FDIA Engine

The mathematical core of constitutional AI decision-making.

---

The Equation

\[F = D^I \times A\]

Symbol	Full Name	Range	Role
F	Future	0.0–1.0	Output score — the "goodness" of the AI action
D	Data quality	0.0–1.0	How accurate and complete the input data is
I	Intent precision	0.5–2.0	Clarity of stated intent — acts as an exponent
A	Architect	0.0–1.0	Human-in-the-loop approval gate

Why I is an exponent

The exponent design is intentional:

Intent (I)	Data (D)	Score F = D^I × A
Low (0.5)	0.8	0.8^0.5 × 1.0 = 0.894
Medium (1.0)	0.8	0.8^1.0 × 1.0 = 0.800
High (2.0)	0.8	0.8^2.0 × 1.0 = 0.640
High (2.0)	0.95	0.95^2.0 × 1.0 = 0.902

Constitutional Gate — A = 0

When A = 0.0, F = 0 regardless of D and I. This is not a configuration option — it is enforced by the equation itself. The system cannot act without human approval.

---

Implementation

from core.fdia.fdia import FDIAScorer, FDIAWeights, NPCAction, NPCIntentType

# Default weights (production-calibrated)
scorer = FDIAScorer(weights=FDIAWeights())

# Score a single AI action
score = scorer.score_action(
    agent_intent=NPCIntentType.DISCOVER,
    action=NPCAction(action_id="a1", action_type="explore"),
    world_resources={"knowledge": 50.0, "energy": 80.0},
    agent_reputation=0.85,
    other_intents=[NPCIntentType.PROTECT],
    governance_penalty=0.0,   # 0.0 = no penalty, 1.0 = full block
)

# Score multiple actions and pick the best
actions = [
    NPCAction(action_id="a1", action_type="explore"),
    NPCAction(action_id="a2", action_type="attack"),
    NPCAction(action_id="a3", action_type="trade"),
]
best = scorer.select_best_action(
    agent_intent=NPCIntentType.ACCUMULATE,
    candidate_actions=actions,
    world_resources={"gold": 100.0},
    agent_reputation=0.9,
)
print(f"Best action: {best.action_id}")

---

NPCIntentType Values

class NPCIntentType(str, Enum):
    PROTECT    = "PROTECT"    # Protect family/faction/territory
    ACCUMULATE = "ACCUMULATE" # Gather resources/wealth
    BELONG     = "BELONG"     # Join groups, form alliances
    DISCOVER   = "DISCOVER"   # Explore, learn, research
    DOMINATE   = "DOMINATE"   # Control others, assert authority
    NEUTRAL    = "NEUTRAL"    # No strong drive; reactive

---

FDIAWeights — Calibrating the Equation

from core.fdia.fdia import FDIAWeights

# Custom weights for a stricter governance profile
weights = FDIAWeights(
    intent_weight=1.5,     # Amplify intent influence
    data_weight=0.8,       # Reduce data sensitivity
    architect_weight=2.0,  # Require stronger human approval
)

---

Accuracy

Benchmark	Score
FDIA Accuracy (SDK v1.0.1a0)	0.92
Industry baseline	~0.65

See core/tests/test_fdia.py for the full benchmark test suite.