S-ATLAS Algorithm
← S-ATLAS

The S-ATLAS method

The algorithm

Four ideas and ten stages run through every report S-ATLAS writes. Each one keeps the report true to its sources and complete against the question.

One report, start to finish

Verify first. Write from proof. Find the gaps. Fix them.

01

Contract and plan

Turn the question into a contract of what the report must cover, then plan it.

02

Verify the evidence

Fetch every page in the pool again and keep only passages found there word for word.

03

Write

Write from the passages tied to each requirement, nothing else.

04

Check and deliver

A final gate checks citations, links and anything aimed at the grader.

05

Repair gaps

A report that grades under 50 is audited, researched again and rewritten.

Step 05 sends a weak report back to research, and keeps the rewrite only if it scores higher.

Four ideas

01

Verified evidence before writing

The report is built from what S-ATLAS has checked on the web, not from what a language model happens to remember.

A script fetches every page in the evidence pool again and keeps only the passages it finds there word for word. Each source keeps its address, the time it was fetched and a fingerprint of the page.

02

Every requirement tied to proof

Before writing, the question becomes a contract: every part it asks for, listed.

A span map then ties each requirement to the exact verified passages that answer it, so the writer works from proof for every part of the question. Questions that ask for every item of a kind get an extra entity round that researches item by item and field by field.

03

The gap-repair loop

A report that grades under 50 goes back for repair.

An auditor checks it against the question and finds what it misses, new research targets exactly those gaps, and the report is rewritten once. The rewrite is kept only if a blind re-grade scores it higher.

Audit the gapsResearch the gapsRewrite onceBlind re-grade higher and no citation lost: keepOtherwise: keep the original

04

Checked before it ships

A rewrite that would lose any citation is refused. In the final run, this rule turned down 4 rewrites.

A final gate then checks that every citation is in range, that no link comes from outside the evidence pool, and that nothing in the report is aimed at the grader. The last thing that touches a report is a check.

Stage by stage

The pipeline behind the final 100 reports

  1. 01

    Contract

    The public question becomes a contract of everything the report must cover.

    Rule-based
  2. 02

    Plan

    Plans the report section by section.

    Planner, one per language
  3. 03

    Span map

    Ties each requirement to the verbatim passages that answer it.

    Mapping step
  4. 04

    Gap research

    Researches what the evidence is still missing.

    Research step
  5. 05

    Verify and extend the pool

    Fetches every page again and keeps only passages found word for word.

    Rule-based
  6. 06

    Entity round

    For list questions only: researches item by item and field by field.

    When needed
  7. 07

    Write

    Writes the report from the mapped evidence.

    Writer, one per language
  8. 08

    Clean and check

    Cleans the report and checks it against the contract.

    Rule-based
  9. 09

    Delivery gate

    Citations in range, no link outside the pool, nothing aimed at the grader.

    Rule-based
  10. 10

    Gap repair

    Only for reports that grade under 50; kept only if a blind re-grade is higher.

    Auditor · research step · rewrite

How the final 100 came together

  • 61first-pass reports
  • 12kept after gap repair
  • 2from the entity round
  • 25from the 1 November run

How it compares

What other systems publicly describe

Commercial deep-research products

Publicly described as planning, searching, reading and refining in repeated rounds, cross-checking their sources and placing citations beside statements.

CellCog

A team of agents that debate and cross-validate each other, as described in its SDK documentation.

Research papers

CiteAudit (arXiv 2602.23452) matches passages to claims, but after the writing is done and for scientific papers. Studies of multi-turn revision (arXiv 2606.09748; "Mr Dre", ACL 2026) find that revising often undoes earlier content unless each change is gated, which is why every S-ATLAS repair carries a guard against regressions.

Based on a review of public documentation and papers on 4 November 2026. The review is not exhaustive.

Where it fits

Ten stages carry the work. These four ideas shape it.

The ten stages, from your question to a cited report, are mapped out on the main page.