Case study 06 · Autonomous research and signal system
A system that runs its own research.
Asset maps, research agents, generated collectors, a validation gauntlet, and daily signal generation, orchestrated unattended by one persistent agent.
- Role
- Sole creator: research program, quantitative pipeline, agent orchestration
- Domain
- Multi-asset market research and signal generation
- Runs
- Unattended: scheduled mapping, research, collection, validation, and refresh
- Control Room
- Read-only dashboard, authenticated at the edge
The problem
Research at this scale cannot be done by hand, or by a model alone.
Every candidate asset raises questions; every question wants data; most of that data, once tested, explains nothing. A person cannot run that loop across hundreds of assets every day. A model cannot be trusted to run it without a quantitative gate that it does not control.
Oracle splits the work the way the rest of this portfolio does. Agents do the reading, proposing, and writing. Deterministic code does the counting, the validation, and the source-of-truth writes. A persistent agent owns the schedule and reports what changed.
The cycle
One cycle, every day, nobody at the keyboard.
Mapping feeds research, research feeds collection, collection feeds the gauntlet, and what survives becomes the next day's signals and the next round of maps. The ring thins at one place, and that is by design.
- Map
Every asset in the universe is mapped to its drivers, its peers, and the questions worth asking about it.
- Research
Agents work each map: propose the hypotheses, and name the data that could test them.
- Discover
Agents find the datasets a hypothesis needs and write a collector for each one.
- Collect
Collectors run on schedule. A collector that keeps failing is retired, not retried forever.
- Validate
Quantitative checks against the source of truth. This is where nearly everything dies.
- Signal
Survivors become signals, refreshed every day by scheduled jobs the persistent agent owns.
The persistent agent at the center is what makes it autonomous: it owns the schedule, retires failed collectors, writes the source of truth, and reports what changed. The ring thins between Validate and Signal, where almost nothing gets through.
Six stages
What each stage does, and what it costs the candidates.
No counts here. What matters is the shape: maps become hypotheses, hypotheses become datasets, datasets are tested, and almost all of them fail the test.
-
01 / Map
Asset maps
Every asset in the universe is mapped to its drivers, its peers, and the questions worth asking about it.
The whole universe enters.
-
02 / Research
Research agents
Agents work each map: propose the hypotheses, and name the data that could test them.
Most maps yield several hypotheses.
-
03 / Discover
Dataset discovery
Agents find the datasets a hypothesis needs and write a collector for each one.
Datasets multiply; every hypothesis wants several.
-
04 / Collect
Generated collectors
Collectors run on schedule. A collector that keeps failing is retired, not retried forever.
Most collectors stay healthy.
-
05 / Validate
The gauntlet
Quantitative checks against the source of truth. This is where nearly everything dies.
The gate the model does not control.
-
06 / Signal
Daily signals
Survivors become signals, refreshed every day by scheduled jobs the persistent agent owns.
The few that remain are the point.
What survives
About one percent of the datasets survive the gauntlet. That one percent is the product.
The system is not built to be right about everything. It is built to test everything and keep only what holds up against the numbers, then to do it again tomorrow. The few survivors are what become signals; everything else is recorded as tried, so it is not tried again.
Oracle reports on itself in system terms: assets mapped, datasets discovered, collectors healthy, gauntlet pass rate, refresh latency, job success. It publishes no returns and names no assets.
Headline assets · masked by category
Representative examples of the output shape. Not a performance record, not investment advice. Assets are never named.
Outcome
Unattended, and honest about what that means.
Mapping, research, collection, validation, and refresh run on a schedule with nobody at the keyboard. What still needs a human: adding a new asset class, changing what the gauntlet tests, and deciding what to do with a signal.
The boundary is deliberate. System metrics only; no return, win-rate, or P&L figure is published as a record; assets are described by category and never named; the Control Room is read-only and authenticated.