KISS — Knowledge Infrastructure for Scientific Simulation

An agent-actionable scaffold that externalizes operational expertise so AI agents can reliably run, check, and extend process-based scientific models.

Explore the live catalogue · Build KI with KDT · Read the paper · Desktop Windows · Desktop Linux · Desktop Mac

Knowledge dissection converts operational expertise into agent-usable Knowledge Infrastructure. Figure 2 from the KISS manuscript.

What is Knowledge Infrastructure?

Running a scientific model requires more than knowing its equations. Practitioners also know how to prepare model-specific inputs, how to decide whether each intermediate state is scientifically plausible, and how to recover when a run fails silently or indirectly. That operational expertise is often scattered across source code, documentation, example cases, and specialist practice.

Knowledge Infrastructure (KI) makes that expertise explicit and usable by a general-purpose coding agent. It separates operational knowledge into three complementary forms:

Knowledge type	The question it answers	Representation in KI	Role during execution
Procedural	How should this operation be performed?	Validated modelling operators	Constrain actions through deterministic tools for format conversion, unit handling, dependency resolution, and execution
Evaluative	Is the workflow still correct and scientifically plausible?	Staged domain protocols	Gate progress using spatial and temporal consistency, physical bounds, mass balance, parameter ranges, and output checks
Diagnostic	Why did the run fail, and how should it recover?	Diagnostic recovery mechanisms	Map symptoms to diagnoses and validated remedies, then repair and rerun

Reliable simulation emerges when an execution path satisfies all three constraints. Without procedural knowledge, agents perform invalid operations. Without evaluative knowledge, erroneous states propagate. Without diagnostic knowledge, agents cannot recover from silent or indirect failures.

From tacit expertise to an executable scaffold

The Knowledge Dissection Toolkit (KDT) extracts operational expertise from model source code, documentation, examples, datasets, and expert practice. It packages that expertise as a self-contained KI that an agent can:

• read to understand the model, its dependencies, and its operational sequence;
• call to execute validated, model-specific operations; and
• query to check intermediate results or recover from failures.

KISS leaves the governing equations untouched. The agent operates the original process-based model; KI changes the operational interface around it.

Model sources ── knowledge dissection ──► KI package ── agent execution ──► valid or repaired simulation
                                               │
                         ┌─────────────────────┼─────────────────────┐
                         │                     │                     │
                    Procedural            Evaluative            Diagnostic
                  constrain action      check plausibility      recover & repair

What this resource provides

The paper introduces four connected resources: