System flow · rev. 2026.09 · illustrative run

Follow one claim through the machine.

Seven stations, from a pasted post to a printed receipt. Scroll to watch a real-world example go through every rule, pass and fallback.

Today's example

“Studies show people who skip breakfast weigh more — so eat breakfast to drop pounds.”

Illustrative run · sources shown as tiers, not real outlets

01Intake

Paste anything. Get claims.

A single claim goes straight to checking. Long text, links, images and video transcripts are dissected by Claude Haiku into individual, checkable claims — prioritized, then you pick which ones run. If Claude is unavailable, a backup model does the extraction.

SINGLE CLAIMSTRAIGHT TO CHECKING
TEXT · URL · IMAGE · VIDEOUP TO 20 CLAIMS
EXTRACTED BYCLAUDE HAIKU

In this runOne post, two claims hiding in it: a statistic, and a promise about cause.

01 / 07 · Intake
Pasted · 1 postLong text

“Studies show people who skip breakfast weigh more — so eat breakfast to drop pounds.”

Claude Haiku extraction · 2 claims found

Claim 01 · statistic

People who skip breakfast weigh more.

Claim 02 · cause

Eating breakfast makes you lose weight.

1 claim → straight to queue · long / media → up to 20 · you pick

02Queue

Six lanes. No waiting room.

Claims run six at a time. Each one is searched, weighed and judged on its own — a strong claim can't lend its sources to a weak neighbour.

CHECKED IN PARALLEL6 AT A TIME
BATCH LIMIT20 CLAIMS

In this runBoth tickets take a lane and are checked side by side, each on its own evidence.

02 / 07 · Queue
Check queue6 lanes
L1
CLAIM 01Checking
L2
CLAIM 02Checking
L3
open
L4
open
L5
open
L6
open

Up to 20 claims per batch
the rest wait their turn

6 at a time

03Search

Search until the record is wide enough.

Two focused queries fire in parallel — ten results each, one per domain — aiming for ten unique domains. Thin coverage triggers a broad research sweep that needs at least eight; still thin, it goes to government, education and published reports.

If the search index is unavailable or comes back empty, the check falls through a side track: a backup web search, then AI search grounding, where the model searches the live web itself and only the pages it actually retrieved become sources.

SEARCH PASS 12 × 10 · 1 PER DOMAIN
PASS 2 · IF < 1010 GENERAL + 8 RESEARCH
PASS 3 · IF < 88 GOV/EDU + 8 REPORTS

In this runPass 1 reaches 6 unique domains, pass 2 only 7, so pass 3 goes to official and academic sources — 13 domains in all.

03 / 07 · Search
Claim 01 · live webup to 3 passes

Unique domains

810
13 unique domains
P1

Two focused queries, in parallel

10 results each · 1 per domain

6 domains < 10 → broaden

P2

Broad research sweep

10 general + 8 research

7 domains < 8 → go official

P3

Official + academic

8 gov / edu + 8 reports

Index down or 0 results? backup web search → AI search grounding · not needed

04Source gate

Weed. Weigh. Limit.

Duplicate domains merge and social posts are blocked. Every source is scored by its domain, never its headline. Search-index candidates are ranked by that score; with six or more strong sources the low tail is dropped. AI-grounded sources keep their retrieval order so their citations stay mapped. At most twelve go to the verdict.

FACT-CHECKERS · GOVERNMENT95
ACADEMIC · MEDICAL92
WIRES · MAJOR NEWS85
REFERENCE · GENERAL WEB75 · 70
SELF-PUBLISHED50

In this runSocial is blocked, the self-published tail is dropped, and 11 ranked sources go forward.

04 / 07 · Source gate
Source gate · 13 inRank by primacy
Fact-checker ×295
Government health agency ×295
Peer-reviewed / medical ×492
Wire service85
Major news85
Reference75
General web70
Self-published50

≥6 strong sources → drop the low tail

Social postblocked—
Duplicate domainsmerged—

To the verdict
max 12, strongest first

11 pass

05Verdict

No support is not the same as false.

Claude reads the ranked, numbered sources and writes the verdict on an eight-point scale, citing them inline. It never searches on its own: every source comes from the gate. FALSE needs a source that actively contradicts the claim; when the record is simply silent, the answer is UNVERIFIED. If Claude is unavailable, a backup model rules on the same evidence; with no live sources, the receipt says the verdict is preliminary and caps its confidence.

