From a research request to a source-grounded report: the stages, in the order they run, and why each exists. The stream below traces a research run, with each stage showing the number of actions or items processed. For the document it produced, continue to part two.
Stage 01 · Intake
Normalize the request, preserve the intent
1Confirmed research brief
Upload a brief, a memo, or a raw prompt, or talk it through with a request-builder model that interviews you and normalizes the request while preserving your original intent. Markdown, text, and PDF uploads can ride along with the request to augment the evidence and context corpus. It then proposes several research angles to accept, decline, or refine; for this run they included benchmark structure and design, regulatory disclosure under EU AI Act Article 50, and the mechanistic theory of why models hallucinate. The confirmed request is the single gateway to a run; Advance Basis is designed to prevent research diversions you didn't ask for.
Method lineage: the clarifying interview and research-plan preparation build on STORM, Stanford's methodology of multi-perspective question asking and retrieval-grounded outline synthesis (Shao et al., 2024).
Output: a confirmed research brief
Stage 02 · Acquire
Search, fetch, extract, retain
273Candidates considered · 35 captured in full
A coordinator model plans the searches and directs auditable workers that do the fetching. Sixteen distinct searches surfaced 273 candidate sources for this run: peer-reviewed papers, primary regulatory texts, benchmark documentation, and lab reports. Search snippets are acquisition metadata, never evidence. Workers fetched, extracted, and retained 35 sources as full payloads in evidence custody, so citations keep working after the original pages change or go offline.
Output: 35 retained source payloads
Stage 03 · Compress
Source-grounded notes, not summaries
16Source-grounded notes prepared
A compression model reads the retained payloads and prepares source-grounded notes: dates, methods, claims, caveats, and locators pointing back into the exact captured material. This run produced 16 notes from its source pack. Notes are the only reading material the later stages are allowed to work from, which keeps everything downstream traceable to a locator.
Output: 16 notes with dates, methods, claims, caveats, locators
Stage 04 · Verify
Extract claims, grade the support
222Claims extracted and graded
A claim-extraction model pulls candidate claims out of the notes. A different model rereads each note and judges semantic support, while a deterministic provenance floor proves that the displayed quote occurs in the retained text. A claim must pass both checks before synthesis can assert it. Refuted claims are quarantined, and what the sources could not settle is recorded as an open question.
Output: 222 verdicts, confidence tiers, open questions
Stage 05 · Deliver
One private artifact
1Private, source-grounded report
A synthesis model writes the report from the verified evidence, and hard quality gates check depth, structure, source tagging, and that no citation points at anything the run never retained. A run that hits genuine limits still ships, with its caveats stated plainly, so the result can be audited rather than lost. The artifact is private to your account until you choose to share it.
Output: a private, source-grounded report
Prepared with Advance Basis
Why the stages are separate
Each boundary in the stream exists to keep 'temptation' and hallucination out. Search results cannot be cited, however convincing a snippet looks; only retained content can. Retained copies keep every citation working after the original pages change or go offline. Claims must come from the captured evidence, with an anchor recorded for each one. And the verification step that grades a claim runs independently of the model that wrote it.
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.