Trust resource
Measure company matching and technology detection separately.
A returned row can be wrong because Kaiju matched the wrong company, missed the relevant public evidence, or interpreted the evidence incorrectly. Test each layer independently.
- Company-match review
- Technology precision
- Technology recall
- Exception analysis
Example evaluation record
A concrete record, not an abstract claim
- Original website
- example.com
- Company match
- Confirmed
- Technology result
- Supported by two evidence types
- Review outcome
- Accepted
- Published
- Jul 20, 2026
- Last reviewed
- Jul 22, 2026
- Reviewed by
- Kaiju research team
- Applies to
- Current CSV company-research product
Start with known answers
Use companies your team can verify
A useful test set includes more than easy positive examples.
- 1Known technology users
- 2Known non-users where the required public checks can complete
- 3Ambiguous brands, subsidiaries, redirects, and multi-domain companies
- 4Blocked, incomplete, and intentionally difficult websites
Separate the errors
Do not hide different failures inside one accuracy number
Each error class points to a different product improvement and a different operational risk.
- 1Wrong company or website match
- 2Missed public evidence
- 3Incorrect detector interpretation
- 4Incorrect handling of incomplete or contradictory evidence
Evaluation outcome
Inspect false positives, false negatives, and unresolved cases
A practical evaluation should show where the product is safe to use and where human review remains necessary.
- 1Record why each disputed result was accepted, rejected, or left unresolved.
- 2Compare errors by field and evidence type rather than only by record.
- 3Set use-specific thresholds before automating a decision.
Continue the evaluation
Follow the next question behind the result.
Move from one trust concern to the methodology, product explanation, or limitation that helps your team decide whether the data is safe to use.
Test the method
Build a benchmark from companies you know.
Include expected positives, expected negatives, ambiguous websites, and difficult company matches before increasing volume.