Methodology & Transparency
How Candr works.
What it does and doesn't do.
Candr holds institutions accountable for their claims. That means we hold ourselves to the same standard. This page explains our methodology, our limitations, and what you should and shouldn't conclude from a Candr analysis.
The 4 laws of Candr
Every analysis Candr produces is governed by four immutable principles. These are not guidelines — they are hard constraints encoded into every analysis Candr runs.
1
Objectivity over opinion
Candr only flags what can be demonstrated with evidence and reasoning. It never speculates, never assumes intent without proof, and never lets ideology influence analysis. Candr applies identical standards to all entities regardless of political alignment, nationality, industry, or size. A strong claim that is genuinely verifiable will score high — even if we dislike the company making it.
2
Motive is always examined
Every claim exists in a context of incentives. Candr always considers what an entity has to gain or lose — financially, regulatorily, reputationally, politically, or in terms of control. Examining motive does not prove guilt. But it always contextualizes claims. A company with massive financial exposure making unverifiable sustainability claims is a red flag regardless of the claim's wording.
3
Language is never trusted at face value
Words are tools of power. Candr decodes what language is designed to do — not just what it says. We ask: who benefits from this specific phrasing? What does this word choice obscure? What is conspicuously absent? What would a lawyer have advised removing?
4
Transparency about limitations
Candr never overclaims. When evidence is insufficient, we say so. When a claim cannot be verified, we flag it as unverifiable — not as false. Honest uncertainty is a feature, not a weakness. Candr's credibility depends on never overstating what it knows.
How trust scores work
Every claim Candr analyzes receives a score from 0–100. This score reflects the credibility of the claim based on available evidence — not the ethical character of the company making it.
| Score range | Credibility level | What it means |
|---|
| 70–100 | High credibility | Claim is specific, measurable, and supported by verifiable evidence or named third-party verification |
| 45–69 | Medium credibility | Claim has some evidence signals but lacks full verification, specific numbers, or independent confirmation |
| 0–44 | Low credibility | Claim is vague, unverifiable, or contains documented manipulation patterns without supporting evidence |
The overall company score is the average of all scorable claims in the analyzed document. Governance boilerplate and noise are excluded from scoring.
Scores reflect one analysis of one document at one point in time. They are not permanent ratings. A company can improve its score by publishing more specific, verifiable claims.
The analysis pipeline
Every Candr analysis runs through the same six-step process.
1
Text extraction
Raw text is extracted from uploaded documents or pasted content. Full documents are never stored — only extracted claim sentences.
2
Claim identification
AI identifies meaningful claims — statements that make a verifiable or falsifiable assertion. Noise, table data, headers, and boilerplate are filtered out.
3
Manipulation taxonomy analysis
Each claim is analyzed against Candr's manipulation taxonomy — 6 categories covering language and rhetoric, psychological tactics, statistical manipulation, structural deception, propaganda techniques, and coercive language.
4
Verified evidence cross-referencing
Claims are compared against Candr's verified evidence layer — a database of legally confirmed enforcement cases from the SEC, EU regulatory authorities, and UK regulators. This grounds the analysis in real-world legal precedent.
5
Live news verification
Each claim is cross-referenced against current news coverage to identify contradictions between what a company claims and what has been independently reported.
6
Scoring and output
Claims are scored, labeled, and presented with plain English explanations, evidence gaps, and the hardest question each company doesn't want to answer.
Manipulation taxonomy
Candr detects six categories of manipulation. These patterns appear across all document types — sustainability reports, government policy, press releases, financial disclosures, and beyond.
🗣
Language & rhetoric
How words are chosen to mislead without technically lying.
Weasel words · Passive voice · Euphemisms · False precision · Scope manipulation
🧠
Psychological tactics
Exploiting cognitive biases to bypass critical thinking.
Authority bias · Bandwagon effect · False urgency · Anchoring · Social proof manipulation
📊
Statistical manipulation
Using numbers to create false impressions of progress.
Restated baselines · Cherry-picked timeframes · Scope exclusions · Misleading averages
🏛
Structural deception
Using document structure to hide unfavorable information.
Footnote manipulation · Report length as camouflage · Burying bad news
📡
Propaganda techniques
Systematic methods to shape public belief at scale.
Glittering generalities · Transfer technique · Card stacking · Repetition
⚖️
Coercive language
Language designed to create pressure without explicit force.
False necessity · Manufactured consensus · Accountability theater
Verified evidence layer
Candr's analysis is grounded in legally verified enforcement cases — not just pattern matching. The following cases are encoded into Candr's knowledge base as confirmed instances of institutional deception.
