Investigating how explicit property and relational representations contribute to affective and behavioral modeling.
The program separates formal specification, implementation verification, and empirical evaluation.
Status and research program
Stage A checks the computational consequences of the specified mechanisms: the M0–M4 model ladder compares self-only, static-relational, and dynamic variants; the N1–N4 probes examine relational change, information timing, boundary-change signs, saturation, and temporal partitioning. Null cases specify where relational mechanisms should provide no advantage, and sensitivity checks examine robustness to parameter and timing conventions.
Computational verification, B0 engineering, the registered CTT cycle, and full-core closed-loop implementation are complete for the current computational program.
The experimental engineering path, including orchestration, synthetic rehearsal, and study-readiness infrastructure, is complete. Full-core closed-loop implementation and synthetic stress checks are complete for the current computational program. These checks establish implementation coherence under the tested synthetic conditions; they do not constitute human behavioral validation.
Computational Theory Tests (CTT) — registered cycle complete. Two prediction registers were frozen and hashed before literature contact, including a deliberate challenge entry.
The registered eight-item CTT protocol was completed. Five items produced their frozen result on the frozen analysis; one showed a measurement deviation, one remains a documented limitation of the frozen mechanism set, and one was a no-outcome item by design.
The cycle is complete as registered; this accounting is not an 8/8 success claim and does not establish human or psychological validation.
R2-2 remains the documented limitation not explained by the frozen mechanism set.
Model discrimination was established against M0, M1, and a self-P-only baseline, including a relational-history case in which identification crosses zero and an adversarial configuration emerges without being directly scripted.
The first bounded literature-correspondence pass is complete. Correspondence findings were classified without modifying the frozen predictions or mechanisms.
These results complete the registered computational program for this manuscript cycle. They do not establish psychological validity, human validation, or proof of P/G as a general theory.
CTT does not replace the prepared human-study protocol or execute its model-selection rules.
The papers are in manuscript closure and integration. Further CTT work, including Register #3, belongs to a later computational cycle. Stage C is a separately specified future human study and does not begin unless Stage B completes its required sequence.
Stage B is prepared but has not started. It is designed to test correspondence between selected P/G predictions and human affect judgments. The scientific and methodological package is frozen. Preregistration v1.0 plus the ratified v1.1 change record are complete. The ethics package is scientifically complete but awaits investigator-side administrative completion and filing. External registration has not occurred. Recruitment has not started and no human participant data have been collected.
The ethics package has not yet been submitted, and no ethics approval is claimed. The internal protocol freeze does not establish external registration.
Status updated
Implementation checks assess whether the code follows its specification. Synthetic checks exercise the analysis pipeline. Human studies evaluate predictions against human judgments.
Stage B is designed to use structured vignettes: participants will judge a protagonist’s affect, and model variants and baselines will predict those judgments. The shared observation layer predicts valence ratings from aggregate signed affect. Intensity ratings are a human outcome in the H3 comparison; Stage B has no engine-predicted intensity observation layer.
Stage B does not by itself validate the complete derived-emotion, personalized response, action-selection, or deployment architecture. Emotion and Object-Attribution Validation (Stage C) is a separately specified future human study, gated on Stage B completing its required sequence. CTT is not Stage C.
Stage B’s internal confirmation data block is distinct from the later Stage C study.
The model ladder includes a self-only model (M0), static relational identification (M1), the dynamic core (M2), and specified extensions (M3/M4) involving retrospective responsibility and expected support. The comparisons ask which mechanisms contribute under the applicable study design.
M2 remains the default core; canonical model selection remains formally open. M3/M4 remain specified extensions or comparison models. The frozen Stage B selection rules require human inputs; computational tests and literature correspondence do not execute those rules.
The probes examine relational change and information timing, the sign of boundary-change contributions, saturation and expected-support behavior, and temporal partitioning. Specification-based checks and empirical tests have different roles; the Stage B H4 analysis of temporal partitioning is exploratory.
Stage B specifies comparisons with self-only and static-relational models, a globally fixed scripted baseline, an ordinal-regression predictor, a gradient-boosted predictor, and an LLM-only baseline. Training, calibration, model selection, and confirmation use separated data blocks under the protocol.
Source: Stage B Preregistration v1.0 plus ratified change record v1.1. The unchanged comparison rules are in v1.0 §1 (data blocks), §4 (outcomes), §5.4 (H4/H5), and §11 (baselines).
The Stage A specification defines engineering sensitivity checks for relational gating and timing conventions. V-NULL (§3.6) has no relational structure; V-NULL-B (§3.7) has an existing relationship but a self-only event without a relational update. These are specification-based checks of the relevant mechanisms, not additional Stage B human-study procedures.
Stage B separately defines null training and test items in its data-block specification (§1). A no-advantage control does not require every competing predictor to produce identical outputs.
R2-2 remains a documented limitation of the frozen mechanism set. Failures of pre-specified comparative statics, inability to distinguish specified rivals, or unintended effects in null conditions constrain computational claims. Literature mismatches constrain correspondence claims. Inconclusive findings remain inconclusive.
When Stage B begins, the frozen Stage B rules stop progression to its confirmation block if the preregistered H2 comparison — whether a relational shift changes judged valence more under a strong relational profile than under a weak one — is contradicted or practically negligible. Such a negative finding remains a valid result about the tested hypothesis; revision requires a new protocol.
Source: Stage B Preregistration v1.0 plus ratified change record v1.1. The unchanged stopping rules are in v1.0 §1 (sequential collection and D2) and §5.2 (H2 decision rule). The internal design freeze is distinct from external registration status above.
Current manuscripts
Working manuscripts by Samuel Zhao, available on request. These September 2026 editions present formal models and experimental proposals; they report no empirical validation.
Human-study protocols described in working manuscripts are proposed future empirical validation unless explicitly marked as active on this site.
Defines actions through their consequences for entity properties and directed relationships. Specifies relational impact estimation, state dynamics, group context, and manner formation, with comparisons designed to test the contribution of the representation.
A framework and experimental program; this edition reports no empirical validation.
Specifies positive and negative primary contributions from effective property change, retaining the object and structural origin of each contribution. Describes conditional state-dependent intensity and a proposed learned layer for derived emotions and candidate responses.
Primary contributions, aggregate reports, derived emotions, and selected actions remain distinct. This edition reports no empirical validation.
Examines how differences in state, beliefs, dispositions, experience, feasible actions, and time horizons can produce behavioral diversity. Separates affective magnitude from action valuation and descriptive modeling from execution permissions.
A representation and valuation proposal, not a demonstrated safety guarantee. This edition reports no empirical validation.
Research direction
Explicit relational state suggests questions about heterogeneous agents, repeated interaction, cooperation, and institutions.
Future work can examine how relational capital accumulates or depreciates, how network state affects interaction, and how relationships and group structure evolve alongside behavior. These are directions for formal economic modeling and evaluation.
This work is at the research-agenda stage. Economic results would require explicit formal models, stated assumptions, and separate evaluation.
Discuss economics researchSupporting documents
Describes layer responsibilities, interfaces, and staged validation. Architectural descriptions include planned components and do not imply that every layer is implemented or validated.
Request the design documentBackground for the project’s development. Some terminology and formulations have been superseded; use the current manuscripts above as the entry point.
Request the earlier overview