Architecture

Game Theory in MARIA OS

Why coupled agents converge instead of oscillate. Nash equilibrium, blocking coalitions, and mechanism design translated into responsibility gates and escalation design.

COORDINATION PROBLEM

Agent organizations are coupled utility systems

The MARIA OS problem is not merely who is correct. It is whether each actor can pursue its own utility without making the organization diverge. Because every agent best-responds to the roles and constraints of others, the whole system must be treated as a game.

Primary Players

Planner

Maximizes exploration and speed

Can overrun constraints if left unchecked

Architect

Maximizes boundaries and replayability

Can reduce throughput if it hardens too much

Operator

Maximizes field results and recoverability

Local wins can damage the global system

Local Utility
U_i(r_i | ρ_t, D_t) = fit_i + reward_i - conflict_i - gatePenalty_i

Each player chooses a role r_i and maximizes utility under the current role assignment ρ_t and governance density D_t.

Where the Game Appears
speed vs quality
local reward vs global safety
exploration vs accountability

Key Point

Workflows only define order. Game theory explains what happens when multiple optimizers collide. MARIA OS matters because coordination cannot depend on goodwill alone.

PAYOFF MATRIX

Governance does not merely persuade behavior. It reshapes the payoff landscape.

A MARIA OS gate is not a checklist added after the fact. It is mechanism design: unsafe choices receive penalties and safe coordination becomes strategically superior. The matrices below show that shift in the smallest possible two-player game.

Ungoverned Game

Base game with only short-term utility

Architect: open gate
Architect: tighten gate
Planner: push fast
(4,4)

Fast but unstable. Dangerous equilibrium.

(2,5)

Architect protects the system, but friction is high.

Planner: gather evidence
(5,2)

Planner wins locally while boundaries erode.

(3,3)

Safer, but weak against short-term incentives.

Each cell is a payoff pair (Planner, Architect). The important part is not the absolute number, but how gate design changes which cell becomes equilibrium.

Governed Game

Game after adding gate penalty λ

Architect: open gate
Architect: tighten gate
Planner: push fast
(1,1)

High risk destroys the short-term gain.

(1,4)

One-sided caution does not produce stable throughput.

Planner: gather evidence
(4,1)

Evidence alone is not enough without structural gating.

(5,5)

Evidence-first becomes the stable equilibrium.

Each cell is a payoff pair (Planner, Architect). The important part is not the absolute number, but how gate design changes which cell becomes equilibrium.

Proposition
λ > Δ_short-term ⇒ (Evidence, Tight Gate) ∈ NE

If gate penalty λ exceeds the unilateral short-term deviation gain Δ_short-term, the evidence-first / tight-gate profile becomes a Nash equilibrium.

Academic Reading

Nash (1950): equilibrium existence.

Mechanism design: redesign payoffs so desirable equilibria emerge.

MARIA OS: GateScore and escalation perform that payoff surgery.

EQUILIBRIUM SHIFT

This state becomes that state

The key is not persuading agent personalities. It is replacing the attractor of the game. MARIA OS uses gates and escalation to make the unsafe short-term equilibrium unattractive and the evidence-first equilibrium dominant.

Before Governance

Short-term utility attracts the unsafe equilibrium

Rush x Open Gate(4,4)

Both players gain in the short term, so an unsafe profile can become equilibrium.

Preferred Equilibrium
Rush x Tight Gate(2,5)

The Architect protects the system, but the Planner still has incentive to deviate.

Evidence x Open Gate(5,2)

Evidence helps, but weak gating leaves an unsafe shortcut in the game.

Evidence x Tight Gate(3,3)

Safer, but too weak against short-term payoff pressure.

Gate redesign

After Governance

The safe equilibrium becomes the attractor

Rush x Open Gate(1,1)

Gate penalties erase the unsafe short-term gain, so this profile stops being attractive.

Rush x Tight Gate(1,4)

One-sided caution still fails to balance throughput and legitimacy.

