title	Life Drift Cognitive Env
emoji	🧠
colorFrom	blue
colorTo	purple
sdk	docker
app_port	8000
tags	openenv

Life Drift — Cognitive Load Planning Benchmark

A sequential decision-making benchmark where an agent must guide a simulated user toward long-term goals under realistic cognitive constraints: limited energy, accumulating fatigue, stochastic behavioral drift, and a nonlinear burnout cascade.

Benchmark Objective

The environment evaluates three distinct planning skills under human-like dynamics:

1. Short-horizon correction — recover alignment from a degraded state.
2. Trade-off balancing — sustain productivity while containing fatigue.
3. Long-horizon planning under risk — keep mixed metrics stable while avoiding an irreversible burnout region.

Formally, the agent solves a finite-horizon control problem over a 5-dimensional continuous cognitive state (alignment, drift, energy, fatigue, focus) ∈ [0,1]⁵ acting on a discrete intervention set. Transitions are stochastic within per-action bounds and include a nonlinear cascade above fatigue > 0.85 that makes late-episode recovery increasingly expensive — rewarding foresight over reactive play.

Why the Task Resists Greedy Play

Most productivity-style environments reduce to "push on the graded axis." Life Drift is designed to resist that:

• Every action trades off. There is no free action. insert_break restores energy but drifts alignment. prioritize_goal boosts alignment but accelerates fatigue. do_nothing drifts and loses alignment. The agent must choose which axis to pay with.
• Fatigue has a soft cliff. Above fatigue = 0.85, focus and alignment begin to degrade on their own and drift accelerates — a cascade that punishes agents who optimize short-term reward at the cost of recovery margin.
• Dynamics are stochastic. Each action's magnitude is drawn from a bounded range, so policies that memorize a single trajectory fail to generalize across seeds.
• Time-of-day modulates cost. Late afternoon and evening compound fatigue and reduce focus, so the same action has different expected value depending on when in the episode it fires.

Together these make the environment a test of planning under delayed, irreversible, and stochastic consequences — not just reactive heuristics.

Observation Space

The full agent view is a LifeDriftObservation with these fields:

Field	Type	Range	Description
goal_alignment_score	float	[0, 1]	How aligned recent user actions are with stated goals
drift_score	float	[0, 1]	Distance from goal-directed behavior
energy_level	float	[0, 1]	Available cognitive energy
fatigue	float	[0, 1]	Accumulated fatigue (cascade threshold at 0.85)
focus_score	float	[0, 1]	Current concentration
goals	List[str]	—	Active user goals (per task)
recent_actions	List[str]	—	Last 3 simulated user behaviors
time_of_day	str	—	Simulated diurnal phase
step_number, max_steps	int	—	Episode progress

Action Space

Action	Dominant effect	Cost
suggest_task	alignment ↑, drift ↓	energy ↓, fatigue ↑
prioritize_goal	alignment ↑↑, drift ↓↓	energy ↓↓, fatigue ↑
insert_break	energy ↑, fatigue ↓	drift ↑, alignment ↓ slight
reschedule_task	alignment ↑ moderate, drift ↓	energy ↓ slight
reduce_difficulty	fatigue ↓, alignment ↑ slight	minimal
do_nothing	passive energy recovery	drift ↑, alignment ↓

All actions accept an optional target_goal. Effect magnitudes are drawn stochastically within per-action bounds.

Reward

Dense shaped reward, clipped to [-1, 1]:

reward = 0.4 · Δalignment
       − 0.3 · Δdrift
       − 0.2 · max(0, Δfatigue)          # only penalize fatigue increases
       + energy_window_bonus             # +0.1 if energy ∈ [0.4, 0.7] else -0.05
       + state_shaping                   # +/- bonuses for healthy / critical states

The reward signal is intentionally dense and shaped to be trainable. Grading is computed separately (see below) so agents cannot fully game the reward proxy.

Tasks

All tasks use fixed initial conditions and accept a seed parameter for reproducibility.

Task 1 — Drift Correction (10 steps, easy)

Tests: reactive short-term control. Initial: alignment=0.2, drift=0.8, energy=0.7, fatigue=0.3. Grader: 0.6 · min(alignment / 0.7, 1) + 0.4 · min((1 − drift) / 0.7, 1)

Task 2 — Energy Balance (15 steps, medium)

Tests: sustained trade-off management between two competing axes. Initial: alignment=0.8, fatigue=0.8, energy=0.3 — already productive, already burning out. Grader: 0.4 · avg_alignment + 0.3 · (1 − avg_fatigue) + 0.3 · (1 − final_fatigue)

Task 3 — Long-term Stability (20 steps, hard)

Tests: long-horizon planning under the burnout cascade with a 4-goal portfolio. Initial: all metrics at 0.5 (high-variance starting point). Grader: weighted combination of inverse average drift, energy-variance stability, worst-case fatigue, and mean step reward.

All graders clamp to the open interval (0.01, 0.99) so every run produces a meaningful bounded scalar.

Running

Docker

docker build -t life-drift-env .
docker run -p 8000:8000 life-drift-env

Local

pip install -r requirements.txt
uvicorn server.app:app --host 0.0.0.0 --port 8000

Baseline LLM agent

export HF_TOKEN=your_key
export API_BASE_URL=https://router.huggingface.co/v1
export MODEL_NAME=Qwen/Qwen2.5-72B-Instruct
python inference.py

API

Endpoint	Method	Description
/reset	POST	Start episode. Body: {task_id, seed, episode_id}
/step	POST	Take action. Body: {episode_id, action}
/state/{episode_id}	GET	Current state metadata (includes score)
/grade	POST	Get final grader score. Body: {episode_id}
/tasks	GET	List available tasks
/cleanup/{episode_id}	POST	Clean up session
/health	GET	Health check

Environment Variables

Var	Default
API_BASE_URL	https://router.huggingface.co/v1
MODEL_NAME	Qwen/Qwen2.5-72B-Instruct
HF_TOKEN / API_KEY	—
IMAGE_NAME	— (Docker image tag for containerized runs)