UnFake
Knesset elections  ·  October 27, 2026  ·  54 days out

Keeping everyone
honest.

This election started dirty. Misinformation is being manufactured and aimed at politicians, journalists and public figures, and which way it flies depends entirely on where the target sits politically.

UnFake scores that content. Two independent engines, one for the facts and one for the reasoning, applied to the left and the right on identical terms. The results are published whether or not we like how they land.

A referee that takes a side is worth nothing. So we built one that cannot.

Two posts, one week, opposite poles
The left
Yair Golan
Facts
52
Reasoning
31
The right
Itamar Ben Gvir
Facts
42
Reasoning
55

Checked the same week, by the same engines, in Hebrew. Neither came out clean. Both results are public and permanent.

The problem

In an election, "false" is a weapon. That is exactly why it cannot be the output.

Every campaign already calls the other side liars. A tool that hands down a one-word verdict just adds a louder voice to that fight, and gets adopted by whichever camp its verdicts happen to favor.

Content fails in two independent ways, and they need separating. It can assert things that are not so. It can also reason badly from things that are perfectly true. Collapse those into a single number and you lose the only distinction a voter actually needs.

REASONING QUALITY HIGH LOW Sound throughout True claims, and the conclusions actually follow from them. Clean logic, false premises Impeccably argued from things that are not true. ERASED BY A SINGLE SCORE True facts, broken argument Every claim checks out; the conclusion does not follow. ERASED BY A SINGLE SCORE Wrong on both counts False claims, and reasoning that would fail anyway. FACTUALLY ACCURATE FACTUALLY WRONG

Campaign season lives in the two amber cells. A politician who says only true things and draws an unsupported conclusion from them is doing something a "false" label cannot describe.

The program

What we are doing between now and October 27

Score political content as it circulates. Publish every result as a permanent, citable page. Where a post scores badly, reply to it with the score and a link to the full working, in public, where the post's own audience will see it.

01

Select posts in circulation

Content from public figures across the political map, weighted toward what is actually spreading. The selection criteria will keep changing as the campaign does, and deliberately so: the point is to follow the misinformation, not to run a fixed list.

02

Run both engines

Text, or a transcript where the post is video. Facts checked against live, cited sources; reasoning checked for the leaps. Each result records the exact post it was a check of, so nothing can drift onto the wrong target.

03

Publish the full working

Every check gets its own permanent URL on unfake.io, carrying the claim-by-claim verdicts, the confidence levels, the sources consulted, the reasoning flags, and what was deliberately set aside as opinion. Indexable, structured for search engines, and open to dispute.

04

Reply where the post lives

A page nobody visits changes nothing. When a post scores below the threshold, the result goes back as a public reply on the post itself, carrying the two scores as an image and a link to the working. The correction reaches the audience that saw the claim.

What actually gets posted
UF
UnFake replying to @YairGolan1

אימתנו את הטענות בפוסט שלך. הציון שקיבלת הוא:

UnFake שני מנועים עצמאיים. שני ציונים. לעולם לא משוקללים יחד. 52 עובדות 31 היגיון unfake.io

לתוצאות המלאות:

unfake.io/u/636bec01f3cc8e7d

Card renders white in both themes because it is an uploaded image.

Note what it does not say

There is no accusation in it. No "false", no "misleading", no adjective at all. Two numbers, and a link to the reasoning behind them.

That is a deliberate constraint. The moment a reply characterizes the post, it becomes another campaign voice and gets treated as one. Handing over a score and the full working, and letting the reader draw the conclusion, is both more defensible and harder to dismiss.

The Hebrew reads: "We verified the claims in your post. The score you received is:", then the card, then "For the full results:" and the link.

Guardrails

Why this is not a brigading machine

Anything that replies to politicians at scale can become harassment by accident. Four constraints stop that, and they are enforced in code rather than left to judgment.

A threshold, not an opinion

A reply is only ever drafted when an engine scores below 35, the bottom band, where the reasoning fails outright or reliable sources contradict the claims. Mediocre content is left alone.

No claim, no reply

If the fact engine found nothing checkable, the post is excluded even when its reasoning score is terrible. The reply opens by saying we verified the claims. Saying that about a post with no claims would itself be a false statement.

Reviewed before it posts

Nothing publishes automatically. Each qualifying result is checked against the exact text and image that will go out before it is released.

Tied to the post it checked

Each result stores the address of the content it analyzed, and the reply target is derived from that record rather than typed in. A reply cannot be aimed at a post that was never checked.

Engine 1

How it decides what counts as evidence

Source selection is where a fact-check earns or loses its authority. In a campaign, where every outlet is accused of bias by someone, the rules have to be stated and held to.

First, work out what to search for

Searching a verbatim snippet fails in a specific and damaging way: a Hebrew post about an internationally covered subject retrieves nothing, because the confirming sources are written in English. The engine then reports the central claim as unconfirmed, which looks like a failure of the tool rather than of the search.

