# Prompt 03 — Georgia Monthly Employment Tracker Report type 3 in `../reports.md` (procedural index/tracker). Modelled on the structure of PMC Research Center's monthly *Employment Tracker* (produced with jobs.ge), whose July 2026 issue was read in full: it has exactly three data blocks, each from a public source — 1. persons receiving a salary, by salary band — **Revenue Service** (rs.ge); 2. vacancies published on **jobs.ge**, total and by category; 3. quarterly labour-force indicators — **Geostat** labour force survey (LFS); plus an optional quarterly Beveridge-curve chart (jobs.ge + hr.gov.ge + Geostat) that this prompt does not attempt. Worked example: `examples/2026-09-21-employment-tracker.md`, built from `examples/2026-09-21-employment-tracker-inputs.md`. ## What "reconstructable" means here Blocks 1 and 3 are *lookups*: the numbers exist as official series and any run on the same date gets the same values. Block 2 is *procedural*: jobs.ge does not publish counts, it publishes listings, and listings expire after about 30 days. A single run therefore yields the trailing ~30-day window only. To reproduce PMC's month-by-month series you must run the count at each month-end (or accumulate daily) and keep the results — the first run gives one data point, not a history. State this in any published output. ## How to run it **Option A — model with code execution.** Run `scripts/jobsge_count.py` (pure Python, no browser), query the two Geostat/rs.ge sources as described, then write. **Option B — model without tools.** Gather the three blocks by hand (the rs.ge form and the Geostat PX-Web table are point-and-click; the jobs.ge count needs the script or a manual tally per category from the listing pages), paste them as an input pack, and run Part 2. --- ## Part 1 — Gather inputs ### Block 1 — Persons receiving a salary (Revenue Service) - Page: https://www.rs.ge/statistics → "ინფორმაცია შემოსავლის მიმღებ ფიზიკურ პირთა შესახებ" (information on income-recipient natural persons). Form: year (`#StatisticsYears`), month (`#StatisticsMonths`, `00` = whole year, `01`–`12`), payment type (`#StatisticsSalaryTypes`, `1` = ხელფასი / salary). Result: a table of income ranges with the number of persons and taxable income per range; an Excel export button exists. - Endpoint behind the form: `POST /RsGe.Module/Statistics/SearchStatistics` with `period=YYYYMM`, `salaryType=1` and the page's `__RequestVerificationToken`; returns `{"Statistics":[…]}` with 31 income ranges. It is slow — 2 to 8 minutes per period in testing — so run `scripts/rs-fetch.js …` in the background (it keeps the page's cookie/token pair) and aggregate the ranges into PMC's bands with `scripts/rs_bands.py`. A manual fetch in a browser is an acceptable fallback for a one-off run; say so in the sources. - Record for the latest month, the previous month and the same month a year earlier: total persons; persons in the bands 0–599, 600–1,199, 1,200–2,399, 2,400+ GEL (and 9,600+ if reported). Data are gross salaries and are revised — note the date viewed, as PMC does. ### Block 2 — Vacancies on jobs.ge - Run `python3 scripts/jobsge_count.py --csv ` (in `georesearch/` the script lives under `scripts/`). It calls jobs.ge's scroll endpoint `POST https://jobs.ge/en/?page=N&cid=&jid=1&for_scroll=yes` (300 rows per page, job type `jid=1` = jobs only), parses id / title / company / published / deadline, and prints counts by category and publish month plus the earliest publish date still listed. - Category ids (jobs.ge `cid`): 1 Administration/Management, 3 Finance/Statistics, 2 Sales/Procurement, 4 PR/Marketing, 18 General Technical Staff, 5 Logistics/Transport/Distribution, 11 Construction/Repair, 16 Cleaning, 17 Security/Safety, 6 IT/Programming, 13 Media/Publishing, 12 Education, 7 Law, 8 Medicine/Pharmacy, 14 Beauty/Fashion, 10 Food, 9 Other. PMC reports five named categories (Sales/Procurement, Administration/Management, Finance/Statistics, Logistics/Transport/Distribution, IT/Programming) plus "Other" — group the rest into "Other" to match. - Record: the count for the latest *complete* month if the run happened at month-end, otherwise the month-to-date count with the cut-off date; the per-category counts; the earliest date in the window. If you have earlier runs, add MoM and YoY changes; if not, say there is no history yet rather than inferring one. ### Block 3 — Labour-force indicators (Geostat LFS, quarterly) - PX-Web table: `Social Statistics/Labour/LFS_by_sex_and_urban_(Quarter_II,_2026).px` (the file name carries the latest quarter and changes each release — list the folder first). Variables: `Sex, Urban/Rural Areas` (0 = Total), `Economic Status` (1 labour force, 2 employed, 3 hired, 4 self-employed, 6 unemployed, 8 unemployment rate, 9 participation rate, 10 employment rate), `Years`, `Quarters` (0 annual, 1–4). Query the latest quarter, the previous quarter, and the same quarter a year earlier, explicit codes only. - Also the latest quarterly earnings release (Geostat "Average Monthly Nominal Earnings of Employees", PDF) for the average salary line. ### Calendar - rs.ge data for month M appear during month M+1 (PMC viewed July data on 20 August). Geostat LFS for quarter Q appears ~7 weeks after quarter end (Q2 2026 → 19 August 2026). Geostat earnings for Q appear mid-month 3 of the following quarter. --- ## Part 2 — Write the tracker ``` You are producing a monthly labour-market tracker for Georgia for professional readers. Use only the INPUT PACK. Do not invent, extrapolate or "estimate" any figure; where a comparison (MoM, YoY) is not in the pack, omit it and say why in one clause. Sections, in this order: TITLE: "Georgia Employment Tracker — " (the latest month with salary data; if the jobs.ge count is month-to-date, say so in the title's subtitle line: "jobs.ge data to "). 1. PERSONS RECEIVING A SALARY — 3–4 sentences: total for the month with MoM and YoY change; share of persons under 600 GEL and its change in percentage points; share at 2,400 GEL or more and its change; the 9,600+ share if present. Then a table: Band | Persons | Share, for the latest month. Footnote: Revenue Service, gross salaries, data not final, viewed . 2. VACANCIES ON jobs.ge — 3–4 sentences: total published in the period with the exact window used; the largest and smallest categories; MoM and YoY only if earlier runs exist. Table: Category | Vacancies | Share, for the five named categories plus Other, sorted by count. Footnote: jobs.ge, job type "Job" only, listings live ~30 days, counted on ; no history before . 3. LABOUR MARKET INDICATORS — Table with columns for the same quarter a year earlier, the previous quarter and the latest quarter: labour force, employed, hired, self-employed, unemployed (thousands); participation, employment and unemployment rates (%); average nominal monthly salary (GEL) where available. One sentence on the direction of the unemployment rate QoQ and YoY. 4. SOURCES — Institution — dataset/release — period — URL or endpoint — date accessed. Mark anything fetched manually. Register: neutral and dense; every sentence carries a figure or a date. Use "MoM", "YoY", "QoQ" after defining them once. Percent changes to one decimal; percentage-point changes labelled "pp". Re-check every number against the input pack before finishing, then output the tracker only. ``` --- ## Known limits - The jobs.ge count is only comparable with PMC's if taken at month-end for a full month; a mid-month run is a partial figure and must be labelled. PMC may also count listing types beyond "Job" — its July 2026 total (8,138) is not directly checkable against a run made in September. - rs.ge's endpoint is slow and token-protected; treat it as "manual-fetch-friendly" rather than reliably scriptable. - The Beveridge curve in PMC's version needs seasonally adjusted series and hr.gov.ge public-sector vacancies — out of scope for this prompt.