How We Inspect

Seven live model calls decide a verdict: five labs evaluating blind, then two masters drawn at random. Below is every stage, including the exact text each model receives.

We publish the prompts because the alternative is asking you to take the method on trust, and a site that asks for trust about truth has already lost the argument. Nothing here is a summary — this is what the models are sent.

How the five numbers become one published score is Station 4, version by version. Every Inspection Report names the scoring model it was reached under and links there.

Station 1 · The panel

All five labs, in parallel. Fired simultaneously so no lab’s latency or failure can shape another’s answer. None of them sees another’s work, and none of them knows who else is on the panel.

The prompt below is printed from the code that sends it, for the current scoring model (v2). It includes the six-band scale, so every lab scores the same ruler the site publishes.

Independently assess this claim: "{claim}".

Provide evidence FOR the claim, evidence AGAINST it, source citations (verified news
outlets, scientific journals, government agencies, fact-checkers like PolitiFact/
Snopes/FactCheck.org) each with a reliability score 0-100, your confidence rating
0-100, and a recommended Baloney Score on the scale below.

THE BALONEY SCALE — what your number will mean when it is published:
1-5 Super Fresh Truth the claim checks out. This is the ONLY band that reads as
TRUE. A statement you consider substantially true belongs
HERE, not at 12 — a small wording caveat is not a reason to
leave this band.
6-20 Fishy Baloney something smells off: missing context, unverified details,
suspicious framing, may be exaggerated.
21-40 Stinky Baloney important facts omitted, misleading implications.
41-60 Rotting Baloney significant portions false, contradicted by evidence.
61-98 Radioactive Baloney overwhelmingly false, potential for harm.
99 Zombie Baloney repeatedly debunked and still circulating.

Score the STATEMENT AS WRITTEN. If it is accurate but phrased slightly more strongly
than the evidence strictly supports, that is a 3-5, not a 12.

1 is true and 100 is false; scores below 1 do not exist. Your number is published
under the band name it falls in, so check it against the bands before you answer:
a claim you have just shown to be false cannot carry a number under 20, and a claim
you have just shown to be true cannot carry a number over 80.

Cite only sources you are confident actually exist. If you cannot recall a real
source for a point, omit the citation and lower your confidence rather than
constructing one. A fabricated citation is worse than no citation.

EVERY SOURCE MUST CARRY ITS URL: the exact page, not the publisher's home page. A
source without a URL is discarded and does not count for you. Every URL is opened
and checked after you answer; a dead link is discarded too.

Return JSON only, matching exactly this shape:
{"evidence_for":["..."],"evidence_against":["..."],
"sources":[{"url":"...","org":"...","type":"news|journal|gov|factcheck|edu|other",
"reliability":0-100}],"confidence":0-100,"recommended_score":1-100}

What the labs received under model v1 (claims inspected before September 2026, until re-inspected)

Independently assess this claim: "{claim}".

Provide evidence FOR the claim, evidence AGAINST it, source citations (verified news
outlets, scientific journals, government agencies, fact-checkers like PolitiFact/
Snopes/FactCheck.org) each with a reliability score 0-100, your confidence rating
0-100, and a recommended Baloney Score (1 = essentially true, 100 = pure
misinformation; scores below 1 do not exist).

Cite only sources you are confident actually exist. If you cannot recall a real
source for a point, omit the citation and lower your confidence rather than
constructing one. A fabricated citation is worse than no citation.

Return JSON only, matching exactly this shape:
{"evidence_for":["..."],"evidence_against":["..."],
"sources":[{"url":"...","org":"...","type":"news|journal|gov|factcheck|edu|other",
"reliability":0-100}],"confidence":0-100,"recommended_score":1-100}

One sentence about the scale and no bands. That is the difference that moved most numbers between v1 and v2 — Model v1, below.

Station 2 · The Primary Master

One lab, drawn at random. It receives all five reports anonymised and shuffled — it cannot tell which lab wrote which, and neither can it defer to a brand instead of to evidence. It writes everything you read on an Inspection Report: the summary, both evidence columns, the ruling on every source, and the correction.