SEC Enforcement · $55.9M settlement
Vale S.A. — False safety claims
Vale claimed "strictest international practices in dam safety" while knowingly concealing that the Brumadinho dam was unsafe. The dam collapsed in 2019, killing 270 people. Pattern: Authority Theater + False Precision.
Italy AGCM · €1M fine · 2025
Shein — Emissions targets contradicted by own data
Published 25% emissions reduction target by 2030 while own data showed rising emissions in 2023 and 2024. Pattern: Temporal Displacement + Motive Misalignment.
Amsterdam Court · 2024
KLM — "Climate-neutral flying" banned
Court ruled climate-neutral flying claims were misleading. Aviation cannot be made climate neutral through offsets alone. Pattern: False Precision + Scope Manipulation.
UK ASA · 2021
HSBC — Positive claims omitting negative impact
Banned for promoting tree planting while omitting financed fossil fuel emissions — the Missing Counter pattern. Positive environmental claims that omit a larger negative impact are misleading by omission.
EU Multiple Regulators · €30B+ fines
Volkswagen — Manufactured test results
Marketed "clean diesel" while defeat device software falsified emissions tests. Real-world emissions were up to 40x higher than claimed. Pattern: False Precision + Systematic Fraud.
Additional verified cases: BNY Mellon (SEC, $1.5M), Keurig (SEC, $1.5M), Shell/BP (ACM Netherlands), H&M (Swedish courts), Drax (Ofgem, £25M), Lidl (German courts).
What Candr cannot do — Law 4 applied to itself
Candr holds institutions accountable for their claims. That means we must be transparent about our own limitations.
Candr cannot independently verify facts
Candr identifies claims that lack evidence and flags manipulation patterns. It does not conduct independent investigations, call sources, or verify primary data. When Candr flags a claim as unverifiable, that means the evidence required to confirm it is not present in the document — not that the claim is false.
Candr scores documents, not companies
A low score reflects the quality of claims in one specific document at one point in time. It is not a permanent rating of a company's character or ethical standing. Companies can and do improve. Scores should be interpreted in context.
AI analysis has inherent limitations
A peer-reviewed study (MDPI, 2026) found that general AI models show low reliability when detecting greenwashing in complex documents. Candr addresses this through domain-specific training, a verified evidence layer, and human feedback loops — but all AI analysis should be treated as a starting point for investigation, not a final verdict.
Candr results are not legal advice
Analysis results surface questions and flag patterns. They are not legal opinions, regulatory findings, or investment recommendations. Always consult primary sources and qualified professionals before drawing legal or financial conclusions.
PDF extraction quality affects results
Complex PDF layouts, scanned documents, and multilingual reports may produce lower quality claim extraction. Results are most reliable on text-based, single-language documents. The quality of input always affects the quality of output.
Data & privacy
📄
Full documents are never stored
When you upload a PDF or paste text, Candr processes it in real time to extract claims. The full document is never stored on our servers. Only short extracted claim sentences are saved (when learning mode is ON) to improve analysis accuracy over time.
🔒
Privacy mode — zero data saved
Professional and Enterprise subscribers can enable privacy mode. When active, nothing is saved — not the document, not the claims, not the analysis. Your report runs and disappears. Designed for sensitive corporate documents and regulated environments.
🧠
How learning mode helps everyone
When learning mode is ON, anonymized claim sentences are saved to Candr's training database. This is how Candr gets smarter over time — learning to recognize new manipulation patterns and improving accuracy for all users. Similar to how spell checkers learn from corrections without storing full documents.
Candr holds itself to the same standard
We analyze corporate transparency claims. That means we must be transparent about ourselves.
⚡
Infrastructure
Candr runs on Netlify's European infrastructure. Our database is hosted in Frankfurt, Germany on servers powered by renewable energy.
🤖
AI model
We use GPT-4o-mini — a smaller, more energy-efficient model than larger alternatives. Deliberately chosen. Less compute, same quality analysis.
💳
Pricing
Flat rate. No tokens. No credits. No slot machine. We don't profit from inefficiency. Our pricing aligns our success with yours.
🏢
Independence
Candr has no corporate investors, no advertising revenue, and no data selling. Our only revenue is subscriptions. We are accountable to users, not funders.
Candr is an AI-assisted analysis tool. Results surface questions and flag patterns for investigation — they are not legal verdicts, regulatory findings, or investment recommendations. Candr applies the same methodology to all entities regardless of size, nationality, industry, or political alignment. Always consult primary sources before drawing conclusions. If you believe a Candr analysis contains an error, contact hello@candrai.com and we will review it.
Candr is built and operated by Luis Roman, Biel/Bienne, Switzerland. hello@candrai.com · candrai.com