RavenEngineering
Engineered to be tested, coordinated and seen
Raven is a working system, run every day. These pages describe the methods behind it at a public level: how ideas earn their way in, how independent parts are kept in step, and how people keep sight of all of it.
Research before change
Raven changes only when evidence supports the change. Research follows a fixed sequence, and an idea has to pass every step in order. Most ideas don’t, and a research process that says no often is doing its job.
Step 1: Question
State the idea, and what result would count as success.
Step 2: Preregistration
Write the test and its pass mark down before running it, so the result can't be tuned afterwards.
Step 3: Historical replay
Run the idea over long periods of recorded data, quiet stretches and turbulent ones alike.
Step 4: Baseline comparison
Compare it with a random baseline under the same conditions. If it doesn't clearly beat chance, it stops here.
Step 5: Shadow run
Run it forward on current data beside the live system, with no effect on live work, and compare outcomes.
Step 6: Decision
Adopt it through a versioned release, or record why it failed. Both outcomes go in the decision log.
What the research covers
- Data processing
- Raw series are cleaned, aligned in time and checked against their source before any study reads them.
- Historical analysis
- Studies cover long periods with different conditions, so a result can't rest on one lucky stretch.
- Simulation
- Replays apply the same rules, costs and timing the live system would face, one step at a time.
- Computational evaluation
- Results are split into periods the idea was built on and periods it never saw, and judged on the second.
- Model-assisted research
- Language models help summarise and compare results. They don't decide what passes.
- Performance evaluation
- Agents in use keep being scored, including against shadow copies of themselves, so a fading result is noticed early.
Research methods are described here in general terms. Raven’s studies, data, models, settings and results are private and are not published.
Independent agents, one coordinator
Raven’s automated work is done by separate agents rather than one large program. Each agent is simple enough to reason about on its own. The hard part, keeping them from getting in each other’s way, belongs to the coordinator.
- One owner per task
- The coordinator grants a short lease on each piece of work, so two agents never act on the same thing at once.
- Narrow, never widen
- Coordination can add restrictions to an agent. It can't remove the agent's own.
- People approve sensitive actions
- Some actions wait until a person confirms them on a phone with biometric approval.
- Safe across restarts
- Agents keep their state on disk and check it against the original source when they start, so they resume correctly after a restart or power loss.
- Everything leaves a record
- Every action, and every refusal to act, is logged with the reason behind it.
- Failures stay contained
- Each agent is its own service. One stopping doesn't stop the others.
Monitoring and control
A system that runs around the clock has to tell people what it is doing without being asked. Raven reports its own health, checks its own records and asks a person before the actions that matter.
Heartbeats
On time
Daily audit
Passed
Last release
Rollback point saved
Heartbeats, last 48 checks
| Service | State | Heartbeat |
|---|---|---|
| Data intake | Healthy | 4 s ago |
| Tiered cache | Healthy | 4 s ago |
| Coordinator | Healthy | 6 s ago |
| Agent A | Working | 2 s ago |
| Agent B | Paused by limit | 3 s ago |
| Model service | Idle | 40 s ago |
Recent events
- 09:42Coordinator renewed a lease for Agent A
- 09:40Alert sent to the team: Agent B paused
- 09:40Agent B paused: a circuit breaker limit was reached
- 06:15Daily audit finished: every check passed
- 06:02Release preflight passed, rollback point saved
- Health and heartbeat
- Each service reports in on a schedule. A missing heartbeat or out-of-date data raises an alert instead of passing silently.
- Daily audit
- Every morning an automated audit compares the system's records with their original sources and lists anything that doesn't match.
- Alerts that reach people
- Problems are sent to the team's phones, with the reason and the affected service named in the alert.
- Release discipline
- Every change passes a preflight check before release and leaves a rollback point behind it.
Engineering principles
Seven rules the team holds Raven to. Each one shows up somewhere concrete in how the system is built.
Modular by design
Each part is a separate service with one job and a defined way of talking to the others, so it can be replaced without rebuilding the rest.
Research-driven development
Changes start as written, testable ideas. Whatever can't be supported with evidence doesn't ship, however promising it looks.
Data-informed analysis
Analysis begins with checked data. If the data doesn't match its source, the analysis waits.
Coordinated architecture
Independent parts share one coordinator, one data foundation and one set of limits, so they work together rather than around each other.
Operational visibility
Anything the system does can be seen, explained and traced back to a rule, a release and a decision.
Continuous improvement
Raven changes in small, reversible, versioned steps. Decisions and incidents are written down so the next change starts from what was learned.
Reliable system design
Services are built for restarts, power cuts and missing data, and they default to doing nothing rather than guessing.
ResortOS
Developed by ResortOS
Raven is a proprietary technology ecosystem developed by ResortOS, reflecting our work in advanced software engineering, quantitative research, intelligent data systems and automation architecture.
ResortOS builds software products, AI systems and client projects. Raven is the company’s own long-running engineering project: the place where its research, data and automation work is built, tested and run every day.
Talk to us about Raven
For questions about Raven, or about working with ResortOS on research, data or automation software, write to the ResortOS team.
Raven is used inside ResortOS. It isn’t offered as a public service, so this site can’t set up accounts or access.