coreason-chronos

Domain: Temporal Reasoning, Time-Series Forecasting, & Longitudinal Reconstruction

Overview

coreason-chronos is the specialized temporal processing unit of the CoReason ecosystem. It enables agents to reason about When, How Long, and What's Next.

Core Philosophy: "Semantic time is fuzzy. Symbolic time is exact. We need both."

It provides three critical capabilities:

1. Longitudinal Reconstruction: Extracting events from unstructured text (e.g., "Patient started Taxol 3 weeks after surgery") and mapping them to a normalized absolute timeline.
2. Probabilistic Forecasting: Predicting future events (e.g., "Clinical Trial Enrollment will hit 100% in 4.5 months +/- 2 weeks") using SOTA Time-Series Foundation Models.
3. Temporal Logic Validation: Ensuring GxP compliance by validating temporal constraints (e.g., "Was the adverse event reported within the 24-hour statutory window?").

Features

• The Timeline Extractor (The Historian):

  • Converts relative dates ("2 weeks later") to absolute timestamps.
  • Resolves "anchored" events based on semantic proximity to reference events.
  • Outputs structured EventSeries JSON.

• The Oracle (The Forecaster):

  • Leverages Foundation Time-Series Models (Amazon Chronos-T5).
  • Zero-shot prediction capability for metrics like Patient Enrollment or Drug Inventory.
  • Provides probabilistic forecasts (P90 confidence intervals).

• The Compliance Clock (The Validator):

  • Symbolic Rule Engine for regulatory checks.
  • Validates constraints like Report_Time <= Event_Time + 24_Hours.

• The Causality Engine (The Sequencer):

  • Uses Allen's Interval Algebra to determine temporal plausibility of causal relationships.

Installation

pip install coreason-chronos

Usage

1. Initialize the Timekeeper

The ChronosTimekeeper is the main entry point for the library.

from datetime import datetime, timezone
from coreason_chronos.agent import ChronosTimekeeper

# Initialize the agent
agent = ChronosTimekeeper()

2. Longitudinal Reconstruction (Timeline Extraction)

text = "Patient was admitted on 2024-01-01. Symptoms started 2 days later."
reference_date = datetime(2024, 1, 1, tzinfo=timezone.utc)

events = agent.extract_from_text(text, reference_date)

for event in events:
    print(f"{event.timestamp}: {event.description}")
# Output:
# 2024-01-01 00:00:00+00:00: Patient was admitted...
# 2024-01-03 00:00:00+00:00: Derived from anchor...

3. Forecasting

history = [10, 15, 20, 25, 30]  # Weekly enrollment
forecast = agent.forecast_series(history, prediction_length=5)

print(f"Median Forecast: {forecast.median}")
print(f"90% Confidence Interval: {forecast.lower_bound} - {forecast.upper_bound}")

4. Compliance Check

from datetime import timedelta
from coreason_chronos.validator import MaxDelayRule

# Check if reporting was done within 24 hours
rule = MaxDelayRule(max_delay=timedelta(hours=24))

# ... (assuming target_event and ref_event are TemporalEvent objects)
# result = agent.check_compliance(target_event, ref_event, rule)

# if not result.is_compliant:
#     print(f"Compliance Violation! Drift: {result.drift}")