← سند dashboard All docsPublished copy (redacted) of sanad/CV-TAILORING-RULES.md · source updated 29 Oct 2026 18:30 Riyadh · synced 4 Nov 2026 05:16
سند — CV Tailoring Rules (draft, doc only)
Status: draft for Bandar's approval. Applies to the Intake & Matching bot (tailoring) and the Apply Desk (approval and submission).
1. Core rule: every line traces back to the source
- The client's original CV, parsed into a profile and confirmed by the client, is the only source of facts.
- Every bullet, title, date, skill, employer, degree and number in a tailored CV carries a
source_ref pointing to the line in the original CV, or to a fact the client confirmed and the confirmation was logged.
- A line with no
source_ref gets blocked: it's removed, or flagged as "needs client input". It is never sent.
2. Allowed
- Reordering sections and bullets to put the most relevant first.
- Rewording a real bullet to use the job ad's terms when it means the same thing (e.g. "managed customer complaints" → "customer experience management").
- Shortening, merging or dropping irrelevant content.
- Translating Arabic ↔ English faithfully.
- A summary/objective line built only from facts in the source.
3. Never allowed
- Inventing or inflating employers, titles, dates, degrees, certificates, tools, languages, numbers or results.
- Adding a skill or keyword just because the job ad asks for it.
- Changing dates to hide gaps; turning "familiar with" into "expert".
- Pasting job-ad keywords as hidden or invisible text.
- If the job requires something the client doesn't have, show it as a gap to the client. Don't paper over it.
4. Automatic checks before review
- Every line has a
source_ref (the check fails otherwise).
- A diff of facts: entity lists (employers, titles, dates, degrees, certificates, numbers) must be a subset of the source.
- An LLM "faithfulness" pass flags any claim stronger than its source line. Anything flagged goes to the client, never auto-passes.
- The language and RTL layout render correctly in the PDF/DOCX.
5. Side-by-side diff before approval
For each job, the client sees: job title + why it matched (AR/EN) | original CV on one side, tailored CV on the other, with changed lines highlighted and each line's source shown on hover | gaps and flags | buttons: Approve and apply / Edit / Skip. The approved version is frozen (we store its hash) and only that exact file is submitted.
6. Approval modes
- Default: approve each. Nothing is submitted without the client approving that job.
- Autopilot (opt-in only): the client switches it on with its own consent text, a daily cap, and filters (city, title, company type, minimum match score). It only covers tailored CVs that passed every check with no flags. Anything flagged falls back to approve-each. The client gets a daily digest and can switch autopilot off at any time.
- Never automated: Jadarat (the client applies through Nafath; we prepare everything), LinkedIn, and [platform]. No bots on those sites.
7. Logging
Store the source CV version, tailored version, diff, flags, approver, timestamp and submission channel for each application. This is kept under the retention rules in DATA-CONSENT-SPEC.md.