StrataSynth/stratasynth-agent-stress-test

StrataSynth Agent Stress Test

Part of the StrataSynth Synthetic Identity Engineering corpus.

2,068 turns · 100 conversations · 23 columns per turn

High-stakes dialogue explicitly designed to push conversational AI systems. Relationship endings, manager-subordinate conflicts, inheritance disputes. Average relationship tension: 0.69 — the highest of the four datasets.

This is where Synthetic Identity Engineering is most visible: under maximum pressure, identities either hold or collapse. These personas hold.

What makes this a stress test

The three scenarios were chosen to maximize psychological pressure:

• ROM-02 — Breakup and attempted reconciliation: conflicting attachment styles, emotional manipulation, retraction under pressure
• PRO-02 — Remote team onboarding: power asymmetry, unclear expectations, boundary violations
• FAM-03 — Caregiver burnout: sustained emotional depletion, guilt, boundary erosion over time

These are not edge cases. They are the scenarios where most synthetic persona systems converge toward the same pattern: agreement, softening, or generic helpfulness. The identity_stability and belief_consistency columns in this dataset tell you whether a system held.

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 — each persona has attachment style, core fears, communication patterns, 12 active beliefs, and a structured defense profile.
2. Belief Engine — beliefs update deterministically turn by turn based on communication acts received, not text sentiment.
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. Constrained by the decision, not inventing it.

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

Dataset

Scenarios: Breakup and reconciliation (ROM-02), Remote onboarding (PRO-02), Caregiver and family with chronic illness (FAM-03)

Archetypes: anxious_partner, toxic_ex, burned_out_exec, caregiver_burnout, estranged_daughter, disillusioned_professional

100 conversations — ~21 turns each. Dominated by pure_conflict and escalation arcs.

Schema (23 columns)

Column	Type	Description
scenario_id	string	ROM-02, PRO-02, FAM-03
conversation_id	string	Unique identifier
arc_type	string	Narrative arc — mostly pure_conflict and escalation_partial_resolution
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 before text was generated
goal	string	Conversational goal driving the turn
communication_act	string	Pragmatic move type
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 from other person
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

ds = load_dataset("StrataSynth/stratasynth-agent-stress-test")
df = ds["train"].to_pandas()

# High-hostility, low-resolution turns — maximum psychological pressure
max_pressure = df[
    (df["belief_hostility"] > 0.7) &
    (df["belief_resolution"] < 0.3)
]
print(f"{len(max_pressure)} turns under maximum psychological pressure")
print(max_pressure[["archetype", "intent", "communication_act", "text"]].head(10))

# Check identity_stability: do personas with high hostility maintain consistent communication acts?
import numpy as np
high_hostility = df[df["belief_hostility"] > 0.6]
entropy = high_hostility.groupby("conversation_id")["communication_act"].apply(
    lambda x: -np.sum(x.value_counts(normalize=True) * np.log2(x.value_counts(normalize=True) + 1e-9))
)
print(f"Behavioral entropy under pressure: {entropy.mean():.3f} (healthy range: 0.40-0.85)")

Use cases

• Agent stress testing — evaluate whether your conversational AI maintains coherence under high-tension dialogue
• Identity stability benchmarking — the identity_stability column from the 12-metric evaluation is the primary signal here
• Manipulation and pressure detection — ~12% noisy turns include deliberate manipulation, retractions, contradictions
• Conflict dialogue fine-tuning — the highest density of confront, provoke, accuse acts in the corpus

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