The reader’s version
We start with a defined question, examine public evidence and test plausible explanations. We make judgements where the reasoning supports them, state the important uncertainties and explain what would change our view. Every principal judgement should be traceable to its evidence.
Likelihood is how probable we judge a proposition to be. Confidence is how strong its evidentiary and analytical foundation is. They answer different questions. Neither is a guarantee.
Start with a proposition
Define the actor, action or outcome, relevant place and period, and what would count as an answer. Separate questions about whether something happened, how it happened, why it happened and what may happen next. Evidence of an effect does not, by itself, establish a motive.
A broad question can yield several judgements. Each needs its own scope, evidence and uncertainty. An assessment does not have a single confidence rating that automatically applies to every claim.
Keep the types of claim distinct
If needed, scroll the table sideways. Keyboard: focus the table, then use the arrow keys.
| Type | What the reader should be able to tell |
|---|---|
| Established fact | What is documented or independently corroborated, by which sources and with what limits. |
| Reported claim or allegation | Who says it, how they could know, and whether it has been verified. |
| Observation or indicator | What was observed and why it might matter; an indicator is not proof of the explanation. |
| Assumption | What is being treated as true to bridge a gap, and what fails if it is false. |
| Inference | The reasoning from the evidence to an interpretation, including plausible alternatives. |
| Estimative judgement | The specific proposition assessed, its likelihood and the foundation for that assessment. |
| Gap or limitation | What is missing, inaccessible, contradictory or outside the research scope. |
Evaluate sources and evidence
Consider authenticity, access, expertise, reliability, incentives, timeliness, context and corroboration. Evaluate the specific claim as well as the publisher: a reliable outlet can repeat a weak claim; a partisan source can provide a verifiable document.
Trace reports to their origin. Several articles quoting the same witness, press release or dataset form one source chain, not several independent confirmations. Keep event dates separate from publication dates. Record contradictory material, missing context and limits of translation or access.
An evidence register links stable evidence identifiers to source identifiers, claim type, provenance, limitations and the judgements affected. A source list alone is not a reasoning trail. For measured claims, explain the population, denominator, sampling period, collection method and classification error. Automated activity, coordinated behaviour and state attribution are separate propositions.
Express likelihood consistently
Use the following ordered vocabulary. These are qualitative judgements, not measured frequencies or a numerical scoring system.
If needed, scroll the table sideways. Keyboard: focus the table, then use the arrow keys.
| Term | Reading |
|---|---|
| Almost no chance | A remote possibility. |
| Very unlikely | A strong expectation against. |
| Unlikely | The proposition is less probable than its alternative. |
| Roughly even chance | No clear lean toward or against; an assessment, not a synonym for ignorance. |
| Likely | The proposition is more probable than its alternative. |
| Very likely | A strong expectation in favour. |
| Almost certain | A remote possibility of being wrong. |
Do not infer exact percentages from the spacing or presentation of these terms. Numerical estimates belong only where a defensible method or explicit, justified estimation process supports them. They must state the proposition, period, assumptions and limits.
Use indeterminate or not assessable when the available evidence and reasoning cannot support a likelihood judgement, and explain why. A weakly supported allegation is not automatically unlikely. Lack of proof is not proof of absence; absence becomes informative only when there was a credible opportunity to detect what would be expected.
Explain confidence for each judgement
Use only High confidence, Moderate confidence or Low confidence. Give the specific judgement first, then its confidence and rationale separately. Consider source quality, relevant coverage, independence, corroboration, inferential strength, unresolved alternatives and consequential assumptions.
If needed, scroll the table sideways. Keyboard: focus the table, then use the arrow keys.
| Rating | Foundation |
|---|---|
| High confidence | Strong, relevant and sufficiently independent evidence; reasoning has survived serious alternative explanations and assumption checks. Remaining gaps are unlikely to overturn the judgement. |
| Moderate confidence | Relevant evidence and coherent reasoning support the judgement, but material gaps, dependence or plausible alternatives remain. |
| Low confidence | A substantive judgement is possible, but its foundation is limited, indirect, conflicting or dependent on consequential assumptions. |
These descriptions guide a reasoned evaluation; they are not additive scores. Source quantity alone cannot establish confidence.
