How reports are built
From one company domain to a source-linked decision brief.
Start with the domain, not a name
A company name is not a reliable identifier. Different businesses can share the same name, and search results often blend them together. We begin with the official domain supplied by the candidate and resolve it against structured company records.
The identity gate looks for at least two corroborating attributes, such as the domain, company record, LinkedIn identity, founding year, or registered name. If the match is too weak, the run stops instead of researching the wrong company.
Collect eleven lines of inquiry in parallel
Once the identity is resolved, eleven workers fan out together. They examine identity, funding, founders and leadership, product and customer evidence, team structure, departures, workplace accounts, compensation and equity context, market risk, adverse evidence, and recent news.
The current pipeline can draw from Coresignal, Firecrawl, Exa, Parallel, TinyFish, PredictLeads, and structured web search. A provider returning nothing is recorded as a coverage gap; another worker does not silently invent the missing answer.
Fill actual gaps, not every possible gap
After the first fan-out, the pipeline checks which questions have no useful records. Only those holes receive a targeted second search. Funding can also receive a later investor-confirmation pass because investor names are not known until the first funding evidence has been collected.
This order limits unnecessary calls while giving thin sections a second chance. It also preserves the difference between “we did not search” and “we searched and found nothing usable.”
Assemble an evidence ledger
Every accepted record is written into a common evidence ledger before analysis begins. The ledger keeps the claim, value, source tier, URL where available, relevant date, retrieval date, confidence, and the report question it belongs to.
Live and replay modes feed the same ledger into the same analysis code. Replaying an existing ledger makes the scoring reproducible without paying for collection again or changing the evidence underneath the report.
Analyse deterministically
The analysis stage does not search the web and does not call a language model. It parses the ledger, reconciles competing facts, computes signals, scores the evidence against the company stage, applies risk caps, and generates consequences and questions.
Because the date can be injected, the same ledger can produce the same report again. Freshness still matters, but it is explicit rather than hidden inside a new search.
Write for the next conversation
The finished report keeps observed evidence, supported claims, inference, contradiction, unknown, and private information visually separate. It includes the source trail, the gaps, the questions worth asking, and people who may have relevant context.
The output is decision support. It does not tell a candidate to accept or reject an offer, and it does not turn missing evidence into confidence.