So a fast pre-pass extracts the claims a reader would walk away believing and writes two to four keyword queries, one per claim, in the language reliable sources on that claim are actually written in, with names in their original spelling. Results merge round-robin so every claim gets evidence.

Then weigh what comes back

Primary over secondhand

Official records, datasets, transcripts, court filings and original footage outrank anyone describing them. Outlets need editorial standards, named authors and a visible corrections policy.

Syndication is one source, not five

Twenty outlets running the same wire story is a single data point. The engine collapses them and looks for reporting that confirms a claim through genuinely different access.

A single source caps confidence at medium

However good it is. High confidence requires corroboration by more than one independent reliable source.

Contested claims get cross-spectrum corroboration

Every outlet carries slant, and in this election that cuts both ways. Where a claim is contested, the engine reports the disagreement instead of picking a side.

No sources means unverified, never false

If retrieval finds nothing, the claim is unverified at low confidence. Inventing a source or a correction is the worst error available to this engine, and it is told so explicitly.

Seven verdicts, kept distinct

Collapsing these into true-or-false is what makes automated fact-checking easy to dismiss. "Nobody has confirmed this" is not "this is a lie," and a campaign will seize on the difference.

trueReliable independent sources confirm it as stated.
falseReliable sources contradict it. It did not happen this way.
partially_trueRight event, wrong number or date. Correct in some respects.
misleadingTechnically accurate, missing the context that changes the impression.
outdatedTrue at an earlier time, no longer accurate now.
disputedReliable sources genuinely conflict. The matter is not settled.
unverifiedNo reliable source addresses it either way. Not a finding of falsehood.

A rubric, so the number means the same thing twice

85–100accurate

Load-bearing claims check out against reliable independent sources.

65–84mostly accurate

The core is accurate. Minor inaccuracies that do not change the picture.

35–64mixed

At least one material claim is false, misleading or unsupported.

0–34inaccurate

The central claims are contradicted by reliable sources. Reply threshold.

nullunverifiable

Could not be checked at all. No score is returned, and no reply is drafted.

Engine 2

The critical thinking it applies

This engine never touches the facts. It asks one question: do the conclusions follow from what is actually offered? Most campaign content is not fabricated. It is true material carrying a conclusion it cannot support, and that is precisely what this engine is for.

01Charity first

Reconstruct the strongest reasonable version of the argument before criticizing it. If a charitable reading removes the flaw, the flaw was the analyst's, not the author's. Especially important when the author is someone you did not vote for.

02Calibration always

Name a problem only when it genuinely fits and actually weakens the case. A tool that finds four fallacies in every paragraph is noise, and in a campaign it is worse than noise: it is ammunition. At most three flags are ever shown, and sound reasoning is called sound.

What it looks for

Unstated assumptions

Arguments missing a needed step. The engine makes the hidden premise explicit, then tests it.

Cause and effect, under scrutiny

Could causation run the other way? Could a third factor cause both? Could it be coincidence? Was the base rate ignored? The staple move of campaign argument is treating sequence as cause.

Scope and overreach

Does the conclusion claim more than the premises license? Named as the most common and least noticed error, and the central one here.

Generalization and analogy

Is the sample large and unbiased enough to carry a claim about a whole population? Are the compared things alike in relevant ways?

Four families of error

Formal fallaciesThe structure looks valid but is not. Affirming the consequent, denying the antecedent.
Informal fallaciesAd hominem, straw man, false dilemma, appeal to ignorance, slippery slope, begging the question.
Bias and manipulationConfirmation bias, availability heuristic, framing, anchoring, loaded language.
Misrepresenting the sourceClaiming the content says or proves something it does not.

Each flag comes back twice: once in everyday language, once with its technical name. Both in the content's own language, so a Hebrew post gets Hebrew findings.

The same rubric discipline

85–100sound

The conclusion follows and is well supported.

65–84mostly sound

Broadly holds. Minor leaps or small gaps.

35–64shaky

At least one significant flaw. Weakly supported as argued.

0–34unsound

The central reasoning fails. Reply threshold.

The handoff. When the reasoning engine meets something whose truth it cannot judge, it does not guess. It writes the claim into a dedicated field for the fact engine and moves on. Neither engine is permitted to do the other's job.

Both sides, on the record

The two results behind the numbers at the top

Checked on September 3, 2026, in Hebrew, from posts on X. Both are public, permanent pages. Nothing below is illustrative. It is the engine output verbatim.

The right  ·  Otzma Yehudit
Itamar Ben Gvir
unfake.io/u/964bfb160a89f16c →
Facts
42
Mixed
Reasoning
55
Shaky
ENGINE 1Facts 42

The central numeric claim drove the score: roughly ten assassination attempts, where the sources consulted describe at least three over three years.

