Boundary
Simulation and evaluation infrastructure for governed AI systems.
Boundary is Deep Bound Research’s simulation substrate — a framework for generating controlled scenarios, tracing agent rollouts, scoring behavior against policy benchmarks, and producing replayable evaluation records. Not a testing tool. An evaluation infrastructure.
Boundary is not publicly released. Current public material covers simulation philosophy, pipeline architecture, and public-safe evaluation abstractions only.
Scenario specifications, scoring rubrics, benchmark suites, and internal simulation mechanics remain private until reviewed.
Research Scope
One of the five core flagship systems of Deep Bound Research, alongside Ex1, Plateau, StrongHold, and Cerberus. Boundary embodies the “test” dimension.
You cannot govern what you cannot evaluate.
Boundary is built on the thesis that reliable AI governance requires high-fidelity simulation — environments that can expose failure modes before they appear in production.
Systems Must Be Tested
Agents cannot be evaluated by final outputs alone. The trajectory matters — what actions were taken, in what order, under what constraints, with what evidence.
Environments Matter
An agent evaluated only in clean, ideal conditions will fail in production. High-fidelity simulation environments expose the failure modes that matter before they occur in the field.
Replayability Is Governance
A simulation that cannot be replayed is not an evaluation tool — it is a black box. Replayable traces allow policy changes to be tested against the same scenario without re-running the real risk.
Traces Create Evidence
Every simulation run produces a structured trace: actions, decisions, artifacts, scores. Traces are the evidence base from which governance policies are derived and validated.
Simulation Pipeline
A public-safe view of Boundary’s five-stage simulation pipeline — from scenario generation through replayable archive.
The pipeline separates four explicit layers: WorldSpec environment generation, rollout-job execution, the Trace Lake of execution records, and evaluation scorecards — with a governed training / SFT export at the end. Agents run within the simulation runtime where their every action is traced. Traces feed an evaluation engine that scores behavior against governance benchmarks, and all records are archived as replayable experiments so systems are evaluated before production exposure. The replay loop allows policy changes to be tested against archived scenarios.
What Boundary measures.
WorldSpec Generation
Environments are generated from WorldSpecs — structured specifications covering environment type, agent setup, adversarial conditions, tool availability, and constraint landscape.
Rollout Jobs & Trace Lake
Agents execute as rollout jobs, with every action logged at the action level into the Trace Lake — decision context, tool calls, intermediate artifacts, and state transitions.
Evaluation Scorecards
Rollouts are scored into EvalScorecards against policy benchmarks covering safety compliance, task completion, evidence quality, recovery behavior, and authority boundary adherence.
Training / SFT Export
Scored traces are exported as structured datasets for policy review, model evaluation, benchmark comparison, and training or SFT pipelines. Export formats are governed and visibility-reviewed.
Replayable Records
Every experiment is archived as a replayable record. Governance teams can re-run the exact same scenario under a modified policy to measure the effect of a change.
Failure Analysis
Boundary is designed to surface failure modes: authority violations, unrecoverable states, evidence gaps, and policy drift. Failure is a first-class output.
The Boundary ecosystem.
Boundary connects with systems that run long-horizon evaluation tasks, apply structured multi-head review, archive trace datasets, and interface with the public workspace product.
Published research artifacts.
Public-safe artifacts associated with Boundary — technical notes on simulation methodology, evaluation doctrine, and trace architecture.