# Automaziot proposal line-item benchmark — publication methodology

Status: internal evidence pack. It must not be rendered publicly while the Phase 1 net-new URL freeze is active.

## Strict publication population

The public benchmark uses 174 distinct Automaziot proposals contributing at least one line that meets every rule:

- `base_packages` line with a numeric price greater than zero;
- currency explicitly equals `ILS`;
- `is_retainer` explicitly equals `false`; and
- category maps to one of the named categories.

The strict population is dated 2026-04-05 through 2026-07-09. Its states at extraction were 70 draft, 89 sent
and 15 accepted. “Accepted” is a workflow state; it does not prove delivery, adoption, revenue or outcome.

The statistical observation is one line item, not one proposal, client, business or complete project. A proposal
can contain several categories and several lines in the same category. Every displayed row therefore publishes both
`n_proposals` and `n_line_items`; categories overlap.

## Why the strict subset is used

The broader source audit found 362 proposals with a positive priced line. Across all positive lines, 855 explicitly
say ILS and 754 older rows have no currency value. Missing currency probably reflects legacy ILS records, but the
public benchmark does not impute it. A row with a missing `is_retainer` flag is also excluded rather than assumed
non-recurring.

This conservative choice reduces the displayed population but removes the two assumptions that caused the first
draft to fail independent review. It also suppresses support: only nine strict support proposals remain, below the
n=10 publication floor.

## Deterministic extraction

The complete read-only PostgreSQL query is versioned at
`scripts/extract-proposal-line-item-benchmark.sql`. It:

1. expands `custom_items.base_packages` only;
2. requires numeric `price > 0`, explicit `currency='ILS'` and explicit `is_retainer=false`;
3. keeps multiple same-category lines as separate price observations;
4. calculates the 25th percentile, median and 75th percentile with PostgreSQL `percentile_cont`;
5. counts both distinct proposals and line items; and
6. returns the population/status audit with the segment output.

The source category is `category_id` when present, otherwise `category`. Aliases are fixed:

- automation: `automation`, `whatsapp_automation`, `admin_automation`;
- foundation: `foundation`, `infrastructure`, `catalog`;
- documents: `documents`, `documentation`;
- operations: `operations`, `management`;
- support: `support`, `maintenance`, `monitoring`;
- AI agent, integration, data, visibility and voice map to their same-named source category.

Other heterogeneous custom work is excluded from named output.

## Price definition

The value is the stored line-item `price`. It is before proposal-level VAT, and the extractor does not apply a
proposal-level discount. It is therefore neither a final invoice nor a full-project total.

## Eligibility and release gates

- An overall benchmark may be named only when at least 30 strict proposals support it.
- A category may be named only when at least 10 strict proposals support it.
- Values must be labeled in ILS and as explicitly non-recurring line-item prices.
- Copy must say the data describes Automaziot proposals, not all Israeli businesses or the Israeli market.
- No claim may imply causation, guaranteed ROI, an industry-wide norm, project completion or future pricing.
- JSON, CSV, chart and rendered table must match the same dated SQL result.
- Before release, rerun the query and have an independent model challenge the wording and method.

## Safe public wording

Hebrew:

> ניתחנו 174 הצעות אוטומציות AI שהכילו לפחות שורת תמחור מפורשת בשקלים וסומנה כלא-חוזרת. הנתונים
> מציגים התפלגות של שורות שירות, לא מחיר פרויקט מלא ולא סקר שוק. הצעה אחת עשויה להיכלל בכמה קטגוריות
> ובכמה שורות באותה קטגוריה.

English:

> We analyzed 174 Automaziot proposals containing at least one explicitly ILS, non-recurring priced line. The
> figures describe service line items, not full-project totals or a market survey; one proposal may contribute to
> several categories and several lines within a category.

## Privacy

The public snapshot contains aggregates only. Client, company, contact, proposal and project identifiers, free text,
transcripts and line descriptions are prohibited. Low-N cross-tabs are prohibited. Public source labels use
“Automaziot proposal system,” not an internal schema or table path.

## Editorial and bilingual use

The strongest initial use is a concise evidence block inside the existing pricing owner, not a new URL. A standalone
research URL requires the query-owner, Phase 1 technical and net-new release gates. A public release must provide
stable HE/EN download links for the versioned CSV/JSON, the methodology and the accessible chart. The English
summary must use the identical snapshot and must not publish before the English emission/hreflang gate clears.

## Refresh cadence

Refresh after the tenth newly strict-eligible proposal or at least quarterly. Preserve dated snapshots. If a
category falls below its proposal threshold, suppress it rather than merge it into an unrelated category.