Unverified or low confidence? Subpoena Mode is an optional, deeper AI review with a more capable model — it reruns retrieval, ranking and the verdict, then settles the question or preserves the uncertainty.

WRITTEN BYCLAUDE · FROM RANKED SOURCES
VERDICT SCALE8 POINTS
FALSE REQUIRESPOSITIVE CONTRADICTION
NO SUPPORTUNVERIFIED

In this runThe statistic holds up: TRUE. The promise doesn't follow from it: MISLEADING.

05 / 07 · Verdict
Verdict scale8 points
  1. ■TRUE · positive evidence confirmsClaim 01
  2. ✓MOSTLY TRUE · core right, a detail off
  3. ✓MIXTURE · true and false in balance
  4. ■MISLEADING · false impressionClaim 02
  5. ✓MOSTLY FALSE · wrong, kernel of truth
  6. ✓FALSE · needs positive contradiction
  7. ✓UNVERIFIED · no support ≠ false
  8. ✓NOT CHECKABLE · opinion or prediction
TRUE ✓MISLEADING ≈

↺ Subpoena mode optional: unverified or low confidence → deeper AI review reruns retrieval, ranking, verdict · not needed

06Framing

Then we grade the spin.

When a post strings claims together, a second pass reads how they're framed: false cause, cherry-picking, missing context and loaded language — the techniques that make true facts carry a false conclusion.

RUNS ONMULTI-CLAIM CONTENT
TECHNIQUES GRADEDFALSE CAUSE +3 MORE

In this runEvery fact can check out and the post can still mislead: correlation wearing a lab coat.

06 / 07 · Framing
Framing passMulti-claim only

Flagged: false cause. People already dieting skip meals — the correlation is real, the cause is flipped.

07Receipt

Print what survives.

The verdict, its confidence and the sources are saved to a permanent public page, with a share card and a QR code so the receipt travels.

VERDICT · CONFIDENCEPRINTED
SOURCESUP TO 12
LINK · SHARE CARD · QRPERMANENT

In this runOne link that settles the chat — and anyone can audit every source behind it.

07 / 07 · Receipt
Mebro · fact checkIllustrative

“Studies show people who skip breakfast weigh more — so eat breakfast to drop pounds.”

MISLEADING ≈Confidence
High
SOURCES11 / ≤12
FRAMINGFALSE CAUSE
  • ✓ Permanent public link
  • ✓ Share card
  • ✓ QR code
  1. Accept a single claim, or have Claude Haiku dissect longer text, URLs, images, and video transcripts into individual claims; if Claude is unavailable, a backup model does the extraction.
  2. Place up to twenty claims into a six-at-a-time checking queue.
  3. Search the live web: through the search index, run focused queries targeting ten unique domains and broaden into research and official sources when coverage is thin; if the index is unavailable, fall back to backup search, then AI search grounding.
  4. Deduplicate domains, block social posts, rank by source primacy, and pass at most twelve candidates.
  5. Claude writes the verdict from the ranked sources on the eight-point rubric, requiring contradictory evidence before FALSE; if Claude is unavailable, a backup model rules on the same evidence.
  6. Escalate unresolved or low-confidence checks to an optional deeper AI review.
  7. For multi-claim content, separately inspect the framing and manipulation techniques.
  8. Save the verdict and sources, then print a permanent public receipt with a share card and QR code.

The whole line

Paste to receipt, at a glance.

  1. 01

    Intake

    1 claim, or up to 20 extracted

  2. 02

    Queue

    6 claims at a time

  3. 03

    Search

    Up to 3 search passes · target 10 domains

    ↳ Index down or 0 results → backup search → AI grounding

  4. 04

    Gate

    Dedupe · block social · ≤12 ranked

  5. 05

    Verdict

    8-point scale · FALSE needs contradiction

    ↳ Unverified / low confidence → Subpoena Mode (optional)

  6. 06

    Framing

    Multi-claim only · spin graded

  7. 07

    Receipt

    Permanent link · share card · QR

Mebro · fine print

Search results are evidence candidates. The diagram does not claim that every full document is fetched or that every cited sentence receives an independent deterministic support audit. The run above is illustrative: its counts show how the rules apply, and its sources are tier labels, not real outlets.