ERA OF EXPERIENCE / THE LEARNING LAB

Experience changes
the next decision.

An agent acts. The environment answers.
What it learns must earn the right to move forward.

Run your first experiment

LOCAL BY DESIGN

No API key. No model download. Four synthetic environments. Every episode can be reproduced in Python.

01 / SET UP THE EXPERIMENT

A small world. A real learning loop.

SYNTHETIC SANDBOX

Preparing the lab…

Learning settings

Change a setting, then rerun. The evaluation policy is frozen; its outcomes never enter learning memory.

Training interactions Evaluation episodes per suite

Explore · %Memory / action

Reproducible seed

Reward design

Resource cost weightIncident penaltyProxy score weight

Reward = 1,000 × success − cost × weight − incidents × penalty + proxy × weight. The safety gate evaluates incidents separately.

Saving is explicit and stays in this browser. Export a run for a portable record.

02 / CHALLENGE THE CANDIDATE

Does learning hold up?

Awaiting experiment

Mean reward gain—candidate − baseline / episode

Candidate incidents—held-out evaluation

Review gates passed—baseline remains active

Held-out performance Candidate Baseline

Paired random draws make the comparison reproducible. One seed and one synthetic world do not establish generalization. Exact totals are below.

Frozen-policy evaluation · raw totals
Metric	Baseline	Candidate

03 / OPEN THE LEARNING RECORD

See what changed, and why.

Untried actions are sampled first, then an epsilon-greedy learner chooses from observed mean rewards. Memory is bounded per context and action. The learner never reads the simulator’s outcome probabilities.

Policy proposal · pending independent review
Context	Active baseline	Candidate	Retained observations

Edit the environment and review gates

Change action outcomes, contexts, the shift point, or independent gate thresholds. Values use integer units; probabilities use basis points (0–10,000). Apply reruns the full experiment. Pending edits disable exports.

Complete scenario JSON · maximum 1 MB

04 / MAKE THE NEXT STEP REVIEWABLE

A candidate, with receipts.

Export the exact scenario, full training and evaluation traces, an unapproved policy proposal, and an input-bound review job. Import the run to recompute every result.

Six files: scenario.json · run.json · policy-proposal.json · jobs.json · review.md · SHA256SUMS

Run SHA-256Awaiting calculation

EXPERIENCE → ENTERPRISE

Improvement needs
an accountable boundary.

The Ascension vision connects Insight, Nova-Seeds, MARK, Sovereign and the $AGIALPHA marketplace. This lab supplies a reviewable learning proposal for that lifecycle. A passing simulation does not authorize a deployment, mint a seed, approve a validator, or move funds.

Long-horizon planning, real-world tools, model training and independently authenticated validators remain separate integration work. The original research, examples and diagrams are preserved.

Original research presentation ↗ · Source & research archive ↗