Research beneath and beyond the product line
Kalu Kodebegan with a source-recovery problem, but the work exposed reusable systems for capture, dependency provenance, runtime grounding, deterministic mutation, change verification, and evaluation. Some have become products; others remain research until their independent boundary and market demand are clear.
Capture and replay
A raw coding agent does not begin with a trustworthy local copy of a dynamic application. Capture research preserves assets, routes, network requests, storage, lazy chunks, service workers, and realtime behavior as a reusable donor and replay workspace.
Client-side exposure
A shipped client application exposes more than its visible screens. Authorized teams may need to understand which dependencies, endpoints, business rules, tokens, identifiers, or sensitive assumptions are recoverable by an observer—and which logic belongs behind server authority.
Runtime and visual grounding
Engineers often know which button, component, mesh, draw call, or visual defect they care about, but not which generated source owns it. Grounding research joins DOM, framework runtime, scene, WebGL, telemetry, and source evidence into bounded source candidates.
Dependency fingerprinting and demixing
Bundled applications mix authored logic, third-party packages, modified vendor code, and bundler runtime. The research identifies what shipped, estimates provenance, finds package-owned spans, and plans evidence-gated externalization.
Longitudinal evaluation
A product designed to govern AI changes must be evaluated more rigorously than a selective demo. We preserve model, harness, task, treatment, environment, cost, raw events, independent grader results, catches, false blocks, and campaign identity.
Take the next step
The journal is where we publish the reasoning, experiments, failures, and measurable boundaries behind the research platform.
