IndisputableMonolith.Verification.RecognitionStabilityAudit.RStoRL
Defines a lightweight moral state for reinforcement-learning agents grounded in Recognition Science. The state packages reciprocity skew σ, an available recognition-cost energy budget, and a value functional V = κ·I(A;E) − C_J*. Downstream RSA code imports it as the RS-to-RL bridge. The module is definitional: structures, norms, virtue actions, and feasibility predicates, not deep theorems.
claimA moral state is a triple $(\sigma, E, V)$ where $\sigma$ is reciprocity skew (imbalance), $E$ is the available recognition-cost budget, and $V = \kappa \cdot I(A;E) - C_{J}^{*}$ is the net value of action $A$ in environment $E$. Virtue actions act on this state by scaling, addition, and energy cost; temperance and $\sigma$-feasibility constrain admissible updates.
background
Recognition Stability Audit (RSA) asks whether agents whose rewards track Recognition Science costs remain stable. This submodule supplies the RS-native state that an RL loop would observe and update.
The essential quantities are taken from the RS cost layer: skew $\sigma$ measures reciprocity imbalance; energy is a budget in units of the $J$-cost $C_J$ (the unique cost forced by the Recognition Composition Law); value is mutual information reward minus residual $J$-cost. Constants such as the tick $\tau_0$ enter only as the ambient RS time quantum.
Sibling definitions introduce MoralState, vector-like operations (zero, norm, scale, add), a finite menu of virtue actions with names and interpretations, an energy-cost map, and predicates for temperance and $\sigma$-feasibility, plus an LA-completion marker.
proof idea
This is a definition module, not a theorem module. It declares the moral-state structure, algebraic operations on it, virtue-action labels and their interpretive map, energy-cost accounting, and feasibility predicates (temperance, $\sigma$-feasibility, LA completion). No substantial proof obligations are discharged here; downstream audit code consumes these types and predicates as the RS-to-RL interface.
why it matters in Recognition Science
The parent umbrella RecognitionStabilityAudit re-exports this bridge so that RSA developments can import a single path and obtain an RL-ready moral state. Paper reference: papers/tex/Recognition_Stability_Audit.tex. By packaging $\sigma$, recognition energy, and $V = \kappa\cdot I(A;E) - C_J^{*}$ in one object, the module lets stability claims be stated as constraints on trajectories in moral-state space rather than as ad-hoc reward hacks. It sits in the verification domain: it does not re-derive $J$-uniqueness (T5) or the forcing chain, but assumes the Cost and Constants layers and exposes only what an RL trainer needs.
scope and limits
- Does not prove convergence or stability of any RL algorithm.
- Does not derive J-cost uniqueness or the forcing chain T0–T8.
- Does not specify a concrete environment dynamics or policy class.
- Does not claim numerical bounds on α, masses, or other RS constants.
- Does not implement training loops; only the moral-state interface.
used by (1)
depends on (2)
declarations in this module (51)
-
structure
MoralState -
structure
VirtueAction -
def
zero -
def
norm -
def
scale -
def
add -
def
virtueNames -
def
interpret -
def
energyCost -
def
satisfiesTemperance -
def
SigmaFeasible -
structure
LACompletion -
def
identity -
def
phiScale -
structure
AuditResult -
def
lexBetter -
structure
LexicographicSelector -
def
selectByLex -
def
filterFeasible -
structure
GibbsPolicy -
def
weight -
def
partitionFn -
def
prob -
def
freeEnergy -
def
default -
def
cool -
def
warm -
structure
EightTickCadence -
def
totalValue -
def
maxHarm -
def
sigmaClosed -
def
totalEnergy -
def
satisfiesTemperanceWindow -
def
exercisedPatience -
structure
RSEnvironment -
def
step -
def
selectAction -
def
SatisfiesConsent -
def
HarmBound -
structure
ActionConstraints -
structure
ParasiticPattern -
def
parasitismScore -
def
parasitismThreshold -
def
isParasitic -
theorem
virtueAction_norm_nonneg -
theorem
virtueAction_zero_norm -
theorem
virtueAction_scale_norm -
theorem
lexBetter_irrefl -
theorem
gibbs_weight_pos -
theorem
gibbs_partitionFn_pos -
theorem
eightTick_value_finite