You are the Primary Master Evaluator for this claim: "{claim}".

Five independent evaluators assessed it blind, without seeing each other's work.
Their reports follow in randomised order. They are anonymised deliberately: weigh the
evidence, not the source of the report. Treat all five as equal peers — none is
authoritative, and none should be discounted for disagreeing with the others.

[the five anonymised reports, shuffled]

Score the STATEMENT AS WRITTEN. If it is accurate but phrased slightly more strongly
than the evidence strictly supports, that is a 3-5, not a 12.

You are also writing the published Inspection Report, so produce the reader-facing
content below. Draw ALL of it from the five reports above. Do not introduce evidence
or sources that no evaluator raised, and do not cite anything you cannot vouch for —
this text is published as the verdict of the panel, not of you alone.

The correction it writes has its own rule: state the fact positively, never as a negation of the false claim, in one sentence of about thirty words. A correction that only makes sense to someone who already read the falsehood is no use to the person who meets it cold.

Station 3 · The Blind Master

A second lab, never the Primary. It sees the same five reports in the same order and scores them itself. This prompt never mentions that a Primary Master exists, what it concluded, or that a comparison is coming — the moment it does, the validation stops being blind.

You are evaluating this claim: "{claim}".

Five independent evaluators assessed it blind, without seeing each other's work.
Their reports follow in randomised order and are anonymised: weigh the evidence, not
its source. Treat all five as equal peers.

[the same five anonymised reports]

[the same Baloney Scale]

Produce your own final Baloney Score (1-100) and the reasoning behind it, based
solely on these reports and the evidence they present.

Return JSON only:
{"score":1-100,"reasoning":"..."}

Station 4 · The score — arithmetic, not judgement

No model call happens here. The published score is the median of the counted labs. Not an average: a median is unmoved by one model being wildly out, while an average lets a single outlier drag a verdict.

That arithmetic is the scoring model. It has a version, every Inspection Report names the version it was reached under in the line above the claim, and the current model is v2. When the model changes, published claims are not silently re-numbered: each is re-inspected under the new model and reviewed before the new verdict replaces the old one. The report as it was stays readable, marked as an earlier version, and a card made from it says so when followed.

The scale everyone scores on

Score	Band	Meaning
1–5	Super Fresh Truth	The claim checks out.
6–20	Fishy Baloney	Something smells off…
21–40	Stinky Baloney	That’s not the whole story…
41–60	Rotting Baloney	Most of this claim has spoiled.
61–98	Radioactive Baloney	Dangerously misleading.
99	Zombie Baloney	DEAD. BURIED. STILL SHAMBLING AROUND.

On a Super Fresh Truth claim the page shows the truth percentage instead — a score of 2 is displayed as 98% true — but the stored number is the same number.

What every model has in common

The published number is a median, never a mean. The middle of the lab scores that count.
The masters own the words, not the number. Their scores are published beside the panel median and used to check it: the two masters are measured against the panel median, not against each other, because two summarisers can agree perfectly and both be wrong together.

  gap ≤ 10   confirmed              publishes
  gap 11-20  moderate disagreement  publishes, flagged on the report
  gap > 20   re-evaluate            blocked — a human reviews it first

Model v1 superseded

The model every claim was scored under before September 2026. It is kept unchanged so that what it published can always be reproduced.

The labs were told one sentence about the scale: 1 is essentially true, 100 is pure misinformation. The six bands, their names and their ranges were given to the two masters but never to the labs.
Every lab counted. The published score was the median of all five recommended scores, whatever each lab had written in words.
The zombie flag was recorded but never changed the number. The Primary Master could flag a long public debunk history; the score stayed the median. Publishing Zombie Baloney needed three labs to independently write 99 on a scale that never named it.

