How sources are checked
What enters the evidence ledger, what gets rejected, and how conflicting claims survive.
Bind every page to the right company
Before a page can support a finding, the pipeline checks whether it is actually about the resolved company. This is especially important for short or ordinary company names that collide with unrelated businesses.
The entity guard checks known aliases, official domains, registered names, URL slugs, and the opening language of page titles. A review page for a namesake company is rejected rather than allowed to contaminate culture or team evidence.
Keep company claims separate from outside evidence
A careers page, founder interview, or company announcement can be useful, but it is the company describing itself. The ledger keeps that provenance. Independent reporting, customer-side evidence, structured company records, and workplace accounts are not treated as interchangeable.
This is why a customer logo is never equivalent to a paid deployment, and a funding announcement is never automatically equivalent to cash received. The source must support the exact claim the report makes.
Extract structured facts without losing the source
Unstructured pages can be converted into typed funding, founder, product, market, or workplace records. The extraction step is constrained by a schema and remains tied to the page that supplied the text.
Structured extraction does not promote a vague sentence into a precise fact. Dates, amounts, people, and claims still need to be present in the source, and records remain subject to entity checks and reconciliation.
Reconcile figures with stated rules
When sources disagree, the reconciler compares source tier, date, and the kind of fact. Measurements within ten percent or two absolute units can be treated as the same observation taken at different times. Material disagreements remain contradictions.
Some facts use explicit rules. A newer funding date can outrank an older one, but language such as “in talks” or “plans to raise” cannot become a completed round. “Raised” and “closed” are treated differently from an announced intention.
Preserve contradictions
An unresolved disagreement between a company claim and an independent record is not buried in a footnote or settled by whichever source arrived last. Both sides remain in the report with their sources.
This is often more useful to a candidate than a single cleaned-up number. The contradiction becomes a question the company can answer directly.
Show the retrieval boundary
Each report is evidence current to a particular date. Sources can change, disappear, or become outdated, so an older report is not presented as automatically current.
The report exposes source links where the data permits it and keeps unattributed or non-public records labelled by source class rather than inventing a URL.