Awas
Pages, not scores .
Paste a job offer. Get the rule, its page and a stamp — decided in code.
Nebius Global AI Hackathon · v0.1.3
$ curl …/api/case/qm5h5l…STOP · 13.6 s · 3 Nemotron calls
Awas answers a forwarded job offer the way a careful clerk would: the rule, the official page it comes from, and a stamp. This is a real recorded run, not a mock-up — the receipt on the right is one curl away.
Someone answers tonight.
as KP2MI described it, 2022
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 .
Customer service in Cambodia, 2,000 USD.
Offers like this arrive through social media — "kerja di IT atau marketing digital", in the ministry's own words. Customer service in Cambodia, two thousand dollars. Someone in the family has to answer tonight, and the rules that would stop it are scattered across police, ministry and embassy pages.
They asked to come home.
12,019 Indonesians from scam networks asked the embassy in Cambodia for help getting home, Jan–Jun 2026.
kemlu.go.id · 2 Jul 2026
“kerja di IT atau marketing digital” how the offer arrives
kp2mi.go.id · 28 Oct 2025
Twelve thousand and nineteen Indonesians reported to the embassy in Phnom Penh and asked for help getting home in the first half of 2026. The offer arrives as "work in IT or digital marketing". The embassy's 1,301-case page also counts people who ran scams, so we claim something narrower: the offers follow a pattern a family can check.
An evidence file, not a score .
1
2
3
1 The stamp, set by a truth table.
2 The rule, with its official page.
3 A quote code found word for word.
The answer is a file, not a number. One: the stamp, which no model sets. Two: the rule that fired, with the official page and date it comes from. Three: context from published pages — highlighted only because code found the sentence on the page.
The model reads. Code decides.
Offer pasted text Extract Nemotron 3 Ultra 16 rules 4 NEVER · licence list 534 Verdict verdict.ts truth table Search Tavily search + extract Propose Nemotron 3 Super Quote gate ≥ 6 words, verbatim Message Nemotron 3 Super Evidence pack Nebius Object Storage
Nemotron on Token Factory code Tavily
Nemotron 3 Ultra reads the offer into typed fields; code checks every field is really in the text. Sixteen rules in code decide — fourteen quote official pages, two check a dated snapshot of the licensed-agency (P3MI) list. Tavily finds context; Nemotron 3 Super proposes quotes; a code gate keeps only quotes found word for word; a truth table sets the verdict.
Verbatim, or it’s dropped.
Found. Word for word, on the fetched page.
this run: 5 of 6 found tranche 2: 6 of 54 dropped
“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
Not there. Struck, dropped.
This is the line I want you to remember. The model proposed both quotes. Code found the first one, word for word, in the page Tavily returned — so it gets the yellow. The second was not on the page, so it is struck and never shown as evidence.
Every NEVER rule has a page.
“Perlu saya tegaskan, Kamboja itu bukan negara penempatan.” kp2mi.go.id · 2025-10-28
Not a placement country kp2mi.go.id Tourist visa for work humas.polri.go.id Fee before a contract kemnaker.go.id Passport held humas.polri.go.id
Any one of four NEVER rules is a STOP, and each one quotes an official page — here the Minister for migrant workers: "Cambodia is not a placement country." 12 more rules lead to VERIFY FIRST, and a named company is checked against a dated snapshot of the licensed-agency list (534 agencies).
One graded run, frozen code.
25 reported scam offers 10 licensed vacancies
STOP VERIFY FIRST NO RED FLAGS FOUND
0/10 licensed vacancies stopped
9/25 reported scam offers got STOP · 15 more VERIFY FIRST
13.7 s median check
$0.0078 median Token Factory cost, list prices
a Nebius Serverless Job
Thirty-five texts we never tuned on, frozen before any model read a benchmark row, and graded in one run — a Nebius Serverless Job sent them through the live app, its src/lib frozen at the tag code-freeze. Top: twenty-five publicly reported scam offers — nine STOP, fifteen VERIFY FIRST, one missed. Bottom: ten vacancies from licensed agencies — none got STOP. Twenty reruns of ten reported cases gave the same verdicts. The reported cases are reporters' descriptions, not the ads.
What failed, said plainly.
RALAT
Our coverage gate failed, 8 of 20 — rules now decide.
Benchmark cases are descriptions, not ads.
Field labels: agent-drafted, unreviewed.
The phone path is unbenchmarked.
We set ourselves a gate before the benchmark: Tavily had to surface a relevant report for twelve of twenty known cases. It found eight. So we changed the claim: the verdict comes from cited rules, and the web evidence is context. A newspaper prints this in a box called ralat — corrections. We print ours first.
Paste one offer.
awas.edycu.dev
No login, no daily cap — 60 live checks an hour. /judge replays 5 recorded runs.
Built with Nebius Token Factory · NVIDIA Nemotron · Tavily
Paste one offer — any offer — and try to make it lie. If you only have a minute, open /judge: five recorded runs, every quote one click from its source sentence.