Hybrid Actor Allocation Engine

Anthrocentrix Fusion

Anthrocentrix is a hybrid allocation engine. It determines whether a domain is Main-Effect Dominated, Mixed, or Interaction-Dominated and recommends allocation strategies accordingly. The platform does not assume execution-fit — it measures execution-fit.

Product position
current scientific position

Hybrid Actor Allocation Engine. Primary mechanism: Actor Quality Allocation. Secondary mechanism: Execution-Fit Optimization. Current evidence supports both actor-quality allocation and execution-fit allocation, depending on domain structure. The platform determines whether a domain is Main-Effect Dominated, Mixed, or Interaction-Dominated and recommends allocation strategies accordingly. The platform does not assume execution-fit — it measures execution-fit.

Case Context + Actor Quality + Actor × Case Fit + Policy Constraints → Allocation Recommendation

Design principle · Never report actor-aware gain as execution-fit gain. Never assume a domain is interaction-dominated or main-effect-dominated — measure the decomposition and classify the domain empirically.

Current scientific position

full evidence →
✓Actor Main Effects
Proven
✓Execution-Fit
Proven to exist
✓Interaction-Dominated Domains
Observed
✓Policy Gain
Observed
△Causal Proof
Domain dependent

Four-axis evidence

actor quality · execution-fit · policy gain · causal confidence
Actor Main Effect
Does actor identity add information beyond case context? A positive result here is necessary but NOT sufficient for execution-fit.
Actor main effects are proven across crossed actor×case datasets (EOIR, STAR, LaborSupply, Grunfeld, Arizona Open Policing).
STRONGLY SUPPORTED
Actor × Case Interaction (Execution-Fit)
Does the actor's contribution depend on the case? This is the execution-fit claim. A main effect alone does not establish it.
Execution-fit is proven to exist. MLB Statcast Umpires (1.47M called pitches, 124 umpires) shows interaction share 67.8% (95% CI 58.8%–75.9%), classified INTERACTION_DOMINATED. EOIR shows weak interaction; Arizona shows main-effect-dominated. Allocation strategy must be measured per domain, not assumed.
SUPPORTED
Policy Gain (actor-aware vs actor-blind)
Does an actor-aware routing/conditioning policy beat the actor-blind baseline in offline/modeled evaluation? Modeled lift only — not yet a causal claim.
Actor-aware allocation produces measurable downstream gains (e.g. Arizona Open Policing: +1.3 pts recall over the actor-blind baseline). Observed across multiple domains; magnitude is domain-dependent.
SUPPORTED
Causal Confidence
Is the policy gain backed by random or quasi-random assignment so it can be read causally? Without this, lift can be confounded by selection.
Causal proof is domain dependent. MLB variance decomposition supports a high-confidence interaction read in that domain; EOIR and Arizona remain associational pending quasi-random assignment.
WEAKLY SUPPORTED

Observed domain types

allocation strategy must be measured, not assumed
Main-Effect Dominated
e.g. Arizona Open Policing
Actor quality drives most of the observed actor value. Allocate on actor quality; execution-fit layer stays OFF.
Interaction-Dominated
e.g. MLB Statcast Umpires
Actor × case interaction drives most of the observed actor value. Activate execution-fit allocation.

Benchmark leaderboard

evidence cards →
DatasetVerdictInteraction shareNote
MLB Statcast UmpiresINTERACTION DOMINATED67.8% · CI 58.8%–75.9%1.47M called pitches · 124 umpires · variance decomposition
Arizona Open PolicingMAIN EFFECT DOMINATED9%+1.3 pts recall; 91% actor main effect / 9% execution-fit
EOIR JudgesWEAK INTERACTION≈30%Interaction statistically significant (p<0.001) but below practical-meaningfulness threshold

Engine Status

live · core v0.7.2-rc
Datasets Imported
3
DecisionEvents
6,429,191
Actors Modeled
12,847
Predictions
2,103,442
Outcomes Resolved
6,201,044
Experiments
2
completed

Pipeline

dataset → calibration
Stage 1
Dataset
3
stable
Stage 2
Domain Adapter
3
stable
Stage 3
DecisionEvent
6,429,191
live
Stage 4
Actor Model
12,847
live
Stage 5
Prediction
2,103,442
live
Stage 6
Outcome
6,201,044
resolving
Stage 7
Calibration
ECE 0.031
passing