Decentralised News · Editorial Standards · Research reviewed 20 November 2026
By Heath Muchena
Direct answer: DN evaluates claims about AI and agentic-finance products by separating documented specifications, independently reproducible calculations, observed product behaviour and independently corroborated outcomes. A product can have impressive capabilities on paper without passing an operational test; a calculator can be mathematically correct without forecasting the future. This article sets out the evidence and disclosure standard DN intends to apply to new and updated research. It does not claim that every historical DN article has already been re-audited against this policy.
What Matters
- Never use “tested” or “verified performance” unless the article records what was actually run, measured and observed.
- A vendor’s documentation is primary evidence about its stated product, but not independent proof of reliability, security or returns.
- Keep observation period, access date and publication date separate.
- Every comparison should have a denominator, scope, countercase and clear “unknown” category.
- Editorial decisions and commercial relationships should remain independent; affiliate availability is never a substitute for evidence.
The Signal: a confident answer is not a research method
A growing number of products describe themselves as “AI-powered”, “autonomous”, “institutional-grade” or “fully audited”. Those labels may be directionally useful, but they are not standardized results from a common controlled experiment. A research publication’s job is to translate such claims into testable propositions, identify which documents establish them and explain what remains unobserved.
Regulatory attention to misleading AI claims reinforces the need for exact wording. The SEC announced settled actions in March 2024 over misleading statements by two investment advisers concerning their purported AI use. That enforcement example does not imply that every AI finance company is dishonest; it demonstrates why an editor must ask what supports a claimed capability. [1]
DN’s standard is to present a reader decision with an evidence record, not to create a long “best” list whose rankings cannot be reconstructed.
DN Alpha Thesis
The publication’s durable asset is the evidence object: a material claim joined to its source, date, scope, calculation and unresolved questions. Models and website interfaces will change. A well-designed evidence record can be revisited, challenged and recomputed when those changes occur.
A document-supported product claim is useful, a sandbox observation is useful and a genuinely independent live result is useful. They are not interchangeable. The most defensible research reports the highest level actually reached and leaves the rest visible rather than manufacturing a single confidence score.
1. Four evidence levels, with distinct wording
| Evidence level | What DN has established | Permitted description | What it cannot establish alone |
|---|---|---|---|
| Level D: Documented | An official specification, filing, contract or dated source says something, with its scope recorded | “According to the issuer’s documentation…” | That the actual product operates as described |
| Level C: Calculated | An original formula applied to disclosed or synthetic inputs, checked for basic boundaries | “DN’s scenario produces…” | That the scenario occurred or is likely |
| Level O: Observed | A specified live or sandbox product action, data import, workflow or browser interaction was performed, with result and environment retained | “In DN’s observed test…” | That all users, chains or future deployments behave identically |
| Level R: Replicated | A result was independently repeated under documented matching conditions or corroborated through a suitable separate source | “Replicated under these conditions…” | Universal safety, profitability or future performance |
These four labels are an original DN editorial scheme, not an official NIST, W3C, OWASP or regulator certification.
2. Define the claim before collecting evidence
“The bot works” is not an adequately specific question. A better claim might be “this version of the bot can run a particular strategy without a live exchange key” or “the order route limits withdrawals at the signing layer.” Specify the product, version, integration, chain or jurisdiction, the relevant behavior and the observation date.
For a product’s economic claims, choose the unit: cost per successful prompt, cost per accepted task, gross transaction volume, recognized revenue, net customer receipt or return after costs. Each answers a different question. Do not combine unlike units into a single score merely because a ranking table needs one.
When a title includes “best” or “safest”, define what those words mean for a named reader segment. A platform that suits a skilled developer may not suit a first-time investor. Without an explicit decision criterion, a ranked list can reward editorial convenience rather than product fit.
3. Preserve provenance through the calculation
The W3C PROV-O specification distinguishes entities, activities and agents, making it possible to describe how a result was derived, which actors performed transformations and which source was used. This is a reference for provenance design, not proof that a cited item is true. [2]
DN’s proposed evidence record includes the claim, source URL, source publisher, original reporting period, date accessed, relevant excerpt or numeric field, units, uncertainty, transformation formula, reviewer notes and next review trigger. A source file’s creation date should not be mistaken for the period it describes. A model’s current timestamp should not be used to make yesterday’s data look live.
Preserve a correction history. If an issuer revises a number or a document becomes unavailable, the publication should update the visible conclusion while retaining enough information to explain why it changed.
4. The countercase is part of the finding
Before drawing a conclusion, state the strongest reasonable alternative. Higher revenue per employee could indicate improved automation or reduced hiring; a lower transaction fee could be outweighed by worse execution; higher token holdings could coexist with diluting common shareholders. The article should identify which evidence would change the interpretation, instead of designing the research to confirm the headline.
Where two sources disagree, do not choose the version that makes an affiliate or sponsor look better. If definitions differ, show both and explain why the comparison cannot be completed. “Not comparable” is a valuable finding when it prevents readers making a consequential mistake.
5. Separate operational testing from an interface demonstration
A screenshot of a dashboard demonstrates appearance at a point in time. A recorded browser calculation with changed inputs, invalid-state behavior and reset demonstrates more. An actual account-level service test needs access permissions, version, date, task selection, expected output, observed result and failed cases. A simulated account does not establish performance on an exchange with real orders.
For trading and payment systems, no publication should claim verified net returns without a suitably complete transaction record that accounts for fees, slippage, financing, rejected orders, changing balances and relevant capital flows. Testing one instance of a transaction does not justify a universal “safe” label.
