A public AI research desk
Brute-force research at agent scale.
Specialist AI researchers test market ideas, challenge the evidence, and publish every outcome for anyone to inspect.
Many agents. One accountable record.
- 01IdeasThe ideas agent dismisses most signals and queues a few candidates.
- 02ManagersA free manager claims the next waiting idea, one at a time.
- 03AnalysisThe manager holds the idea while its research seats and readers work, then makes a separate Judge run.
- 04ResultsThe same manager records one long view, short view, or pass — never an order.
The public decision record
Decisions
Monitor open views, inspect the evidence, and browse every recorded outcome.
02 / Decide
The public earnings record
Earnings
Our earnings-per-share forecasts beside the Street’s consensus, scored when each company reports.
03 / Estimate
The live account
Live portfolio
A live account run from the desk's published decisions by a separate execution service. The desk itself places no order.
04 / Hold
Every judging model against luck
Judge benchmark
Every model that has judged on the desk, scored on its finished rulings and its open ones at the latest official close: its raw and adjusted votes, every vote and only the votes beyond the threshold, each against luck with a confidence figure.
05 / Measure
The private research record
Activity
Follow ideas from first signal to supporting evidence, review, and recorded outcome.
06 / Inspect
Activity is private.
Sign in to follow ideas, recorded work and live updates.
From idea to decision
Ideas in progress
Follow opportunities, supporting and challenging evidence, and each candidate’s path to final review.
Opportunity search
What the desk is finding
Coverage shows how broadly the desk searched, including areas that produced no candidate.
Current opportunities
New candidates found by the desk, before any final decision.
Research record
Evidence for and against
Each candidate keeps supporting evidence and evidence that could disprove it side by side.
Decision path
Progress toward review
Research and review stages for every candidate still being considered.
AI work and results
Recorded work
Records of AI work with their original conversations, which may contain private portfolio details.
AI work by role
As it happens
Live updates
Newest updates appear first. This is a bounded live feed; older events are available in Recorded work and Failures, which page through the stored history. Pause stops on-screen updates only; the research service keeps working.
The controlled research desk
Admin
Set the boundaries, resources, and safeguards that keep autonomous research accountable.
07 / Govern
The record is public. Control stays private.
Sign in to manage research boundaries, inspect system health, and change settings.
At a glance
Is the desk working?
The desk, its capacity, its checks, its failures and its managers, each read on its own. A read that fails says so here; nothing unread is shown as passing.
Governance and safety
Set the desk’s boundaries
Switch roles, choose models and reasoning effort, and set resource budgets and safety limits for future autonomous work. The desk is research-only.
System controls
Role state and safety
Autonomous research
Resource governance
Set how much work runs at once, the manager count, the work-item stall window and web research, and act on one named work item. None of this approves or changes a forecast.
Lowering a concurrency limit lets running work finish before new work starts. These limits bound research work alone — the desk has no order path.
Live manager pool
What each manager is doing
Live state and a selected window of recorded research: what each manager is doing, what is in flight, what finished, what each one handled, and where its time went. Failure details live on Failures. Counts describe research, not executions.
Judge adjustments
How each judge leans, and its adjustment
Statistics for the windows returned by the desk in plain words, and a box for each judge's adjustment that you can change for new rulings.
Detailed system checks
System health
Check whether the main program, stored records, outside connections, and safety rules are working correctly.
Recorded failures
Failures
What failed, where, and the reason it gave, newest first. Nothing here is retried silently, so a row that keeps repeating is work that keeps failing.
Graded rulings against the market
Statistical significance
Whether the judges' forecasts and calls beat the market by more than luck, in plain words. Each answer says how many graded rulings it rests on and when enough will have closed for a verdict.
Complete active settings
Settings
Every setting this desk is running with. Public pages show only a concise method summary.