Evidence sources and limits
What it can retrieve
Retrieval has four layers, searched from the inside out:
| Source | Content | Used for |
|---|---|---|
| Internal knowledge base | Your uploads plus accumulated domain knowledge | Searched first, before any external source |
| General web search | Public pages and news | Filling in outside facts, finding leads and primary sources |
| Literature and registry data | Biomedical literature, clinical trial records, patent literature, filing and registration data | Mechanism evidence, human evidence and dosing precedent, regulatory status, precedent and design-around |
| First-party e-commerce and social data | Product listings, price tiers, creator rankings, social posts | Verifying market facts firsthand |
Formulation and ingredient work additionally draws on specialized sources such as physicochemical prediction, safety and registry data, and supply/retail availability; industry news and trend work also uses one narrow corpus: Chinese cosmetics-industry WeChat official accounts (9 accounts, roughly the last 90 days, Chinese-language text matching). With a corpus this narrow, few matches is normal, and no match never means “none exist”.
Sources are graded
Every retrieved item carries a source grade, which decides what it may be used for:
- Citable primary sources — regulatory and registry publications, peer-reviewed literature, vendor and official pages: these may support a number in the body text.
- Neutral sources — ordinary web pages and media: usable, but must trace back to the primary source they cite.
- Leads only — report-aggregator sites (catalog pages offering little beyond “market size / share / forecast”), press-release wires, and self-published content or UGC: kept, not dropped, but they never support a number in the body text. Their value is telling you what to go check, not what the answer is.
Grading only downgrades, never upgrades: dropping one real source costs a perspective, but letting one bad source through puts it in the conclusion.
What it cannot retrieve
The platform can run general web search, but things beyond public pages are still out of reach: paywalled database full text, internal enterprise systems, offline channel data, and unpublished clinical or regulatory material. The Chinese industry WeChat corpus covers only 9 accounts and roughly 90 days.
Such gaps are written into the report’s methodology-and-limits section rather than filled with something that merely sounds right. When a report says information could not be obtained, that is deliberate — it is more useful than a fabricated number.
How to treat the conclusions
- Key claims carry an evidence grade. Strong and weak evidence are not written in the same voice.
- Citations are traceable. Reports ship with a bibliography, and the mapping between citations and claims is self-checked (dangling numbers, cited-but-unused and used-but-uncited sources all get flagged).
- There is a pre-delivery audit. After drafting, a delivery audit checks structural completeness, dangling citations, and gap labeling. When it finds problems it degrades and warns rather than blocking — you get a report with warnings, not nothing.
- Computational predictions inform ranking, not proof. Docking scores, similarity, and connectivity analyses order candidates; they do not establish activity.
Output is for R&D reference. Verify the primary evidence behind any decision you act on.
Orphic