SignalRivet Research Notes
Technical notes from the systems we actually build and test.
A working-paper series on agent reliability, model evaluation, operating controls, and the architecture around AI-assisted work. Notes separate measured results from interpretation and state their limitations explicitly.
Current papers
Research Note Series
Notes are published when there is enough evidence to make the method, result, and limitations worth inspecting.
Delegation Without Drift: Exception-Based Supervision in Human–AI Operations
A bounded retrospective case study proposing task ownership, tiered escalation, and evidence contracts. It reports isolated passes and explicit holds without claiming production autonomy or measured performance gains.
Intelligence at the Edges, Determinism at the Core
A retrospective, failure-driven case study of BOSS as an internal system. It reports bounded measurements, negative evidence, and design hypotheses; it is not peer reviewed and does not claim general superiority.
From Conversation to Control Plane
Building reliable AI agent systems with deterministic orchestration, durable state, and evidence-backed execution.
Why Reliable AI Work Needs More Than a Good Model
Governed worker completion, machine-verifiable procedures, and fail-closed semantic validation.
Publication standard
A result should be easier to inspect than to hype.
These notes are not presented as peer-reviewed scholarship. They are technical working papers that make our internal questions, methods, evidence, and limitations legible to other practitioners.
- 01Question
State the specific thing we were trying to learn.
- 02Method
Describe the procedure, tools, task configuration, and evaluation path.
- 03Results
Report observed outcomes separately from interpretation.
- 04Limitations
Say what the evidence does not establish.
- 05Revisions
Version notes when later evidence materially changes a conclusion.
Applied research