עיתונאי פרסם מקומות שבהם שהה ובילה יאיר נתניהו בזמן מלחמה.

truehigh confidence

ישראל היום, July 23, 2026  ·  מעריב

בן גביר הוא אחת האישיויות המאוימות ביותר בישראל, לאחר כ-10 ניסיונות חיסול.

partially truemedium confidence

מעריב  ·  מעריב, November 2024

בן גביר מוביל מדיניות קשוחה בבתי הכלא, בהריסת מבנים בלתי חוקיים ובקו נחרץ מול האויב.

unverifiedlow confidence

No source consulted addressed it. Not scored as false.

6 further statements were set aside as opinion or rhetoric, and listed openly rather than quietly graded.

ENGINE 2Reasoning 55

הטיעון מעלה חשש לגיטימי לגבי עיתונות שמסכנת חיים, אך מקפץ למסקנות רחבות על בסיס הנחות שאינן מוכחות.

majorunsupported leap

מסקנה שאומרת הרבה יותר ממה שהתוכן מוכיח

majorunwarranted assumption about intent

ייחוס כוונות זדון ללא הוכחה

minorhasty generalization

הצגת ניסיון אישי כהוכחה לתופעה רחבה

Note what the engine credits: the underlying concern about journalism that endangers lives is called legitimate. The score reflects the leap, not the politics. Neither engine crossed the reply threshold, so no comment was drafted.

The left  ·  The Democrats
Yair Golan
unfake.io/u/636bec01f3cc8e7d →
Facts
52
Mixed
Reasoning
31
Unsound
ENGINE 1Facts 52

Two load-bearing claims verified true at high confidence, against named international outlets and an Israeli research institute.

ראש הממשלה היה בניתוק מהרמטכ"ל בזמן שמחבלי חמאס כבשו יישובים ורצחו אזרחים ב-7 באוקטובר.

truehigh confidence

The Jerusalem Post, February 12, 2026  ·  Yahoo News

לא הוקמה ועדת חקירה ממלכתית לחקר אירועי 7 באוקטובר.

truehigh confidence

NPR, March 5, 2025  ·  המכון הישראלי לדמוקרטיה, April 23, 2026

נתניהו זלזל בהתרעות מערכת הביטחון לפני ה-7 באוקטובר.

disputedmedium confidence

NPR  ·  N12. Reliable sources genuinely conflict

נתניהו ניהל כבר שלוש שנים קמפיין כדי להאשים את מי שיצאו להילחם על הבית.

partially truelow confidence

Roya News HE  ·  N12

ENGINE 2Reasoning 31

טקסט פוליטי תוקפני שמציג מסקנות חמורות כעובדות מוכחות, מבלי לספק ראיות לטענות המרכזיות שלו.

majorunsupported leap / assertion as fact

המסקנה אומרת הרבה יותר ממה שהתוכן מראה

majorad hominem / poisoning the well

תקיפת האדם במקום הטיעון

minorfalse dilemma

הצגת שתי אפשרויות בלבד כשיש יותר

This is the whole argument in one specimen. The facts largely hold, two at high confidence with citations. The argument built on them does not, and at 31 it crosses the reply threshold. A blended score would have landed near 40 and told a voter nothing about which half was broken.

The balance rule
Yair Golan 52 facts · 31 reasoning Itamar Ben Gvir 42 facts · 55 reasoning LEFT RIGHT

The first question anyone asks is whose side this is on. The answer has to be demonstrated rather than asserted, which means checking both poles on identical terms and publishing whatever comes back. The engines are instructed to judge the structure of the reasoning, not the politics or the person, and the operators do not get to withhold a result they dislike.

Standing

Built, deployed, and running now

2independent engines, never combined
6→24live sources per section
<35score that triggers a drafted reply
7claim verdicts, kept distinct
~$0.09model cost per section analyzed
54days until the vote
Shipped

The web app

Text, links, YouTube, and uploaded audio or video, in any language the model reads.

Permanent, citable results

Every check gets an indexable URL with ClaimReview structured data and a sitemap entry, recording the post it was a check of.

The reply queue

Qualifying results become drafts carrying the exact text and card, released once reviewed.

Moderation and dispute handling

A safety gate holds sensitive content from the public feed, and every published result carries a dispute route.

What makes scale affordable

Content-hash caching

Re-checking identical content is free to serve. Viral content, checked by many people, converges on one cached analysis. The cost curve improves exactly where the volume is.

Model routing by complexity

Short, simple posts stay on the cheaper model. Long or number-dense content is routed up.

Claim-driven retrieval

Two to four targeted queries rather than broad crawling. Cheaper, and better evidence.

The replies are the distribution

Each one puts a permanent unfake.io page in front of an audience already engaged with the claim, at no media cost.

Everyone gets the same two engines

The value of a referee is that it does not care who is playing. Every result is published with its sources, its confidence levels, and the statements it deliberately refused to grade, and every one is open to dispute. Between now and October 27, that is the whole proposition.