Research
Researching how intelligent systems can be measured, improved, and trusted with real work.
Research at Syntur is about evidence: measuring what an agent actually did, testing whether a model can learn the right behavior, building somewhere to try work safely, and learning from what happened. Most of it is early, and each piece says how early.
Focus areas
Four questions, each at a different stage.
Every area names its stage: a direction is intended and not built; a hypothesis is being tested; a prototype exists and is used internally; a validated result has evidence behind it. There are no validated results yet, and nothing on this page should be read as one.
Agent evaluation
Can a change to the system be measured before it reaches anyone?
Eval sets and harnesses exist for the router; whole agent runs are next.
Training and data
Can a small open model hold Varra’s response format without a
long prompt? Tooling exists to test it offline. No trained Syntur model is in use.
Simulation and environments
Where can agent work be exercised safely before it touches a real
business? Today the answer is ordinary testing.
Learning from outcomes
Can what actually happened improve how the next piece of work is
handled, inspectably and reversibly? Outcomes can be recorded; nothing learns from them.
Current work
What exists in each area today, and where it stops. Labeled Internal tooling when it exists and only the team uses it, Active development when it is being built now, and Future direction when it is intended and not built.
01
Agent evaluation measuring a system before it changes anything
Internal tooling Active development
What exists is internal: eval sets and harnesses for the router, covering routing decisions, answer quality and a shadow mode that records what the router would have done without applying it, run offline and in continuous integration. Acceptance tests bind the agent-run and Assurance foundations to each other. Evaluating a multi-step agent run end to end is active work.
There is no public evaluation service and no benchmark result being claimed. An evaluation says where something stands; it does not by itself make a model or a system better. Agent evaluation →
02
Simulation and environments somewhere to test work before it touches a business
Future direction
The direction is environments in which agent work can be exercised against realistic systems and documents, safely and repeatably, before it runs against anything real. What exists today is ordinary testing: database-backed integration tests and bounded simulations inside individual projects. Those are tests, not a simulation platform.
03
Training and data offline preparation, not a trained model
Internal tooling
Offline tooling exists for preparing a dataset and running a fine-tuning experiment on a small open model, to test whether it can hold Varra’s response format without a long prompt. It teaches behavior, not facts; facts stay in checked reference data.
There is no approved production training data, no approved trained Syntur model, and no proprietary model integrated into Varra. Nothing retrains itself, and no model is trained on customer conversations. Varra runs on Anthropic’s Claude models.
04
Reliability systems that fail safely and recover honestly
Active development
Much of the infrastructure work is reliability work: records and spend accounting that survive a restart, bounded leases, deadlines and timeouts on every run, explicit recovery rules, and failure paths chosen deliberately — the router fails open to an ordinary answer, while the paths that can act are built to stop when a required check cannot be completed. This is active development on foundations described on Infrastructure.
05
Continuous improvement learning from what actually happened
Future direction
The direction is for recorded outcomes — an approval overridden, a figure corrected, a check that failed — to improve how work is handled, in a way that is scoped to the company it came from, inspectable and reversible rather than baked into weights. Today an outcome can be recorded against a Work Item by hand, and nothing consumes it. No routing, model or behavior changes because of it.
Publications and notes
No papers or benchmark results have been published.
When research changes something, it is recorded as a dated entry on News, taken from the build status log.
A skill-routing layer, with its own evaluation set
Development direction
The themes the next stretch of work is organized around, each with its maturity. The order is a dependency order, not a schedule: verification comes before coordination, and both come before anything is allowed to act on its own. No dates are given because none would survive contact with a company this size.
- Evaluation Internal tooling — Widen evaluation from the router to multi-step agent runs, so a change can be measured before it reaches anyone.
- Training Internal tooling — Keep fine-tuning experimental and offline until a tuned model demonstrably holds the contract and the data behind it is approved.
- Simulation Future — Build environments to exercise agent work safely before it runs against real systems.
- Infrastructure Building — Give the run, Assurance, approval and Relay foundations a product surface, one at a time.
- Execution Building / future — Widen where approved work can be carried out: connectors first, then a sandbox, then browser, then desktop. Always behind checks and an approval.
- Reliability Building — Keep every path that can act bounded, recoverable and fail-closed as capability grows.
- Agents Foundation built — Move from durable run records to agents a person can start, watch, pause and approve inside Varra. No autonomous background operation is planned ahead of that.
- Assurance Building — Widen what can be checked: from values against reference tables, to arithmetic and consistency, to a claim against its source document.
Where each component actually stands is on build status.