# Weights fixed first, evidence second, scores last, and the scorer never knew the names

METHOD, read 10 Nov 2026

**The short answer.** Upstream scores 16 lineage tools on seven criteria: depth of lineage (25), lineage that carries meaning (20), served to agents (20), coverage of the stack (15), stays current (10), safe to deploy (5) and price transparency (5). Weights were committed on 10 Nov 2026 before any score, and the scorer saw coded names only.

The weights came first. They were written for a buyer who has never heard of any vendor here and wants lineage that goes deep, carries meaning and can be served to agents, then committed to the repository on 10 Nov 2026 (commit 091b5f9) before any evidence was scored. A scoring guide that interprets the anchors followed in commit cebd972, and the first scores in commit 7dbe012. No weight and no criterion has changed since.

The evidence came second: 398 facts across the 16 tools, each read from the vendor's documentation, release record or public code, with its source URL, its read date and its kind. A vendor's own claim is labeled as the vendor's. A fact confirmed in public code, a schema or a third-party write-up counts as a second source.

## The criteria and their weights

| Criterion | What is scored, with evidence | Weight |
| --- | --- | --- |
| Depth of lineage | Column-level lineage that follows transformation logic through to BI fields and metrics | 25 |
| Lineage that carries meaning | Definitions, ownership and governance rules travel with the lineage | 20 |
| Served to agents | Context delivered to AI agents over MCP, SDK or API, scoped to who is asking | 20 |
| Coverage of the stack | Warehouses, transformation, BI and security tools covered, as documented | 15 |
| Stays current without work | Automated collection, freshness, no manual curation | 10 |
| Official and maintained | For MCP servers: published by the vendor, versioned, recently released | 0 |
| Safe to deploy | Metadata-only access, permissions respected, audit | 5 |
| Price transparency | Published prices or pricing model | 5 |

Weights fixed 10 Nov 2026 in euno-sites/data/weights/datalineagecompare.json, before the first score. Official and maintained applies to MCP servers and carries 0 on this site.

## What each score means

- **0:** none found on the read date
- **7:** the most a vendor claim or vendor documentation alone can earn: detailed, specific, covering the whole criterion
- **1-3:** mentioned, partial, or limited to one system or one level
- **4-6:** documented and working, with gaps the documentation shows
- **8-10:** detailed and confirmed by a second source (third-party write-up, or observed directly in public code, schema or release record)

## The rules every score follows

- Every vendor is scored 0 to 10 on every criterion, whole numbers, from the evidence file only.
- A vendor claim scores at most 7 until a second source confirms it (kind third-party or observed).
- None found scores 0 and is written "none found on" the read date, never as a missing feature.
- Breadth facts (module counts, integration counts, headcount, funding) carry at most 10 percent of any ranking; integration counts inform Coverage only through the systems they name.
- The total is the weighted sum of the scores divided by 10 (0 to 100 scale, published to one decimal).
- Ties break on the higher score on the heaviest criterion, then by name.
- Once the first score exists, no weight and no criterion changes. A score changes only when a sourced fact changes, logged with the fact.

## The five situations and their own weights

Each situation page ranks the same scores under weights a buyer in that situation would use. These were fixed and committed with the site weights, before any score.

| Situation | Depth of lineage | Lineage that carries meaning | Served to agents | Coverage of the stack | Stays current without work | Safe to deploy | Price transparency |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [A data platform lead whose analysts no longer trust the dashboards](/best-for/untrusted-dashboards/) | 35 | 20 | 5 | 20 | 15 | 5 | 0 |
| [An analytics engineering team on dbt and Snowflake or Databricks](/best-for/dbt-warehouse/) | 30 | 15 | 10 | 20 | 15 | 5 | 5 |
| [A company putting AI agents on its data this year](/best-for/agents-this-year/) | 15 | 20 | 40 | 10 | 5 | 10 | 0 |
| [A governance lead who must show where a regulated field flows](/best-for/regulated-field/) | 30 | 25 | 5 | 15 | 5 | 20 | 0 |
| [A BI team with thousands of reports and no map of what feeds them](/best-for/bi-sprawl/) | 30 | 10 | 5 | 30 | 20 | 0 | 5 |

Situation weights, each summing to 100.

- **Untrusted dashboards:** Trust comes back when a wrong number can be traced column by column from the dashboard to its source, across the whole stack, without someone maintaining the map by hand.
- **dbt and a warehouse:** A dbt team needs column lineage through its models into the warehouse and the BI layer, refreshed on every run, at a cost it can plan.
- **Agents this year:** An agent can only use lineage that is served to it, scoped to the person it acts for, with the definitions attached.
- **A regulated field:** Showing where a regulated field flows needs column depth, the classification and owner carried along the path, and access that an auditor accepts.
- **BI sprawl:** Thousands of reports need lineage that reaches the BI fields of every BI tool in use, collected automatically.

## How the scoring was kept blind

Every tool was scored on every criterion by a scoring agent that received the rubric and an evidence packet in which every catalog, context and observability vendor name was replaced by a code, URLs were removed, and no entity was identified as the subject of the site. Warehouse, dbt and BI names stayed visible because coverage is scored on them.

A second agent with fresh context, given the same rubric and packet, checked every score against its evidence and changed 7 of 112 for rule or consistency reasons, each change logged with its reason. Names were mapped back only after both passes. A score changes only when a sourced fact changes, and the change is logged with the fact.

## What the scores do not cover

The scores read documentation, release records and public code. They do not test a tool on a live stack, and a feature that exists but is not documented scores as none found until the documentation shows it.

Breadth facts such as integration counts, headcount and funding carry no weight beyond the systems they name, which inform Coverage of the stack only. Prices are scored on what is published; a price on request scores lower than a published one.

## Sources

- Upstream scores, CSV: https://datalineagecompare.com/datasets/upstream-scores.csv (read 10 Nov 2026)
- Upstream evidence, CSV: https://datalineagecompare.com/datasets/upstream-evidence.csv (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://datalineagecompare.com/method/
