Skip to content

Methodology · karat-0.1

How we judge, in full.

A verifier should be verifiable. Here is every rule ORO applies, and the exact formula behind each Karat.

Karat calculator · runs the production formula

Supported claims2
Conflicting claims0
Unverified claims0
Hedged or unattributed0
Medium scam signals0
Severe scam signals0
15
Karat / 24
Low risk

12 + 6 − 3 − 0 − 0

base      = 12
evidence  = +3 per supported, −5 per conflicting (bounded ±9), × tier weight (1, 1, 0.7, 0.4)
source    = +3 filing or regulator origin, 0 tier-1 news, −3 unknown or anonymous
language  = −1 per hedge or missing attribution (max −3)
scam      = −2 per medium signal, −6 per severe signal
karat     = clamp(round(sum), 0, 24); 24 only if every checkable claim is supported by tier 1–2

risk band: High if any severe signal or a conflict on a material claim;
           Elevated if medium signals or most claims unverified; otherwise Low.
           Bands describe the information, never the stock.

1 · Personal details go first

Phone numbers, PAN, Aadhaar, emails, bank accounts and UPI ids are removed before anything is stored or sent to a model.

Screenshots are processed and discarded. Tip text is kept only if you opt in; otherwise ORO keeps a fingerprint for Forward Trail.

2 · Scam rules run before any AI

Guaranteed returns, urgency, secrecy, fake registration, payment requests, look-alike links, .apk files, fake authority, surveillance lists, SME/low float and unusual volume.

Each rule is severe or medium. Results stream to you in under two seconds.

3 · Claims are judged one by one

Each factual claim is matched against BSE/NSE filings, SEBI records and tier-1 news using keyword and meaning search, filtered by company and weighted by recency.

A claim is Supported or Conflicting only with a quoted line from a source. Unverified is a normal, honest answer.

4 · Code checks every quote

The quote must exist, word for word, in the cited document. Every number in a Supported claim must appear in the quote. If not, the claim falls back to Unverified.

Tip text can't instruct the system: it is treated as data, the judge has no tools, and output is schema-only.

5 · Advice is filtered out

Every sentence ORO writes passes a filter for buy/sell/hold language, targets, stop-losses and their Hindi and Hinglish forms. A match blocks the sentence.

ORO answers “is this true?”. It never answers “should I?”.

6 · Mistakes are corrected in public

Anyone can report a wrong verdict. A person reviews it; if it was wrong, the verdict changes, the correction is logged publicly, and everyone who saw it is told.