Methodology
How the agentic reports are produced
MacroRadar's AI Search Reality Check reports are drafted by a controlled agent workflow and approved by a human editor before publication. "Produced by agents" must never mean automated slop, especially on a finance topic — so this page documents exactly how a report is made, what each step may do, and what is never left to a machine.
The principle
Agents amplify the work; they do not replace the judgement. Every report is agent-drafted and human-approved before publication. The transparency on this page is not a footnote — on a finance domain, it is the credibility.
The signals in every report come from MacroRadar's own models, documented in the methodology and bounded by the limitations. The audit records what AI systems say in their own words.
The workflow
Each edition moves through these steps. A draft never becomes a published edition without passing the critic and the human editor.
Demand
Mines the macro questions people and agents actually ask, so each report is built around real query intent rather than a topic chosen in a vacuum.
Data
Pulls MacroRadar's own derived outputs — recession probability, regime classification, sentiment — from the internal warehouse. No third-party series are re-hosted; raw data stays with its source.
Narrative
Turns the live signals into a plain-English read. Qualitative framing only; specific figures come from the data step, not from the model writing the prose.
AI-search audit
Runs a fixed prompt panel across the AI search engines and records, verbatim, what each says — whether MacroRadar is cited, which competitors appear, and whether the answer is accurate, stale, or incomplete against the live signals.
Critic
An adversarial pre-publication pass that flags unsupported claims, stale dates, hallucinated relationships, missing methodology links, and any promotional or overconfident language. Anything it flags blocks publication until resolved.
Human editor
A person reviews and approves every edition before it is published. This step is never skipped.
What is never automated
- The decision to publish. A human editor approves every edition; no report goes live on a schedule or by a model's say-so.
- The macro figures themselves. Signals are read from MacroRadar's pipeline, not written by a language model — a model never invents a probability or a regime label.
- Claims about what AI engines said. Audit answers are recorded as observed, not paraphrased into something more convenient.
- Resolving a critic flag. A flagged claim is fixed or removed by a person, not waved through.
Allowed data sources
Reports may use MacroRadar's own derived outputs and the public economic data behind them, documented on the data-sources page, plus the verbatim answers returned by the AI engines in the audit panel.
Reports do not draw on unlicensed third-party market or analyst consensus data. A capability is only described once it has actually shipped — no economic-calendar, central-bank, or inflation-pressure claims appear until those features exist.
Corrections
A published edition is dated and reflects the data and AI answers as known at publication. When we find a material error, we correct it and note that the edition was updated, rather than silently rewriting the record.
If you spot something that looks wrong, tell us — a flagged error is a better outcome than a confident one.