Methodology

How every Delvant deliverable is built and verified

Delvant ships custom clinical intelligence under a fixed methodology. Every claim is tier-classified for provenance. Every PDF passes three pre-delivery gates before it leaves the workshop. Anchor sources are locked. The disclosed process is the product.

1. Per-claim provenance tagging

Every claim in every deliverable carries a label. Readers always know the analytical posture. No silent inference, no unattributed forecasts.

SOURCED
Direct primary citation

Claim is supported by a named, page-locked citation in an anchor source. Re-verified at the cited URL before delivery (Gate A).

e.g. “Phase III pivotal enrolled 217 patients (NCT04562519, ClinicalTrials.gov as of 2026-04-30).”
INFERRED
Cross-trial or indirect

Claim derives from cross-referencing two or more sources, or applying a known framework to fresh data. Basis stated explicitly.

e.g. “Based on platelet response curves in TPO-RA pivotals vs the rozanolixizumab Phase III readout, FcRn ranks favourably on speed-of-onset.”
OPINION
Analytical interpretation

Author judgement, declared as such. Not data, not forecast. Useful for framing and recommendation, never as evidence on its own.

e.g. “Reading the EMA assessor concerns alongside the label wording, the post-launch real-world cohort will likely shape Year-2 reimbursement.”
HEDGED
Low-confidence signal

Forward-looking, partial-data, or single-source claim where confidence is explicitly low. Flagged to prevent overreliance.

e.g. “A single congress abstract (EHA 2026 P1457) suggests a possible signal in pediatric ITP, not yet replicated.”
When no claim is possible - the MATERIAL GAP marker: if the assembled evidence does not support the analytical step the scope requires, Delvant surfaces the absence explicitly with a [MATERIAL GAP] marker instead of silently filling from model parametric memory. Resolution paths: reduce scope, pull a new source (priced separately if outside the locked anchor set), or replace the claim with an analyst note explaining what would change the read. Pattern adapted from the open-source Anti-Leakage Protocol.

2. Three pre-delivery gates

No PDF ships without all three gates green. Each gate is run by a separate process, not a single review pass.

Gate A Deep-research verification

Every Sourced claim is re-verified at the cited URL or page reference. Broken links, retracted papers, and outdated data trigger removal or re-tagging. A claim survives Gate A only if the source is reachable, current, and supports the wording as written.

Gate B LLM-council adversarial review

Every analytical claim (Inferred + Opinion) is reviewed by a council of independent LLM instances briefed to attack the reasoning. The reviewer cohort tries to falsify, find counter-evidence, or surface alternative readings. Claims survive only if they hold against adversarial pressure or get rewritten with stronger basis.

Gate C Originality.ai plagiarism scan

Every T1, T2, T3, and T4 deliverable is scanned with Originality.ai Pro before PDF render. Target: under 2 percent match excluding direct citations. The scan PDF is logged in the engagement folder; the scan result line is printed on the deliverable's methodology slide.

3. Seven failure modes the gates explicitly block

Each Delvant deliverable is classified against seven AI-research failure modes before ship. Any mode flagged Suspected or Insufficient Evidence blocks delivery until resolved. Taxonomy adapted from Lu et al. (Nature 2026, 651:914-919) and the open-source Academic Research Skills checklist.

1

Data-extraction error

A figure mis-parsed from a source PDF or table reads plausibly. Decimal shifts, wrong denominator, units mismatch, ITT vs per-protocol confusion.

Caught at Gate A spot-check + page/cell anchor requirement
2

Citation hallucination

Non-existent NCT ID, wrong congress abstract number, miscited regulatory decision, finding attributed to a paper that does not contain it.

Caught at Gate A URL re-verify at every cited source
3

Anomaly without literature predicate

A finding framed as surprising or counter-intuitive that the surrounding class literature would have predicted. Narrative amplifies a non-anomaly.

Caught at Gate B council novelty challenge
4

Shortcut comparator

A competitive claim that survives only because of one dimension choice. Different endpoint, time-window, or population cut reverses it.

Caught at Gate B council robustness challenge
5

Artefact-as-insight

A parsing artefact (Mode 1) becomes the centrepiece of a narrative section. The deliverable reads more interesting than a correct version would have.

Caught at Gate A headline-anchor spot-check + Gate B challenge
6

Methodology fabrication

Methodology paragraph claims searches, registry pulls, or KOL probes that were not actually performed. Reads competent, is fictional.

Caught at Gate A audit-trail check per claim
7

Frame-lock at intake

Intake scope commits to a framing that excludes an emerging mechanism or comparator now relevant. The right question goes unanswered.

Caught at Gate B council reframing prompt
Decision rule: seven of seven marked Clear ships. Any Suspected or Insufficient Evidence blocks delivery. The classification log is retained in the engagement audit trail and the per-mode result line is printed on the deliverable methodology slide.

4. Locked anchor source set

Delvant claims attach to a closed list of authoritative sources. Anything outside this set is flagged secondary, and analytical weight is reduced accordingly.

Regulatory

  • EMA (European Medicines Agency)
  • FDA (U.S. Food and Drug Administration)
  • NMPA (China), PMDA (Japan), MHRA (UK) - regional cuts
  • HTA bodies: NICE, IQWiG, HAS - payer context

Trial registries + literature

  • ClinicalTrials.gov (NIH/NLM)
  • EU Clinical Trials Register (EMA)
  • PubMed (NIH/NLM)

Congress proceedings

  • EHA, ASH (hematology)
  • AACR, ASCO, ESMO (oncology)
  • EULAR (rheumatology)

Adjacent (secondary)

  • Company press releases (flagged as sponsor-sourced)
  • SEC filings (10-K, 8-K, S-1)
  • Trade press (FiercePharma, BioPharma Dive, Endpoints) for context only

5. Staleness rule

Clinical intelligence ages fast. Delvant deliverables carry the snapshot date front-of-PDF. Snapshots older than 90 days carry a staleness banner. Re-snapshot is a separate engagement, scoped and priced.

Why this matters: a Phase III readout, a label expansion, or an EMA assessor concern can invalidate an entire competitive read in days. Locking the snapshot date prevents the reader from acting on superseded analysis. Re-engagement triggers a Gate A re-run on every Sourced claim.

6. What Delvant will not publish

Hard rules. None of these are negotiable in commissioned scope.

Exclusions

7. EU AI Act Article 4 posture

Delvant uses LLM-driven retrieval and synthesis under documented human oversight, and is classed as an AI deployer under the EU AI Act.

Disclosure stack: AI assistance disclosed on every public surface and every deliverable methodology section. Client data is never used to train external models. Founder maintains a training log under Article 4 (AI literacy). Human pre-delivery checklist sign-off is required before any PDF ships. The full posture is referenced at Article 4 of the EU AI Act.
Adjacent reading: How we work (engagement tiers + pricing) - Public sample (T2 demonstrator with methodology in practice) - Glossary (acronyms used across deliverables).

Have a question that needs this rigour applied?

30 minute scoping call. We tell you which tier fits, or that none of them do.

Book a scoping call View public sample