Worked example. Five labs score a debunked claim 3, 85, 90, 95 and 99. The 3 came from a lab whose evidence was entirely against the claim at confidence 94 — it read the scale backwards. v1 publishes the median of all five: 90, Radioactive Baloney. The zombie flag is true and does nothing.

What was wrong with it. A lab that thought a claim was “substantially true with a caveat” would reasonably write 12 or 15, which on this scale is Fishy Baloney — it was scoring a ruler it had never seen. And a lab saying “false” in words and “true” in its number counted the same as one that had read the scale correctly.

Model v2 current

Four changes to v1, and nothing else. The panel, the masters, the evidence questions and the report are identical.

1. The labs get the scale. The same six-band guide the masters have always had — names, ranges, “score the statement as written” — goes to every lab word for word (it is the prompt printed under Station 1), with one explicit check: a claim you have just shown to be false cannot carry a number under 20, and a claim you have just shown to be true cannot carry a number over 80.
2. A lab that contradicts itself is not counted. A report whose only evidence is against the claim, with confidence 80 or above, and a score of 20 or below has read the scale backwards. The mirror — only evidence for, confidence 80 or above, score 80 or above — is the same error the other way. That lab’s report is still shown, marked NOT COUNTED with the reason; its number is left out of the median. If fewer than two labs survive, all are counted and the report says so.
3. Zombie is a call. When the Primary Master flags a long, well-documented debunk history spanning years, AND the counted median is 61 or above, AND the Blind Master is 61 or above, the published score is 99. All three are required. The rule can only ever raise a number already at Radioactive.
4. Fresh is a call. The mirror of rule 3. When the Primary Master scores 5 or below, AND the Blind Master scores 5 or below, AND the counted median is 20 or below, the published score is the higher of the two masters’ numbers — inside Super Fresh Truth, and the more cautious of the two. Any gate fails, the median stands. The rule can only ever lower a number already on the true side.

Worked example, the false side. The same five scores: 3, 85, 90, 95, 99. The lab that wrote 3 with all its evidence against the claim at confidence 94 is not counted. The median of the four counted is 92.5. The Primary Master flags a long debunk history and the Blind Master scores 92, so the Zombie rule applies: 99, Zombie Baloney. The report shows all five labs, marks one NOT COUNTED, and prints every step.

Worked example, the true side. Five labs score a true claim 2, 4, 8, 12 and 3. Nothing is excluded; the median is 4. The Primary Master scores 3 and the Blind Master 4, so the Fresh rule applies and the published score is the higher master, 4, Super Fresh Truth — shown as 96% true. Had the Blind Master written 9, the rule would not apply and the median of 4 would stand on its own.

Why the numbers moved. Claims re-inspected under v2 often land in a different band from their v1 report. That is mostly rule 1: the labs are now scoring the ruler the site publishes, so their numbers cluster in named bands instead of a vague middle. It is a calibration change, not a change in what the labs think about the claim.

Reading a report

The line above the claim names the model: “Scoring model v2” links to the section above.
“How the number was reached” lists every decision the model made on that claim, including any lab not counted and any rule applied.
If the claim has been re-inspected, the same line offers “Earlier report, model v1”: the report as it was, marked as superseded and pointing back at the live one.

Station 5 · The correction, chosen before it is paid for

When a claim scores 95 or above, the truth it displaces deserves a page of its own. Several wordings are drafted — each by a different lab, all reading the same anonymised evidence — and each is then scored by a single model that did not write it. A model grading its own sentence is grading its own homework.

A person picks one, and only then does a second full panel run. That page is a real inspection: five models evaluating that statement, which never happened while they were evaluating the false one.

What we refuse to do

Scores are never typed in by hand. There is no field anywhere in our admin panel that can set a score or a rating band. A verdict can only be changed by running the panel again, and when that happens the revision is recorded. Wording and publication status can be edited; the number cannot.

A correction that does not verify is not published. If the panel scores a proposed correction anything other than Super Fresh, nothing goes live and no pair is created — the panel is telling us the sentence is wrong, and a human rewords it.

Browse the Truth Deli →Our Mission →