- 01OriginExplicit work-mode or schedule language
Sentences in unverified stories and comments.
- 02Company matchDistinct source excerpts
Source ID, URL, role, and observed date remain linked.
- 03TransformationRules count explicit mentions
Coverage, conflicts, and extraction confidence stay visible.
- 04Product useWork-setup clues
Reported mode and schedule become questions to confirm.
- 05LimitNever current company policy
Confidence is support—not accuracy. Unknown never becomes onsite.
How the evidence travels
Every insight keeps a trail back to its source.
b4joinacompany preserves what each source can support, names every transformation, and turns recurring terms and missing facts into questions—not conclusions.
✓Reported, submitted, derived, and destination evidence never collapse into one score.
The provenance thread
Follow one label from origin to limitation.
2,464 stories across 781 companies.
Two paths, kept distinct
Prepared questions are not the same as AI answers.
The checkpoint engine is rules-based. Ask is optional and uses a bounded set of company story excerpts. Keeping the paths separate makes the limitation of each one visible.
Rules prepare what to verify.
- InputCompany stories + joined comments
Role and date context are retained.
- ScanEight fixed candidate concerns
Term recurrence ranks management, pay, stability, growth, workload, culture, work setup, and hiring.
- OutputQuestion + guidance + rationale
Up to three source labels are attached. Thin evidence keeps an Evidence Gap.
Generated text needs source review.
- RetrieveA bounded set of company story excerpts
If lexical matching finds none, early company stories may supply fallback context.
- RequestThe provider is instructed to stay cited
Retrieved excerpts leave b4joinacompany only after Ask is used.
- ReturnAnswer with [S#] labels—or labeled retrieval fallback
Generated prose is not independently fact-checked. Inspect the cited stories.
The language boundary
Careful wording is part of the method.
Personal accounts can show what deserves investigation. They cannot establish policy, intent, or one final judgment about everyone’s experience.
b4joinacompany can say
- Several dated reports mention a related concern.
- A submitted role range provides unverified negotiation context.
- Explicit work-mode language appears in named source excerpts.
b4joinacompany will not say
- The company has good or bad culture.
- A submitted amount is an HR-approved salary band.
- Missing work-mode evidence means the role is onsite.
Change without rewriting history
A later release does not silently rewrite saved context.
Evidence is published in versioned releases. Corrections and new source material require manual maintainer review and a dataset update before they can appear in a later release.
- 01Published releaseVersioned evidence and lane-specific dates
Public research reads from the active release.
v1.3.0 - 02New source or correctionPrepared for maintainer review
The correction form stores a review record; it changes nothing automatically.
Not published - 03Integrity reviewIdentity, labels, joins, and files are checked
Integrity checks do not fact-check personal claims.
Maintainer run - 04Later releaseAccepted changes can enter a new version
Earlier release identifiers remain meaningful.
New version
Use the method
Put a company through the evidence trail.
Build one brief, compare two companies on the same unknowns, or carry the research into Deshi Mula with the extension.