"""MemoryStress dataset schema — dataclasses for the benchmark format.
The top-level JSON file (memorystress_v1.json) contains all facts, sessions,
questions, and contradiction chains needed to run the benchmark.
"""
from __future__ import annotations
from dataclasses import dataclass, field
@dataclass
class Fact:
"""A single ground-truth fact planted in a session."""
fact_id: str # "F001"
content: str # "User's favorite language is Rust"
category: str # preference|decision|technical_fact|personal_info|event|relationship
phase_planted: int # 1, 2, or 3
session_planted: str # "S003"
simulated_date: str # ISO date
importance: int # 1-5
mention_count: int # how many sessions reference it
mention_sessions: list[str] = field(default_factory=list)
contradiction_chain: str | None = None # chain_id or None
recoverable: bool = True # should this be answerable after eviction pressure?
tags: list[str] = field(default_factory=list)
difficulty: str = "medium" # easy|medium|hard
@dataclass
class ContradictionChain:
"""A fact that evolves across sessions (updated, reverted, accumulated)."""
chain_id: str # "C007"
original_fact_id: str
versions: list[dict] = field(default_factory=list)
# [{version, content, session_id, simulated_date, phase}]
final_ground_truth: str = ""
pattern: str = "fact_update"
# fact_update|fact_update_revert|fact_accumulate|fact_partial_update
@dataclass
class Session:
"""A single conversation session to be ingested."""
session_id: str # "S003"
session_index: int
phase: int # 1, 2, or 3
simulated_date: str # ISO datetime
topic: str
agent_id: str # "agent_a" or "agent_b"
turns: list[dict] = field(default_factory=list) # [{role, content}]
planted_fact_ids: list[str] = field(default_factory=list)
noise_level: str = "low" # low|medium|high
token_count_estimate: int = 0
@dataclass
class Question:
"""A benchmark question targeting specific facts."""
question_id: str # "Q001"
question: str
target_fact_ids: list[str] = field(default_factory=list)
target_chain_id: str | None = None
answer: str = ""
answer_detail: str = "" # context for grader
phase_asked: int = 1 # checkpoint where this is asked (1-4)
target_fact_phase: int = 1 # when the fact was planted
age_sessions: int = 0 # distance in sessions
age_days: int = 0 # distance in simulated days
question_type: str = "fact_recall"
# fact_recall|preference_recall|contradiction_resolution|
# temporal_ordering|cross_agent_recall|cold_start_recall|
# single_mention_recall
scoring_dimensions: list[str] = field(default_factory=list)
difficulty: str = "medium"
agent_scope: str = "agent_a" # which agent to query as
cold_start: bool = False # if True, run with fresh agent instance
@dataclass
class BenchmarkData:
"""Top-level container for the full benchmark dataset."""
version: str = "1.0"
phases: dict = field(default_factory=dict)
# phase boundaries + date ranges
stats: dict = field(default_factory=dict)
# total counts
facts: list[Fact] = field(default_factory=list)
contradiction_chains: list[ContradictionChain] = field(default_factory=list)
sessions: list[Session] = field(default_factory=list)
questions: list[Question] = field(default_factory=list)
