A system for agents that learn in the open

From live work to a better policy.

AREAL2.0 makes the missing system visible: a universal signal protocol, a safe task proxy, and an automatic trigger for updates.

Run a workload trace

LIVE WORKLOADS · STEP-BY-STEP SIGNALS · SAFE TRAINING DATA · AUTOMATIC UPDATE TRIGGERS · REAL-WORLD DEPLOYMENT ·

The system is the unlock.

Self-evolution needs a loop that observes deployed work, converts it safely, and decides when a change is worth making.

Protocol

Standardize each step into a learning signal.

Proxy

Turn real tasks into safe training data.

Trigger

Decide when evidence earns an update.

Policy

Return improved behavior to deployment.

Watch one workload become a learning loop.

Choose a workload and advance each stage. This local trace demonstrates the named architecture without inventing live data.

Workload

Stage 1 of 3

Capture the work

CHECKOUT

Keep the claim close to its sources.

Grounded in the visible post by Jiqizhixin and linked to the named paper and project.

Read the paper Visit the AReaL project

A scripted three-stage trace preserves its local history rules

Your browser does not support video.

Read the explanation

The source local trace has three stages indexed zero, one and two. Advancing zero to one to two displays stages one, two and three. Common six pixel per percentage point bars show thirty three point three three three, sixty six point six six six and ninety nine point nine nine nine percent, because the code multiplies by thirty three point three three three. This is a scripted demonstration, not actual task observation, training or deployed policy change. Checkout, research and routing are the three scripted workloads. Each has the same three stage positions, so three times three is nine workload-stage combinations. Three equal boxes split into three equal marks show that count. Changing workload resets stage and history; Advance at the final stage cycles to zero. The displayed safe-signal and policy-ready text is fixed source content and does not prove safe training data or model learning. After Reset, three Advance clicks record stages one, two and zero: three history entries but only one current stage, zero. Common one hundred fifty pixel per entry bars compare zero and three actions. Arrow keys and direct stage buttons also append entries, including repeated boundary selections. Export says ready solely when current stage is two. Local persistence restores selection, stage and history, with no independent quality evidence, external deployment or learning receipt.