From a Question to a Dossier: The 51-Section Master Index in Deep Research

Ask most AI research tools a question about a stock and you get an essay: fluent, variable in shape, and silent about what it left out. NataPulse’s Deep Research takes a different route. Every run now produces the Master Index — a fixed 51-section U.S. equity dossier in which each section carries an explicit coverage status, every material number must cite its source, and the confidence score is computed from the evidence rather than asserted by the model. The result is a research document that can be audited section by section. As with everything NataPulse publishes, it is research evidence, not trade instructions.

A fixed table of contents, not a free-form essay

Deep Research is NataPulse’s multi-agent investigation workflow: eleven specialist roles — from fundamental and technical analysts to bull and bear researchers, a risk reviewer, and a final synthesizer — working across six phases, from planning to synthesis. The Master Index sits on top of that team as a contract: whatever the run finds, the output is organized into the same 51 sections, covering deterministic company facts, fundamental and technical evidence, news and macro context, debate, risk, and synthesis.

The fixed structure is the point. A free-form report hides its gaps inside prose; a fixed index makes them visible. If a run has nothing to say about a section, the section still appears — with a status explaining why. When the research scope is not a single issuer (a market-wide question, a comparison, a watchlist), the single-issuer sections are marked not_applicable rather than silently dropped. Nothing is ever omitted; it is accounted for.

Six statuses instead of false completeness

Each of the 51 sections declares one of six coverage states: complete, partial, missing, not_applicable, stale, or conflicted. A section is complete only when its mandate is answered with cited evidence. Stale flags evidence that exists but is no longer fresh enough to lean on. Conflicted is arguably the most useful state of all: it tells the reader that credible sources disagree, instead of papering over the disagreement with a confident average.

This status layer changes how the dossier reads. Instead of asking “is this report good?”, a reader can ask “which sections are solid, and which are thin?” — the same question a human research lead would ask before circulating a draft.

Provenance at the claim level

The Master Index enforces citation discipline where it matters most: numbers. Every numeric or material claim in the dossier must carry a citation. If a section makes an uncited numeric claim, the section is automatically downgraded from complete to partial — the claim survives, but the section’s standing does not. There are no invented figures: real data only is a platform rule, and the dossier’s claim ledger is how Deep Research enforces it at the sentence level.

The source ledger follows the same logic. It lists only the sources the run actually cited — not everything it retrieved or could have consulted. A bibliography padded with unread references would inflate apparent rigor; the ledger records what the evidence actually rests on.

Confidence is computed, not claimed

Language models are notoriously willing to state a confidence level on request. Deep Research does not ask. The confidence attached to a Master Index run is derived — a function of coverage (how many sections are genuinely answered), citation (how well claims are sourced), and freshness (how current the evidence is). The model never asserts its own confidence; the number falls out of the audit.

This has a practical consequence: a run built on thin or aging evidence gets a low score no matter how fluent its prose is. Confidence becomes a property of the evidence base, which is the only thing it should ever have measured.

Two modes, one labeling rule

Deep Research runs in two modes. NataPulse mode is grounded in the curated dataset — published events, clusters, reports, SEC filings, market, social, and on-chain evidence — and a grounded run must select at least one NataPulse source before it will start. General (“Normal”) mode reasons from broader model knowledge plus permitted live web research, with cited web sources.

The rule is that the label always travels with the output. Every non-grounded run carries an explicit disclaimer that it is not grounded in curated NataPulse data. A general-mode answer is never allowed to pass itself off as a statement of the curated record — the reader always knows which evidence regime produced the page in front of them.

The dossier is the deliverable

Completed Master runs are artifact-first: the structured result is rendered into a branded HTML report with scorecard and coverage charts, plus PDF and DOCX exports, delivered through authenticated, workspace-scoped endpoints and surfaced as a summary card in the Report Center. The dossier is designed to be filed, shared, and challenged — sections, statuses, citations, and derived confidence intact.

What it is not designed to do is tell anyone what to trade. The final stance in a Deep Research run is a synthesis of evidence and disagreement — invalidation conditions and monitoring points included — never an order to buy, sell, or allocate. Fifty-one sections, six statuses, one standard: if the dossier can’t show its evidence, it says so.