Links others miss
Connections drawn across sources — e.g., recognizing several indications as one shared mechanism — that reframe where the real opportunity sits.
Process
Quality engine
Research-use only
The process is the moat. The dossier is what it produces.
A Zemi Dossier carries more depth, traceability, and decision structure than a conventional research summary because of how it is built: hundreds of sources synthesized for non-obvious connections, every claim put through an adversarial AI audit, and every number traceable to its origin. It is months of disciplined analyst work, compressed into one package you can verify rather than trust.
What the process guarantees
Gather broadly, connect across fields, audit every claim against its source, preserve what was excluded, reconcile the workbook, and release only when the package can support professional diligence.
Intake scale
Synthesis at a scale no individual reader can match.
A specialist tracking a fast-moving field reads what time allows and remembers a fraction. Zemi ingests the field — primary papers, trial registries, regulatory records, datasets, and company signals — and holds it all in a structured evidence memory so claims can be compared, reconciled, and connected at once. That breadth is the precondition for insight: the connections that matter usually sit between sources, where no single reading lands.
Hundreds of sourcesPrimary literature, trials, registries, regulatory records, and company signals, ingested per dossier.
14 frontier threadsThe Engineering Memory sample ran forward search across fourteen distinct frontier lines of inquiry.
313 identifiers verifiedIn that sample, every customer-facing identifier was resolved against live biomedical databases.
Months → one packageWork that would take a team months arrives decision-ready — and more thorough than a manual review.
Scale is not the product; it is what makes the product possible. The deliverable is the synthesis and the decisions it supports — not a literature dump.
Synthesis
The value is in the connections, not the recap.
Once the field is held in one place, the work is deliberately pointed at the non-obvious: cross-field couplings, missing measurements, failure modes, and near-discovery gates. This is the deep, cross-domain reading that turns a pile of citations into a thesis a buyer can act on — and that a single analyst, however expert, rarely has the bandwidth to produce.
What will make or break the field
The specific upcoming readouts most likely to confirm or refute a paradigm, each with the result that would settle it.
Claims with a kill switch
Forward propositions, each tied to a primary endpoint, the result that would break it, and a power-calculated study — not assertions.
Classifiers, not opinions
The synthesis is distilled into a reusable classifier or decision stack that routes a program to its binding gate.
The audit
Every claim faces an adversarial AI auditor before it ships.
This is the credibility layer. An automated auditor takes each claim and tries to break it: does the citation resolve, does the source actually support the claim, how direct is it, and how mature is the evidence? What it finds routes back to the dossier agents, which correct, re-anchor, scope-fence, or quarantine — iterating until the package passes clean.
Release-audited is not peer review or journal validation. It is an internal quality process for research information products used in professional diligence.
01
Citation resolution
Every identifier is resolved against live databases — NCBI E-utilities, Europe PMC, ClinicalTrials.gov, Crossref — and read at the abstract or full-text level. Retractions and expressions of concern are screened.
02
Source-support verdicts
Each claim gets an independent verdict on whether its cited source genuinely supports it — the check that separates a real citation from a decorative one.
03
Three-axis grading
Evidence layer (established / inferred / hypothesis), translation maturity (in-vitro → rodent-causal → human), and citation quality are graded separately and never collapsed into one another.
04
Figures & formulas
Figure provenance, displayed numbers, and workbook formulas are reconciled against their sources — including a live-vs-mirror check on every power calculation.
05
Correction loop
Findings become an issues log with a status and a closure note; the dossier agents fine-tune and correct until the audit is clean and the release is signed off.
What the audit removes
The funnel from candidate claim to customer-facing finding.
The audit is subtractive by design. In the Engineering Memory sample, more than two thirds of candidate claims were excluded or quarantined — and crucially, none of that negative work was discarded. It stays in the workbook with its reason code, so a reader can see exactly what was rejected and why.
| Stage | Count | What it means |
|---|---|---|
| Candidate claims audited | 239 | Every claim entering review, before any was kept or cut. |
| Findings kept | 87 | Survived the audit; 82 are customer-facing, 5 appendix-only. |
| Excluded / quarantined | 152 | 96 off-topic, 39 beyond report scope, 13 ungroundable, 3 near-duplicate, 1 below-quality — all logged. |
| Identifiers verified | 313 | 198 PubMed, 72 DOI, 28 PMCID, 15 trial registrations — resolved against live databases. |
| Retractions found | 0 | Zero retractions and zero expressions of concern across the registry. |
Counts are from the Engineering Memory sample dossier; every dossier carries its own reconciled denominator tree in the workbook's Source Quality Summary.
Traceability
Trace any number back to its source. That is what you are buying.
Complete traceability is what lets a purchaser use a dossier in real diligence: you do not have to take a claim on faith, because the chain from conclusion to source is in the file. If a partner, an investment committee, or a regulator asks "where does this come from?", the answer is one lookup away — and the places where the evidence is thin are labeled, not hidden.
Claim Ledger & Traceability Certificate
One row per finding: verbatim excerpt, report page anchor, source type, evidence layer, support verdict, grade, directness, and limitation note.
Figure Source Map
Every figure and displayed value maps back to the identifiers and calculations behind it, with a reproduction log.
Score Basis & Version Registry
How each score was assigned, and a revision history showing what changed between audit passes.
What it means for a purchaser: the dossier is auditable, not just readable. You can defend its claims to a third party, and you can see exactly where it stops short of proof.
The trust engine
The report carries the thesis. The workbook carries the proof of work.
The Evidence & Decision Workbook is where the dossier earns its credibility: claim ledger, evidence table, source-support verdicts, score basis, excluded-claim dispositions, salvage ledger, hypotheses ledger, most-valuable next studies, power calculations, classifier modules, version registry, and audit issues log — the machinery behind every sentence in the report.
Inspectable infrastructureUp to 45 sheets of evidence, claims, verdicts, and decision tools per dossier.
Negative work preservedExcluded and quarantined claims stay in the record with their reason codes.
Live, editable formulasPower calculations are working spreadsheet math you can drive with your own assumptions.
Audit-trailed releaseVersion registry and audit issues log show what was checked, found, and closed.
Execution layer
Power-calculated next studies turn the thesis into a plan.
Every major hypothesis is converted into a powered, costed, falsifiable study — endpoint, effect assumption, alpha, target power, computed sample size, design n, budget, and timeline — as live spreadsheet formulas your team can edit. That layer has its own page.
Return to the dossier library
The process supports the product. The homepage library shows the current domain-mapped dossier catalog.