Evidence x Open Gate(4,1)

Evidence-first is not enough if the gate remains structurally weak.

Evidence x Tight Gate(5,5)

Evidence-first plus tight gating becomes the desirable stable equilibrium.

Preferred Equilibrium
Previous attractor
Rush x Open Gate

Local speed wins, so the unsafe short-term equilibrium is selected.

Operators of the shift
GateScore penalty λ
Evidence requirement
Escalation credit
New attractor
Evidence x Tight Gate

GateScore and escalation make the evidence-first equilibrium strategically preferred.

NASH EQUILIBRIUM

Stable specialization can be defined as a fixed point

A Nash equilibrium is a state where no agent has incentive to change roles unilaterally. In MARIA OS this becomes a fixed point of the role assignment vector. Governance density must be high enough relative to interaction strength so the equilibrium converges instead of oscillating.

Best-Response Map
BR(ρ_t) = (argmax_r U_1, ..., argmax_r U_N)

The simultaneous map where every agent recomputes its best role.

Stability Condition
λ_max(A_t) < 1 - D_t

If the interaction strength λ_max is dominated by governance density D_t, cycling is replaced by stable specialization.

ρ* = BR(ρ*)

Equilibrium is not inactivity. It is a dynamic balance in which no one needs to switch roles under the current boundaries and incentives.

Reading Convergence
k0
88%
k1
66%
k2
49%
k3
35%
k4
24%

If the role-change magnitude shrinks each iteration, the team is approaching equilibrium. If it amplifies, governance is too weak.

01

Compute best responses

Each agent chooses the role that maximizes utility given the current roles of everyone else.

02

Search for a fixed point

A role assignment ρ* is a Nash equilibrium when no single agent can improve by deviating alone.

03

Impose a convergence condition

When the spectral radius of interaction stays below governance damping, the equilibrium becomes stable.

COALITIONS & ESCALATION

If the equilibrium is wrong, change the game itself

An equilibrium can exist and still be unacceptable for company values or audit constraints. MARIA OS overlays gates, evidence requirements, and escalation rules on top of local payoffs so undesirable equilibria become non-executable.

01

Peer negotiation

Local conflicts are resolved at the same layer first. If they fail, the disagreement is preserved and escalated.

02

Universe gate

GateScore and responsibility boundaries change the payoff itself, not just the message between agents.

03

Human escalation

Even if an equilibrium exists, humans intervene when that equilibrium is unacceptable for the organization.

Blocking Coalition

A blocking coalition is a coordination pattern that is attractive to a subset of players but dangerous for the organization. MARIA OS detects it through responsibility gates and prioritizes system safety over local payoff.

payoff'_i = payoff_i - gatePenalty_i + escalationCredit_i
MARIA Escalation Rule
Escalate with the conflict preserved. Never smooth it away.
Terminate peer negotiation when responsibility boundaries cross, then hand off to a Universe Gate.
Humans enter as mechanism designers, not as ad-hoc answer generators.
MARIA OS IMPLEMENTATION

Game theory becomes runtime structure

The value of this page is not to add more theory vocabulary. It is to map strategy spaces, equilibria, and mechanism design into concrete runtime artifacts: candidate generation, GateScore, and responsibility escalation.

Strategy space
Candidate set C_t / role set r_i
The decision graph constrains what can actually be chosen.
Best response
Planner / Architect reassignment logic
Roles update in response to others and to gate conditions.
Nash equilibrium
Stable role assignment ρ*
Only configurations with no profitable unilateral deviation remain.
Mechanism design
Gate penalties / approval rules / escalation budgets
Payoffs are redesigned so desirable equilibria become dominant.
Blocking coalition
Conflict cards + responsibility escalation
Locally beneficial coalitions are stopped in an auditable way.
Related Theory Pages

Game theory does not stand alone. It becomes operational only when combined with stability proofs, quality gates, and team design.