01
Frame the claim
One narrow, testable proposition—not a mood, label, or bundle of allegations.
Use a relevant base rate or documented prior knowledge when possible. Avoid 0% and 100%: both make updating impossible.
Reality before allegiance
Do not buy the narrative. Extract the claims and test them.
Build a falsifiable hypothesis, audit the source chain, record evidence on both sides, and update confidence without pretending uncertainty has disappeared.
02
Audit the information source
Ownership and incentives inform reliability; they do not settle whether a specific claim is true.
?
No source audit yet. Record the publisher, ownership, funding, relevant interests, and the original source behind the item.
03
Build the evidence ledger
Add evidence that supports, challenges, or fails to discriminate between H and not-H.
HypothesisNot set
AlternativeNot set
∅
The ledger is empty. Search for evidence that could prove you wrong—not merely material that resembles your theory.
04
Write the calibrated finding
Separate established facts, inference, plausible risk, and unknowns.
Field guide
A repeatable way to vet information without choosing a team.
The object is not perfect certainty. It is a working model whose confidence matches the available evidence.
MOVE 01 / ATOMIZE
Extract the claim
Strip away the music, adjectives, montage, moral framing, and implied conclusion. Rewrite the material as small propositions that can each be checked.
- Replace “secret plot” with the specific alleged action.
- Separate existence, capability, opportunity, action, and intent.
- Record dates, names, documents, quantities, and direct quotations.
MOVE 02 / ORIENT
Read laterally
Before investing in the page, leave it. Learn what independent sources say about the publisher, author, funder, owner, and cited evidence.
- Map ownership, investments, political ties, and business incentives.
- Look for patterns across the outlet’s output.
- Do not confuse a polished page with a reliable institution.
MOVE 03 / TRACE
Follow the dependency chain
Ten articles may reduce to one press release, anonymous briefing, wire story, preprint, or clipped video. Count independent origins—not URLs.
- Find the primary document, dataset, recording, filing, judgment, or study.
- Compare the quotation with its original context.
- Ask what each new outlet actually verified or added.
- Use trade publications and domain experts to test whether the mechanism works in practice.
MOVE 04 / COMPETE
Generate alternatives
A fact does not strongly support H when several ordinary explanations predict the same fact. Write the strongest alternative before scoring evidence.
- What mundane cause fits?
- What rival actor or process fits?
- What observation would distinguish them?
MOVE 05 / ADVERSARIAL TEST
Try to be wrong
Search deliberately for disconfirming evidence. Steelman the opposing explanation instead of attacking its weakest version.
- What should exist if H is true?
- What should not exist if H is true?
- What evidence would force a meaningful update?
The stopping rule is not “I feel vindicated.” It is “I can state the probability, the uncertainty, and what would change my mind.”
Evidence Ledger principle
MOVE 06 / CALIBRATE
Update, do not flip
Evidence can move confidence up, down, or barely at all. Strong claims need evidence that is much more expected if the claim is true than if it is false.
MOVE 07 / STOP
Match the threshold
You may need only moderate confidence to investigate or add safeguards, but much higher confidence to accuse, publish, punish, or act irreversibly.
Method and sources
Bayesian in spirit, explicit about its limits.
The workbench converts qualitative judgments into an auditable update. The number is a disciplined estimate—not a machine that manufactures truth.
The core update
Bayesian reasoning begins with a prior probability and updates it when new evidence arrives. The evidentiary question is comparative: how expected would this observation be if the hypothesis were true, and how expected would it be if the alternative were true?
posterior odds = prior odds × [ P(E | H) ÷ P(E | not-H) ]
A likelihood ratio above 1 supports H; below 1 challenges H; exactly 1 does not discriminate. The app asks for qualitative likelihoods, then discounts the update according to your rating of source reliability and independence.
Why the discount exists
Real-world evidence is rarely clean. Reports repeat one another, witnesses share information, and uncertainty in the evidence itself can be substantial. The app therefore applies a transparent heuristic: it shrinks the log-likelihood update when reliability is low or the item is derivative.
effective LR = exp[ ln(raw LR) × reliability × independence ]
This adjustment is not canonical Bayes. It is a safeguard against false precision. If you have validated likelihood ratios from a technical field, use a proper statistical model instead.
What the number cannot do
It cannot repair fabricated inputs, determine whether your hypotheses are exhaustive, detect hidden dependence, or replace domain expertise. The ledger is most useful because it exposes your assumptions and makes revisions visible.
A disciplined conclusion
Report facts and inference separately. A good endpoint may read: “The mechanism is technically plausible and opportunity is established, but public evidence does not establish action or intent. Confidence remains moderate, and these findings would change it.”