“We cannot establish whether the event occurred” is a statement about the evidence available. Do not append “Low confidence” as though a likelihood judgement had been made. If an evidence-coverage conclusion is itself assessed, scope and justify it separately; it does not establish the event’s probability.
Test explanations, not just supporting evidence
For every assessment, identify the strongest plausible alternative and the evidence that would weaken the preferred account. Use structured analytical techniques when they help answer the question:
- Key Assumptions Check: expose consequential assumptions, test them and explain what changes if they fail.
- Analysis of Competing Hypotheses (ACH): compare explanations against the same evidence, especially evidence that discriminates between them.
- Indicators and Signposts: specify observable developments, expected timing and what they would strengthen or weaken.
- Alternative Futures: explore consequential scenarios when the question concerns divergent paths; a scenario is not a forecast merely because it has been described.
- Challenge review: argue the strongest alternative, look for missing evidence and ask how the preferred explanation could be wrong.
For complex investigations: the complete ACH process
- Build a broad inventory of relevant events, facts, claims, context and potential evidence before narrowing the argument. Retain exclusions and their reasons. Breadth helps expose omissions; a large inventory is not evidence of quality by itself.
- Evaluate each item’s provenance, access, reliability, limits and source-chain dependence. Separate evidence of an event from evidence of its cause or intent.
- Define plausible hypotheses before favouring one. Include a mixed explanation when a mechanism and motive could coexist. State what distinguishes it; an elastic hypothesis that explains every outcome is not a useful test.
- Compare every included evidence item with every hypothesis. Use C (consistent), I (inconsistent), N (not diagnostic here) and U (uncertain relevance or interpretation), with written reasons. C means compatible, not proved. U is not a blank or an inconsistency. This is WSC’s working notation, not a universal ACH code.
- Examine the most diagnostic contradictions and their reliability. Do not count matrix cells as independent votes or convert them to probabilities. Common-source rows and repeated descriptions of one event must not accumulate artificial weight.
- Rank from least to most inconsistent when the evidence discriminates. Explain decisive evidence and assumptions. Allow tied tiers or a partial ranking. Least inconsistent is a comparative result, not an automatic absolute probability.
- Test sensitivity: remove a key source chain, weaken a disputed claim, recode an ambiguous cell, and change a consequential assumption. Record whether the ranking and final judgement survive.
- Give the final assessment using calibrated likelihood and judgement-specific confidence. State unresolved alternatives, gaps and the evidence that would change the result. Preserve the complete matrix and process when they materially support the conclusions.
Incomplete information is normal. It does not prohibit judgements. It also does not justify an ordering that the evidence cannot sustain.
Make the reasoning inspectable
Every assessment should provide a precise question, evidence cutoff, principal judgements, important supporting and contradicting evidence, reasoning, significant assumptions, alternatives, gaps, review indicators, inspectable references and revision history.
For complex investigations, add a complete public evidence and source register, full hypothesis definitions and matrix, cell rationales, rankings and sensitivity results. Keep the broad inventory and excluded items available in an annex when material to the audit. Explain withholding or redaction without exposing private research or personal information.
The opening summary gives readers the answer. The final assessment synthesises the reasoning after the full analysis. They must agree. Important caveats belong beside the judgement, not hidden in an expandable annex.
Check, publish and revisit
Before release, check the proposition, factual claims, source provenance and independence, contrary evidence, assumptions, alternatives, likelihood, confidence rationales, citations, internal consistency and privacy. Verify AI-assisted claims against original sources. A model response is not corroboration.
Disclose material AI assistance and what was verified. An editor remains responsible for the assessment. Date and version the release, then revisit it when stated indicators or new evidence warrant review. Distinguish an error correction from an update and from continuing analytical disagreement. See Corrections & Challenges.