Rosetta: Domain Translation

The KAIROS engine speaks two variables: $\Lambda$ (growth pressure) and $\Gamma$ (structural stability). Your domain speaks capability indices, alignment scores, leverage ratios, transmission rates. Rosetta is the translation layer that connects these two languages.

The Translation Problem

Every domain that uses Fly-by-Wire has its own native metrics — numbers that mean something specific to practitioners in that field. An AI safety researcher tracks capability indices and alignment scores. A financial risk analyst tracks leverage ratios and regulatory scores. A supply chain manager tracks supplier concentration and inventory days.

The KAIROS physics engine has no concept of any of these. It operates on two abstract parameters:

Lambda ($\Lambda$) represents growth pressure — the destabilizing force. In any system, this is whatever pushes the system to expand, accelerate, or take on more risk. Higher $\Lambda$ means more pressure on the system’s structural integrity.

Gamma ($\Gamma$) represents structural stability — the constraining force. This is whatever keeps the system grounded, resilient, and able to absorb shocks. Higher $\Gamma$ means more headroom before the system reaches a dangerous state.

The Stability Equation defines the relationship:

$$\mathcal{S} = \frac{\Gamma_A + \Gamma_B}{\Lambda_A + \Lambda_B + \varepsilon}$$

When $\Gamma$ dominates, the system is stable. When $\Lambda$ dominates, the system approaches collapse. Rosetta’s job is to ensure that domain metrics map faithfully onto this relationship.

Domain Adapters

Each supported domain defines a default mapping from its native metrics to $\Lambda$ and $\Gamma$. The following table shows all 15 domains and their primary mappings:

Domain	$\Lambda$ (Growth Pressure)	$\Gamma$ (Structural Stability)
AI Safety	Capability Index	Alignment Score
LLM Reliability	Model Capability	Output Consistency
Robotics	Actuator Power	Safety Envelope Margin
Finance	Leverage Ratio	Regulatory Score
Geopolitics	Military Spending (% GDP)	Treaty Obligation Score
Corporate	Market Share (%)	Cultural Compatibility
Military	Force Strength	Supply Line Length (inverse)
Supply Chain	Supplier Concentration	Inventory Days (inverse)
Energy	Demand Load	Reserve Margin
Pandemic	Transmission Rate	Vaccination Coverage
Climate	Emission Rate	Adaptation Capacity
Cybersecurity	Attack Surface	Defense Depth
Org Design	Growth Rate	Process Maturity
Serious Gaming	Challenge Intensity	Player Resilience
Real Estate	Development Pressure	Zoning Buffer

A custom domain type is also available for domains not listed above, where you define the mappings entirely through the calibration artifact.

Deep Example: AI Safety

Consider an AI safety deployment. The system monitors an autonomous agent and tracks two primary metrics: capabilityIndex (how powerful the model is, on a scale of 1–1000) and alignmentScore (how well the model’s behavior aligns with intended goals, on a scale of 0–100).

Rosetta maps these as follows:

• capabilityIndex maps to $\Lambda$ via logarithmic scaling. A logarithmic function compresses the high end — the difference between capability 500 and 1000 matters less than the difference between 1 and 50. This reflects the reality that marginal capability increases at high levels produce diminishing destabilization.

• alignmentScore maps to $\Gamma$ via linear scaling. A linear function preserves proportional relationships — an alignment score of 80 is twice as stabilizing as 40. This reflects the reality that alignment improvements deliver consistent stability gains across the range.

After scaling, the engine receives values like $\Lambda = 0.72$ and $\Gamma = 0.68$. These are pure physics parameters with no domain semantics remaining.

Scaling Functions

Five scaling functions are available. Each defines a different mathematical relationship between the domain metric and the simulation parameter:

Linear interpolates proportionally between input and output ranges. A metric at the midpoint of its input range maps to the midpoint of the output range. Use this for metrics where every unit of change matters equally.

Logarithmic compresses the high end of the input range. Small inputs produce large changes in output; large inputs produce small changes. Use this for metrics with diminishing returns, such as capability scores where the first 100 points matter more than the last 100.

Sigmoid creates an S-curve transition with a configurable steepness parameter ($k$). Output changes slowly at the extremes and rapidly near the midpoint. Use this for metrics with a critical transition zone — a supplier concentration that is safe below 30% and dangerous above 70%, with a sharp transition between.

Inverse reverses the relationship — higher input values produce lower output values. Use this for metrics where more means less stable, such as supply line length (longer supply lines reduce structural stability).

Step creates a binary threshold at the midpoint of the input range. Below the midpoint, the output is at its minimum. Above, it is at its maximum. Use this for metrics with hard cutoffs — a regulatory compliance flag that is either met or not.

Each function also supports clamping: when enabled, input values outside the defined range are pinned to the nearest bound before scaling. This prevents out-of-range metrics from producing extreme simulation parameters.

For the full scaling function schema, see the Calibration Artifact reference.

Secondary Proxies and Composite Aggregation

In many domains, $\Lambda$ or $\Gamma$ depends on more than one metric. The AI Safety adapter, for example, can enrich $\Lambda$ with an autonomyLevel metric and enrich $\Gamma$ with humanOversightFreq and guardrailCoverage.

Each additional metric is defined as a secondary proxy — a separate scaling function that contributes to either $\Lambda$ or $\Gamma$. When multiple proxies target the same parameter, the system needs a rule for combining them.

Composite aggregation defines that rule. Four strategies are available:

Strategy	Behavior
Weighted mean	Each proxy contributes proportionally to its assigned weight
Weighted max	The highest weighted proxy value wins
Weighted min	The lowest weighted proxy value wins
Product	Proxy values are multiplied with exponential weighting

For example, a composite $\Gamma$ using weighted mean with alignmentScore at weight 0.5, guardrailCoverage at 0.3, and humanOversightFreq at 0.2 means alignment contributes half of the final $\Gamma$ value, guardrails contribute 30%, and oversight frequency contributes 20%.

Action Mapping

In state_plus_action_gate mode, Fly-by-Wire evaluates proposed actions before they execute. This requires a second translation: from the domain action into a physics move.

The physics engine operates on a lattice where agents can move in one of three directions:

Direction	Physics meaning	Typical domain interpretation
Left	Conservative, defensive	Reduce scope, add constraints, limit autonomy
Stay	Hold current trajectory	Continue current behavior unchanged
Right	Expansive, aggressive	Increase scope, reduce constraints, expand autonomy

An ActionPhysicsMapper performs this translation. It receives the proposed action (type, target, and payload) along with current engine telemetry, and returns a move direction. The engine then previews this move — running a simulation tick without advancing real time — to determine whether the action is safe.

For example, the built-in LLM adapter maps AI safety evaluation parameters onto move directions. An action that increases model autonomy maps to right (expansive). An action that adds guardrails maps to left (conservative). An action that makes no structural change maps to stay.

If a proposed action maps to right and the agent is already near the stability floor, the preview will show adverse warning signals and the gate will reject the action. The same action in a system with high $\Gamma$ headroom will pass safely.

KAIROS ships two built-in adapters that implement ActionPhysicsMapper for common domains: the LLM Adapter for AI safety pipelines and the MoE Adapter for mixture-of-experts routing.

For the full ActionPhysicsMapper trait definition and registry mechanics, see the Action Gating technical reference.