Facthink
Facthink
You bring a claim or research goal, your expertise and your topic priorities. They seed a shared claim graph. Four stages, run by seven specialised agents and a set of fixed rules, read from that graph and save their results back into it.
One timing wrinkle: the stages are numbered the way this page explains them, but in the running system the evidence check ③ happens early: every link with short-listed evidence is verified right after ①, before any hypothesis is drafted.
Seven specialised agents do the reading and writing. Everything that decides — which verdict an edge gets, which hypotheses pass, how they rank, what gets saved — is a fixed rule, so a run can be replayed and audited step by step.
Your expertise steers the proposing stage, so what is mined and proposed stays close to your interests.
The Extraction Agent turns your claim or research goal into a small graph: concepts joined by directed links, each recording a relationship a source asserts. Every link starts out unverified. Your topic priorities are recorded on the matching concepts.
The Research Synthesist works in one ongoing conversation with the retrieved papers always in hand: it first mines the concepts the papers add, each tied to the exact phrase it was read from, then proposes candidate hypotheses, each with a new concept, typed links, an explicit mechanism chain and quoted source spans.
A Critic Panel then reads the whole pool: two independent critic agents each grade field novelty and establishedness, audit every mechanism step and term, and rank the proposals on their own; a third judge joins only when the two disagree. Fixed rules merge the grades into a pass gate and a ranking, and the critique goes back to the Synthesist, which repairs what was flagged and derives new candidates. The best few are then shown to you, and nothing is saved until you, or the selection rule in your profile, confirm it.
For every link with retrieved evidence, the Evidence Reviewer reads the short-listed passages in one sitting and grades each one: how it bears on the link, and how sound its study design is. It is not allowed to state a verdict. A fixed rule fuses the grades, together with the open risks and scope limits the reviewer lists, into one verdict and a confidence. A verdict is accepted only if its confidence clears 0.5 and beats the runner-up clearly; otherwise the link is recorded as insufficient evidence, an abstention rather than a refutation.
For every hypothesis you confirmed, the Experiment Designer expands the one-line sketch into a full test plan grounded in a targeted search of the methods literature: design, operationalisation, baselines, datasets and metrics, controls drawn from the graph's confounder and mediator concepts, and the outcomes that would confirm or refute it. The independent Experiment Validator grades the plan on five criteria (clarity, validity, robustness, feasibility, reproducibility) and sends it back for one round of revision. The plan is saved into the graph, bound to its hypothesis.
Every proposed hypothesis runs a three-stage gauntlet before any reach you. The thresholds and weights below are the configured defaults, echoed in each run's audit, not fixed truths.
Each scored 0–1. Most blend two ingredients: the proposing agent's rating and a hard semantic-similarity measurement.
Plus four judgments the proposing agent gives directly, with no measurement blended in:
plausibility, value of testing, importance, mechanism specificity.
Eligible only if all four hold at once. Missing any single bar disqualifies the hypothesis no matter how high the others are.
Two critic agents, separate from the proposer, each grade every proposal on its own relative to the field, not just this graph; a third judge is consulted only when the two disagree. Grades combine by the median, which resists a single outlier reviewer. Each judge also audits every mechanism step and term against the quoted sources; what it flags goes back to the Synthesist for repair.
Establishedness is a 0–1 rating of how strongly the literature already treats the relationship as a known result. It demotes the familiar; it never disqualifies.
Eligible hypotheses sort by this score, the higher the better. The best few are shown to you, and only the ones you confirm, or your profile's selection rule keeps, are saved into the graph and sent on to experiment design.