The OWASP Agentic Applications framework identifies common attack classes involving tools, delegated authority and workflow hijacking. It provides useful risks to investigate, but citing that framework alone is not equivalent to penetrating a particular product or verifying its deployed contract. [3]
6. Disclose which calculations are synthetic
DN frequently builds calculators that help readers compare assumptions. A calculator can clarify the economic pressure in a scenario without predicting which price, failure frequency or customer demand will occur. Synthetic defaults should be labelled beside the controls. Where a rate is compounded or a percentage is applied, the denominator and unit must be visible.
For example, our white-label agent cost investigation models a hypothetical $299 retainer and a cost bridge; it does not claim a survey found that price or margin. Our startup evidence protocol supplies questions buyers can ask; it does not claim DN tested every startup with those questions.
An updated calculation should not silently overwrite an earlier assumption. If the formula changes materially, record what changed and whether previous conclusions require review.
7. Editorial versus commercial boundaries
Affiliate links, referral codes and paid services can help sustain a publication. They also create a potential incentive to emphasize products that pay commissions. DN’s policy is to separate the evidence underlying a comparison from the availability of an affiliate relationship and visibly disclose relevant commercial links.
Products with no affiliate relationship may be included when relevant; those with a referral link may fail the comparison or be excluded if inactive. A referral link should never be represented as an independent endorsement, and absence of commission should not automatically make a product trustworthy. Existing link attributes should be changed only when their actual legal or contractual treatment has been established.
8. A test receipt needs a date, configuration and boundary
A lightweight operational receipt can name the product and version, testing environment, test account or sandbox status, specific user task, whether keys or funds were exposed, allowed permissions, pass criteria, observed results, failures and the date. Keep credentials and private personal data out of the public record. Where no hands-on test occurred, the receipt should explicitly say so.
For an article with a calculator, independently check the default calculation, a changed assumption, a boundary case, invalid input and reset. For an exported dataset, confirm the download is actually available and preserves the intended columns or JSON fields. A live URL alone does not establish that each control works in a browser.
9. Source reliability and editorial judgment are not the same
An official company document is a strong source for a company’s statements and legal terms. It may still be self-interested about product reliability. A government rule can establish an obligation but not prove every firm complies. A blockchain transaction can establish a state change while saying little about the beneficial owner or off-chain cost. The editor must label the boundary each record supports.
NIST’s voluntary AI Risk Management Framework Playbook includes governance, measurement and risk-management activities, but does not require every organisation to use every suggested action. DN draws on its emphasis on documenting scope and risk without claiming NIST endorses our editorial process. [4]
10. Corrections and review dates
A newly published rule, platform release or pricing change can materially affect a previous conclusion. The article should carry a research cutoff and update trigger for any claim likely to change. When evidence changes, correct the record, retain the prior date and explain the reason when the update changes the reader’s decision.
Not all articles need a scheduled monthly refresh. A comparison of pricing and active services may need one sooner than a timeless explanation of mathematical mechanics. The cadence should reflect how rapidly the evidence changes and how consequential an outdated answer would be.
DN Publication Evidence Packet Checklist
For an editor or researcher, this local checklist shows which eight record categories are ready. The count is a completion aid, not a certification or score of truthfulness. No data is transmitted or saved.
Checkboxes signify a recorded review step; they do not prove that the evidence is correct. Some items require an explicit “not applicable” or “not tested” disclosure rather than a claimed success.
11. What DN cannot promise
DN cannot certify that every crypto wallet is secure, every trading bot is profitable or every AI model produces reliable financial decisions. No amount of source markup can eliminate underlying market uncertainty or undisclosed software defects. A model is a tool for analysis, not a substitute for the issuer’s underlying financial records or a regulator’s official text.
We can promise to distinguish what we have actually examined, calculated or observed, publish material limitations and correct identified mistakes. The quality of an investigation should be judged by its evidence and reproducibility, not by a claim that the publication has achieved an arbitrary search-engine ranking.
Methodology and Sources
This article is a prospective DN editorial policy grounded in a review of official methodological, provenance, security and enforcement sources. It is not evidence that every legacy article or referral was audited retroactively. The D/C/O/R four-level labels and eight-item evidence packet are original DN conventions. The standards cited below do not certify a publisher merely for using them.
- SEC’s 18 March 2024 AI-washing enforcement announcement — settled adviser statements about AI use, not a market-wide survey.
- W3C PROV-O provenance ontology — entities, activities, responsible agents and derivation relationships.
- OWASP Top 10 for Agentic Applications 2026 — security risk classes and mitigations, not a live product certification.
- NIST AI Risk Management Framework Playbook — voluntary governance and measurement suggestions.
Frequently Asked Questions
Does every DN product review include live testing?
No. Each article should disclose whether the research was a documentary comparison, a calculation, a live test or a replicated test. Absence of a recorded hands-on test means it should not be presented as one.
Does DN accept AI-generated content as independent evidence?
No. A model summary may help locate sources or organise questions, but material facts need the appropriate underlying record and analysis. A model-generated answer is not an independent primary source.
How are sponsored and affiliate products treated?
Commercial relationships require appropriate disclosure and should not determine the research outcome. An affiliate relationship is not a safety or performance endorsement.
Does passing the evidence checklist guarantee an accurate article?
No. It shows that review categories were completed; an error or new information may still require correction.
Continue the Research
For deeper application, see DN’s financial-audit approach to trading agents, source-defensible crypto AI search and startup due-diligence evidence records.
Editorial methods may evolve as evidence and technology change. This page describes a disclosure standard, not an external certification.




