The Advance Basis research pipeline, explained

See how Advance Basis turns a research request into a source-grounded report. This article follows a real example through planning, source discovery, evidence review, claim verification, and final delivery. Part two takes a closer look at the finished report.

One autonomous research run

Question in. Auditable report out.

The stages exchange structured records, not a growing chat transcript. Each boundary limits what the next stage may treat as evidence.

01Plan
02Explore
03Select
04Distill
05Resolve
06Deliver

Full research Scan Same evidence standard No editorial approval queue

Stage 011 research plan

Shape the research before it runs

Begin with a question or detailed prompt and add any material you already have. PDF, Markdown, text files, and pasted notes can travel with the request. Intake turns that material into an editable research plan with an objective, scope, research lenses, questions, and a source strategy.

You review the plan before launch, then choose Full research or Scan. Scan is a leaner preliminary pass. It uses the same quote-provenance and independent-support standard as Full research.

The planning approach is anchored in STORM, including multi-perspective question development and retrieval-grounded organization (Shao et al., 2024).

Research requestDraft

How should model evaluations be compared, and what can the results actually establish?

PDFevaluation-notes.pdf

Research planEditable

✓Benchmark design
✓Disclosure duties
✓Mechanisms and limits

Stage 0216 searches

Explore broadly, then bring the sources back

The system expands the plan into multiple searches across the research angles, gathers candidate results, and retrieves original pages and documents. In the example run, 16 searches surfaced 273 candidates and 35 were captured as full source payloads.

Search snippets help decide what to fetch. They are leads, not evidence. Only material actually retrieved and retained can support later claims.

Research questionWhat does each evaluation measure?

Primary textregulatory record Peer reviewbenchmark study Documentationscoring method Counterpointreplication analysis

273candidates35full captures

Stage 0316 selected sources

Build the evidence set

Captured material is evaluated for relevance, authority, full-text availability, source role, and coverage of the plan. The example run selected 16 sources for its evidence set. Captured sources that were not selected remain outside the material used to write the report.

Coverage is a live research signal. If an important angle is still thin, the system can continue bounded exploration on its own. There is no operator review or customer-approval queue inside the run.

35 full captures

S03Primary

S11Analysis

S18Not selected

S27Counterpoint

Evidence selection relevanceauthoritycoveragefull text

16-source evidence set

S03

S11

S27

Coverage recorded

Stage 0416 source notes

Distill without losing the passage

Each selected source becomes a bounded, source-grounded note. The note preserves relevant claims, dates, methods, units, caveats, locators, and contiguous quote material. Downstream evidence work reads these retained notes rather than search snippets or model memory.

What retention guarantees: Advance Basis preserves the source material the run actually read and the passage used for support. It cannot guarantee that the publisher's original URL will remain live or unchanged.

S11Retained source

Exact passage

S11Source-grounded note

Date
Method
Caveat
Quotepinned

Stage 05222 claim verdicts

Resolve each claim against the evidence

Claim extraction proposes atomic candidate claims; it does not approve them. A separate model rereads the retained note and judges whether its meaning supports each claim. Independently, a deterministic check proves that the recorded quote is a real verbatim span in that note.

Both checks are required before a claim becomes assertible. Contradictions and insufficient evidence stay out of the report's supported claim set, while unresolved questions remain visible to the writer and the reader.

C-118Candidate claim from retained note S11Awaiting both checks

Quote provenanceS11

The recorded excerpt is found as a verbatim span in the retained note.

Matched ✓

Semantic supportIndependent

A different model judges the claim against the note without seeing the extractor's label.

Supported ✓

C-119Claim exceeds the evidenceInsufficient evidence

Stage 061 private report

Write from resolved evidence, then deliver

Synthesis receives the research plan, assertible claims, retained evidence, and unresolved gaps. It writes a structured draft, then backend validation checks the artifact contract, source tags, citation bookkeeping, and grounding before a private report is delivered.

Recoverable coverage or editorial limitations can ship with candid caveats when a valid report can be produced. Grounding failures, fabricated citations, and invalid artifact contracts remain hard failures. Reports stay private by default; the owner can publish a fixed share or revoke it later.

Resolved evidence

C-118S11

C-204S03

U-03Open

Advance Basis

Source-grounded report

Section01Decision context

S11

PrivateConfidence and unknowns included

Evidence DossierPDFSlides PDFWordMarkdownHTML

Complete package: report HTML, Markdown, and individual source files

Why the stages are separate

Each boundary prevents one stage's confidence from becoming the next stage's evidence. A convincing search snippet cannot become a citation. A fluent extracted claim cannot approve itself. A semantically supported claim cannot enter the report without a real quote, and a real quote cannot establish support without an independent judgment of its meaning.

The run is autonomous after launch. No operator approves sources, unblocks the workflow, or reviews the report as a product step. Full research and Scan differ in breadth and speed, not in their evidence standard. If the evidence supports a valid but limited report, the limits ship with it. If the system cannot satisfy grounding, citation, or artifact-integrity requirements, it stops rather than deliver an untrustworthy report.

The result of this run is public. Part two dissects it part by part, or you can read the full report exactly as it was delivered.

Part two: the anatomy Read the full report