Investigate the right layer
Greenhouse’s parsing documentation and Gupy’s data notice describe different parts of the journey. The former helps investigate conversion of a document into fields; the latter distinguishes application data and employer-configured features. Neither provides a universal public hiring formula. That difference suggests a diagnostic method: locate the observation, form a bounded hypothesis and choose a check capable of testing it.
Case A — experience entered the wrong field. A fictional candidate sees the employer name under education. There is evidence of a representation problem in that form. She can correct the field, check the attachment and simplify the next version’s structure. If the field then shows the employer correctly, that check succeeded. It does not establish improved ranking: that second outcome was not observed.
Case B — the file is correct, but a condition is missing. A fictional applicant has relevant experience, but the role requires on-site work in another city and they only accept remote work. Repeating ‘operations’ in the document does not resolve that condition. The useful decision is to clarify flexibility or select another opportunity. The same distinction applies to a required skill they cannot yet demonstrate: naming it does not replace learning it.
Case C — no response. Several explanations remain possible: competition, a changed vacancy, review timing, unclear evidence or unknown criteria. Without feedback, choosing ‘ATS problem’ as the cause is guesswork. Improve observable defects while retaining uncertainty about the rest.
| Observation | What it supports | What it does not support |
|---|
| Copied text appears out of order | Investigating file construction | Claiming automatic rejection |
| The form displays correct dates | Those fields were represented correctly | Guaranteeing interpretation of the whole document |
| A recruiter asks for a leadership example | Preparing a situation and your actual role | Inventing team-management responsibility |
| An external tool returns 92 points | That tool applied its own criteria | Knowing another system’s internal score |
When comparing versions, state the purpose of the change: correcting a date, clarifying a contribution or explaining an acronym. If you simultaneously change target role, selected experience, format and market, a different response cannot identify which change caused it. Even a careful comparison across a few applications is not a controlled experiment: jobs and evaluators vary. Use records to learn about your process without turning correlation into a promise.
This also limits ‘ATS testing’: local extraction can reveal a real defect, but a simulation does not reproduce every employer field, criterion or decision. The useful outcome is a more accurate, verifiable application together with an honest account of what remains unknown.