya
  1. extractnvidia/Nemotron-3-Ultra-550b-a55b · 1,081 tok · 2.9 s
  2. rulesnon_placement_country · NEVER · “Kamboja” · kp2mi.go.id 2025-10-28
  3. searchTavily · “WNI korban penipuan lowongan kerja Kamboja” · 5 pages · 1 credit · 1.8 s
  4. proposenvidia/nemotron-3-super-120b-a12b · 4,814 tok · 3.7 s
  5. gate6 proposed · 5 found word for word · 1 dropped
  6. messagenvidia/nemotron-3-super-120b-a12b · 1,567 tok · 4.7 s
  7. verdictSTOP — rule: the destination is not an official placement country for Indonesian workers
  8. total13.6 s · 7,462 tokens · 1 Tavily credit · kept in Object Storage

condensed from that JSON · no key needed

kemlu.go.id kp2mi.go.id

12,019 asked to come home.

Cut from the ministries’ own pages; each sentence is in docs/SOURCES.md.

12,019Indonesians from scam networks in Cambodia asked the embassy for help getting home, Jan–Jun 2026kemlu.go.id · KBRI Phnom Penh · 2 July 2026
1,301 · 85%cases at the embassy in Cambodia, Q1 2025 — 85% online scamskemlu.go.id · KBRI Phnom Penh · 24 April 2025
“kerja di IT atau marketing digital”how the offer arrives, in the ministry’s wordskp2mi.go.id · KP2MI · 28 October 2025

catatanThe embassy’s 1,301-case page counts people who ran scams, not only victims. Awas claims less: the offers follow a pattern a family can check before anyone leaves.

src/lib frozen

The model reads. Code decides.

Nemotron reads and proposes; code decides — unchanged since code-freeze, under 144 unit and 56 browser tests.

Reads the offer

Nemotron 3 Ultra copies typed fields; code checks each is in the text.

nvidia/Nemotron-3-Ultra-550b-a55bcall 1

Looks it up

Tavily searches the exact company and number, and scam reports for the destination: context, never a rule input. Only a verified line naming the offer’s own company beside a harm word can raise it to STOP.

tavily search · exact_matchtavily extract

Checks every quote

Nemotron 3 Super proposes lines; code shows a quote only if 6+ words sit verbatim in a page Tavily returned. Else it is dropped, on screen.

Awas — an official Polri page opened from a gate link: the browser scrolls to the quoted sentence and marks it
A gate link, opened: the page scrolls to the sentence and marks it.
nvidia/nemotron-3-super-120b-a12bcall 2gate.ts

Decides, then writes

A truth table sets the verdict; Super writes the family message from verified lines.

STOPVERIFY FIRSTNO RED FLAGS FOUND

verdict.tsnvidia/nemotron-3-super-120b-a12bcall 3
Nebius Token Factorysrc/lib/tf.ts
Tavily searchsrc/lib/tavily.ts
Tavily extractsrc/lib/tavily.ts
Nebius Object Storagesrc/lib/nebius/storage.ts
Nebius Serverless Jobssrc/lib/nebius/jobs.ts

tanya jawab

Questions a judge would ask.

Who decides STOP — the model?

Code: verdict.ts — a sourced NEVER rule, or a verified line naming the offer’s own company beside a harm word. 8 of 9 tranche-2 STOPs came from rules. No score, no green; the mildest verdict says it is not a guarantee.

What if the model invents a quote?

The gate drops it, on screen — 6 of 54 in tranche 2. Flip one word in the lab and watch.

How accurate is it?

Out of sample: 0 of 10 licensed vacancies got STOP; of 25 reported scam offers, 9 STOP, 15 VERIFY FIRST, 1 missed — reporters’ descriptions, not the ads. Every row: bench/results-t2.jsonl.

Does it accuse companies?

Its own sentences are checked in code not to name the offer’s company. A name appears only in the pasted offer, a verified quote and its translation, or beside “Dilaporkan oleh <sumber>, bukan tuduhan kami” — reported by the source, not our accusation.

What happens to what I paste?

The text goes to Nemotron on Token Factory; only a company name or number goes to Tavily. Each check is stored in Nebius Object Storage so its link keeps working; the app says disimpan 75 hari, tidak diindeks mesin pencari, nomor telepon disamarkan — kept 75 days, not indexed, phone numbers masked (the expiry rule’s receipt is still pending). Don’t paste personal data.

Paste one offer.

No login, no daily cap (60 checks an hour per connection) — or replay 5 recorded runs on /judge.

Open the app Recorded runsPitch deck