Pose a research question. A budget-bounded research agent builds a belief ledger — reading real sources, verifying claims, and recalibrating its own confidence — which you can steer live (directions, scope, assumptions) as it runs.
budget $
Experiments default to gpt-5.4-mini.
Your investigations
loading…
SRA
Stanford Research Agent
151 nodes · 383 papers · ● live
Home / new investigation
Switch investigation
Contents
Fork
Log & revert
Compare / Diff
Inject evidence
CDH1 synthetic lethality151 nodes
Talazoparib resistance in SUM149PT86 nodes
↳ Fork: recency-weighted (your fork)live
↳ Fork: PARP-only re-investigationqueued
+ Start new investigation
click any claim chip to drill in · highlight text to ask
Agent trajectory hover a node for detail · click for the full inspector →
Click any node — search query, paper citation + findings, note, the stop-judge verdict, the digest table, or the answer.
Agent activity & cost every action the agent took, what each cost, and how the belief ledger evolved
Loading…
expandexplorepaper-minesynthesizeyour nodes
loading…
—
Papers
Sorted by reads × claim supports. Click any to open in evidence view.
Compare investigations
→
Inject evidence
Drop a PDF here or paste a DOI / PMID / PMC ID below
Upserts the paper, creates a digest, and inserts a paper-mine node. Originals are read-only — recommended to fork first.
Creates a new hypothesis, plus a user expand node that supports it. If refines is set, also adds a refines link to the named claim.
Adds a contradicts row against the named hypothesis. The claim's status becomes contradicted and its confidence chip turns red across the UI.
Same as a claim but tagged as private lab data in the node rationale.
Fork investigation
Branch a copy of the current investigation. Both forks remain live; you can compare them side-by-side and merge later.
Fork from current state ⇨ recommended
Clones every node, paper, and claim at this exact moment. The fork inherits your active filters and user-injected evidence. Cheap (~$0, no re-run).
Fork from an earlier round — "rewind & explore"
Rewind to round of 25. Useful when a recent nudge took the investigation somewhere unfruitful and you want a do-over with different weights.
Fork a claim into a competing hypothesis
Pick a high-traffic claim (e.g. "AKT3 is universal") and spawn a parallel sub-investigation that assumes the opposite. The planner is constrained to test the counter-hypothesis.
Three-way merge view
Open a comparison surface for two existing forks and the main investigation. Resolves claim-level conflicts row by row.