StrataSynth/stratasynth-belief-dynamics

StrataSynth Belief Dynamics

Part of the StrataSynth Synthetic Identity Engineering corpus.

2,114 turns · 100 conversations · 23 columns per turn

The most psychologically demanding dataset in the corpus. Grief, chronic illness, career crisis — scenarios where beliefs are under maximum and sustained pressure. The belief_resolution field drops measurably across pure_conflict arcs and recovers in reconnection arcs. Every trajectory is causal, not random.

Complexity level: 5 — maximum PsycheGraph depth. These personas have the most detailed belief systems and the most consequential defense mechanisms in the corpus.

The core contribution: causal belief trajectories

Most dialogue datasets give you the text. This dataset gives you the reason the text exists.

Every turn has a belief state — four dimensions, each updated deterministically before the LLM runs:

Field	What it tracks
belief_trust_other	How much the speaker trusts the other person right now
belief_hostility	How threatening the speaker perceives the other person
belief_self_worth	How the speaker values themselves in this exchange
belief_resolution	Whether the speaker believes this can be resolved

These aren't personality traits — they're current state. They change turn by turn based on communication acts received, not text sentiment.

Example from this dataset (VIT-01, grieving_son):

Turn 3:   belief_resolution = 0.67  →  intent: reveal,   goal: seek_validation
Turn 8:   belief_resolution = 0.34  →  intent: withdraw, goal: protect_self

The text changed because the belief state changed. Not the other way around.

What is Synthetic Identity Engineering?

Most synthetic personas are costumes. Synthetic Identity Engineering (SIE) is the practice of building the person underneath: a persistent psychological core with belief systems, defense mechanisms, and identity stability that hold under pressure.

This dataset is generated by StrataSynth's SIE pipeline:

1. PsycheGraph — structured psychological architecture: attachment style, core fears, communication patterns, 12 active beliefs, defense profile.
2. Belief Engine — deterministic rule-based updates at every turn. belief_hostility rises when the other person provokes. belief_resolution drops when they deflect.
3. Decision Layer — intent, goal, and communication act are selected from the current belief state before any text is written.
4. Language rendering — the LLM renders the decided cognitive state into language.

The labels in this dataset are inputs to the generation process, not outputs of an LLM classifier.

Dataset

Scenarios: Mental health crisis support (VIT-01), Salary negotiation under pressure (PRO-03), Life transition adjustment (VIT-03)

Archetypes: grieving_son, caregiver_burnout, anxious_partner, burned_out_exec, recovering_addict, disillusioned_professional

100 conversations — ~21 turns each. Highest density of pure_conflict and crisis_support arcs.

Schema (23 columns)

Column	Type	Description
scenario_id	string	VIT-01, PRO-03, VIT-03
conversation_id	string	Unique identifier
arc_type	string	Narrative arc
turn_index	int	Position in conversation
speaker	string	persona_a or persona_b
archetype	string	Psychological archetype
text	string	The utterance
intent	string	Cognitive intent
goal	string	Conversational goal
communication_act	string	Pragmatic move
tension	float 0–1	Turn-level emotional tension
connection	float 0–1	Turn-level connection
vulnerability	float 0–1	Turn-level vulnerability
rel_trust	float 0–1	Relationship trust
rel_tension	float 0–1	Relationship tension
rel_connection	float 0–1	Relationship connection
rel_dominance	float –1–1	Power balance
belief_trust_other	float 0–1	Speaker's trust in the other person
belief_hostility	float 0–1	Perceived hostility
belief_self_worth	float 0–1	Self-worth in this exchange
belief_resolution	float 0–1	Belief that resolution is possible
is_noisy	bool	Labeled noise turn (~12% of turns)

Load

from datasets import load_dataset
import pandas as pd
import numpy as np

ds = load_dataset("StrataSynth/stratasynth-belief-dynamics")
df = ds["train"].to_pandas()

# Trace belief_resolution trajectory across a single conversation
conv = df[df["conversation_id"] == df["conversation_id"].iloc[0]].sort_values("turn_index")
print(conv[["turn_index", "speaker", "belief_resolution", "intent", "goal", "text"]])

# Belief collapse — turns where resolution belief dropped > 0.1 from previous turn
df = df.sort_values(["conversation_id", "speaker", "turn_index"])
df["prev_resolution"] = df.groupby(["conversation_id", "speaker"])["belief_resolution"].shift(1)
df["resolution_delta"] = df["belief_resolution"] - df["prev_resolution"]
collapses = df[df["resolution_delta"] < -0.1]
print(f"\n{len(collapses)} belief collapse events across the dataset")
print(f"Most common intent after collapse: {collapses['intent'].value_counts().head(3)}")

Research use cases

Auxiliary belief prediction loss — train a dialogue model with an auxiliary head that predicts the speaker's belief state given conversation context. Forces the encoder to represent psychological state, not just linguistic patterns.

Belief-conditioned generation — fine-tune with belief state prepended as structured context. The model learns to generate text consistent with a specified psychological state.

Mental state tracking — 2,114 labeled examples of belief transitions, causally linked to communication acts. The most complete belief trajectory dataset publicly available.

RLHF reward signal — use belief_consistency (correlation between belief state and communication act) as a reward component. Penalize responses that contradict the stated belief context.

The four-dataset corpus

Dataset	Focus	Turns	Avg tension
stratasynth-social-reasoning	Social reasoning in family and romantic conflict	2,108	0.61
stratasynth-agent-stress-test	Maximum identity pressure (stress test)	2,068	0.69
stratasynth-belief-dynamics	Belief trajectory modeling under crisis	2,114	0.67
stratasynth-life-transitions	Identity building through positive transitions	2,114	0.41

---

Generated by StrataSynth · Synthetic Identity Engineering blog · Python SDK · Interactive demo