Research for consequential questions
insum investigates complex questions, builds a structured account of the evidence, and gives you the tools to examine the sources, disagreements, and limitations behind every conclusion.
The report holds the synthesis. A structured account of what the investigation found, with every passage cited.
The Atlas holds the evidence. Open any claim to read the verbatim quote, see its sources and what conflicts with it, and judge its support yourself.
Why it matters
AI can produce persuasive research without reliably distinguishing established facts, weak evidence, conflicting findings, and unsupported assertions. insum is designed around that distinction. It discovers evidence, verifies extracted claims against their sources, identifies conflicts, and gives you direct access to the material behind its reports.
Shown with every quote that supports it, and how many sources stand behind it.
Flagged wherever it appears, so a conclusion is never stronger than its evidence.
Both sides are kept, set side by side, with the sources for each.
Held or rejected before the report is written, and recorded rather than silently dropped.
What you receive
Every investigation ends in a written report and an interactive companion built from the same verified evidence.
A structured synthesis of what the investigation found, including an executive summary, thematic analysis, answers to individual research questions, a bibliography, and a claim-to-source ledger.
An interactive companion for exploring the evidence behind, and sometimes beyond, the written report. Inspect supporting quotations, conflicting findings, source concentration, research coverage, and areas where conclusions remain fragile.
Evidence Atlas
Seven views of the same evidence, each answering a different question a careful reviewer would ask. Choose a view to see it in the product.
Select a view in the screenshot to switch lenses. Strength scores rank claims within one investigation; they are not probabilities that a claim is true. Screenshots are from a completed insum run used as a placeholder.
Inside a claim
Stated exactly as it was extracted and verified.
The full recipe behind its ranking: source type, independent sources, corroboration and any conflict penalty. A ranking aid, not a probability that the claim is true.
The passage the claim rests on, word for word, with its source and source type.
Findings that pull the other way, each with its own support, so disagreement stays visible.
Related findings that support it or add context.
The workflow
Nothing is researched until you approve the plan, and you can change direction while the work is under way.
insum performs preliminary reconnaissance and suggests research angles. You specify what matters and what to exclude.
insum develops a Research Brief and a detailed investigation plan, with research questions and acceptance criteria.
You revise priorities, questions, exclusions, and depth before authorizing research.
Parallel workers discover sources, extract and verify claims, identify disagreements, and pursue gaps in coverage.
You receive a structured report and Evidence Atlas. Research can also be steered during execution without necessarily restarting.
While an investigation runs, you stay in control of it.
Under the hood
Four connected stages turn a question into evidence you can check. Each one records what it did, so the result can be audited.
Combine a reusable source library, live web search, and specialist scholarly, legal, governmental, medical, and scientific repositories.
Confirm source quotations exist, use semantic verification to assess whether they support claims, and hold or reject unsupported records.
Build relationships between claims to identify corroboration, contradiction, tension, dependencies, and context.
Check completed prose against cited evidence, including numerical assertions and unsupported statements.
Open-weight models handle planning and writing. Narrower tasks, such as checking whether a passage supports a claim, are assigned to smaller models trained and calibrated for that one judgment against human-labelled examples.
Pages are stored as captured, with retrieval time and location. A quotation can always be checked against the exact version the system read, even if the live page later changes.
Grounding and precision measure whether each statement is supported by its cited passage, without addition. Completeness and correctness measure coverage against a reference and are reported separately, never as a measure of truth.
Built for consequential research
For work where a conclusion has to survive scrutiny from a colleague, a client, a reviewer or a court.
Fact-intensive investigations, complex regulatory questions, and research requiring detailed source examination.
From a completed runHow do different states address tax exemptions for data centers? What is the current local, state and federal legislation?
Synthesizing literature, studies, datasets, and conflicting scientific findings.
Example questionWhat do trials report about long-term outcomes of a treatment, and where do they disagree?
Exploring unfamiliar fields, identifying research directions, developing literature syntheses, and examining contradictory evidence.
Example questionWhich open questions in a field are under-studied, and what evidence would settle them?
Investigating industries, technical questions, market developments, and competing claims.
Example questionWhich competing claims about a technology's readiness are backed by primary evidence?
insum supports professional research. It does not replace legal, medical or other professional judgment, and its reports are a starting point for that judgment.
Deployment and privacy
insum uses open-source models and supports deployment on customer-controlled infrastructure. With sufficient computing capacity and locally available data, organizations can run the research stack without sending sensitive research material to external inference providers.
The quickest start. insum operates the research stack on infrastructure it controls.
The full stack, including models, runs on infrastructure you own or control.
Live source discovery requires network access to the sites being searched, so a deployment that uses it is not offline. A deployment limited to locally available data can run without external connections.
See how insum turns complex questions into structured research, and gives you the evidence to examine its conclusions.
