Skip to content

Three Stages

Jump to bottom
b2o2i edited this page Apr 23, 2026 · 1 revision

The Three Stages — GATHER · VERIFY · INTERPRET

Every Cross workflow belongs to exactly one of three stages. Each stage answers a categorically different question, and each tool is named for the stage it serves. Knowing which stage you're in tells you which tool to reach for next.

Stage Question Tools
GATHER Where does the content come from? st-fetch · st-gen · st-bang
VERIFY What are the facts? st-fact · st-cross
INTERPRET What do the facts mean? st-verdict (one report) · st-analyze (across reports)

GATHER — Where does the content come from?

The content is the substrate the rest of the pipeline operates on. There are three ways to get one into a subject.json container:

  • st-fetch — pull external content into a container. Accepts a URL (web article), a file (PDF, Markdown, plain text), the clipboard, or a tweet. Use --prep to also extract the prompt so downstream stages know what topic to evaluate against.
  • st-gen — ask one AI to write a fresh report from a prompt. The simplest gather: prompt in, single-author report out.
  • st-bang — ask N AIs in parallel to each write a report from the same prompt. Use this when you want to compare authors before any verification — a wider sample at the GATHER stage makes the VERIFY stage more informative.

After GATHER, your container has at least one entry in story[] and (if you used st-fetch --prep or st-gen) a prompt in data[0].prompt.

VERIFY — What are the facts?

Verification is purely a data-collection stage: each tool walks the report claim by claim and writes a structured verdict (true / partially_true / opinion / partially_false / false) into the fact[] array on each story. It does not interpret anything.

  • st-fact — fact-check one report with one AI. Appends one fact[] entry per story. The fastest path to a single second opinion.
  • st-cross — fact-check every report in the container with every AI provider configured (the full N×N cross-product matrix). Appends one fact[] entry per (story × checker) combination.

After VERIFY, every story has one or more fact[] entries with parseable claims and verdicts. The container is now ready for interpretation.

Architectural rationale (cross-st 0.7.0): st-fact is now a pure verifier — its --ai-* interpretive flags moved to st-verdict (see the migration table on the st-fact page). The smell behind the move was that st-fact --ai-caption produced an interpretation of the chart, not a fact-check verdict — two categorically different operations conflated under one tool.

INTERPRET — What do the facts mean?

Interpretation reads the verdicts produced by VERIFY and turns them into something a human can act on — a chart, a written analysis, or a recommendation for the next step.

  • st-verdict — interpret one report's fact-check evidence. Produces the stacked verdict bar chart, AI-written captions / summaries / stories at five word-count contracts, and four focused lenses:
    • --what-is-false — focused breakdown of inaccurate / disputed claims
    • --what-is-true — focused breakdown of verified / supported claims
    • --what-is-missing — what important aspects of the prompt the report omitted
    • --how-to-fix — recommend exactly one next action: st-fix / st-bang -N / st-merge / publish-as-is
  • st-analyze — interpret patterns across reports in a container. Useful when you've run st-bang or st-cross and want a synthesis that spans every author.

After INTERPRET, you have a chart, a written analysis, or a one-line recommendation. Decide your next step from there.

Three killer workflows

Every common Cross use case is one of these three GATHER + VERIFY + INTERPRET combinations. Full copy-pastable transcripts on the Showcase Workflows page.

Workflow Question Pipeline
A — "Is this fake news?" What is wrong with this report? st-fetch <url> --prep → st-cross → st-verdict --what-is-false --ai-summary
B — "What's missing?" What did the report fail to cover? same gather → st-verdict --what-is-missing --ai-summary
C — "What can I trust here?" Which claims are well-supported? same gather → st-verdict --what-is-true --ai-caption

The fourth lens — --how-to-fix — sits one step further down the pipeline; run it after any of the three above to get a concrete next-action recommendation rather than a description.

Why the three stages matter

Before cross-st 0.7.0, the boundaries between VERIFY and INTERPRET were blurred — st-fact had interpretive flags that let it generate written analysis, and the same code paths handled both verdict-collection and verdict-interpretation. The result was that a tool's name no longer told you what it did. The 0.7.0 refactor fixes that: every tool now answers exactly one of the three questions, every flag belongs to the stage that owns its question, and every workflow is a clean GATHER → VERIFY → INTERPRET pipeline.

If you find yourself wanting a tool to do work from a different stage, you're probably reaching for the wrong tool. Look at the table at the top of this page first.

Related: Showcase Workflows · st-fetch · st-cross · st-verdict · Container Format

Clone this wiki locally