Demos — choose a path, run it, inspect the evidence
Version 1.18.0. The canonical directory is
alpha_factory_v1/demos.
All 26 original entries, their research narratives, flowcharts and media remain available.
Start in the browser for an interactive tour, or run a finite local example below.
Start in the browser
Built-in scenarios need no account, wallet or API key. Each experience identifies its inputs, computed results and modeled assumptions. Export work you want to retain.
| Your question | Open | What to inspect |
|---|---|---|
| Can a meta-agent teach a solver without leaking evaluation answers? | Curriculum Lab | Adaptive tasks, hypothesis search, lineage, independent gates and verified exports |
| Which workflow design deserves independent review? | MATS Search Lab | Branching search, rollouts, oracle gap, held-out schedules and five gates |
| Should a governance proposal advance to independent review? | Governance Workbench | Nine gates, adversarial cases, exact dossier and verification jobs |
| Which enterprise projects fit capital, staff and reviewer capacity? | Business 3 Enterprise Studio | Exact portfolio, downside stress, evidence bundle and Ascension job specs |
| What should I do with a constrained budget or schedule? | Decision Studio | Edit a case, compare feasible plans and export its dossier |
| Which opportunity deserves further investigation? | Insight Atlas | Scenario assumptions, holdout results and claim evidence |
| How does a claim become reviewed work? | Proof Bloom | Jobs, exact evidence, review, promotion and revocation |
| Does retained experience help a new task? | Compounding Lab | Frozen policy, held-out tasks, costs and negative controls |
| How does the complete Ascension vision fit together? | Ascension Lab | Editable scenarios, cryptosealed seeds, FusionPlans and modeled economics |
| What does the executable enterprise protocol enforce? | Protocol Desk | Local-EVM receipt, role checks, funding, auctions, validation and settlement |
| Where are the original experiments? | Full gallery | All 24 demos and two reference/support entries, with explicit execution modes |
Legacy replay charts display bundled traces. They are illustrations of the retained experiments. Real local model inference and native training have separate, labeled launch paths and prerequisites. See the browser guide for downloads, offline use and recovery.
Run a local example
Use Python 3.11–3.13. The checksum-verifying release installer creates the environment from matching release assets. Activate that environment before continuing. The wheel includes the catalog, CSV samples and Insight scenario fixtures; these commands work outside the source checkout.
For source development, run this setup from the repository root:
python3 -m venv .venv-agent
source .venv-agent/bin/activate
python -m pip install --require-hashes -r requirements-agent.lock
python -m pip install --no-deps -e .
On Windows PowerShell, use python -m venv .venv-agent and then
.\.venv-agent\Scripts\Activate.ps1. The remaining commands are the same.
Now inspect and run a bounded governance review:
python -m alpha_factory_v1.demos list
python -m alpha_factory_v1.demos show solving_agi_governance
python -m alpha_factory_v1.demos check solving_agi_governance
python -m alpha_factory_v1.demos run solving_agi_governance --output-dir my-demo-runs/governance
The command prints the execution mode, expected result, command and output directory. The review
prints nine gate decisions and saves five evidence files, including unsubmitted validator jobs.
See the governance guide for replay and model assumptions. Try macro_sentinel for risk calculations over bundled macro samples, or
era_of_experience for a reproducible learning experiment and independent review gates. Replace the name in show, check and run.
For bounded workflow design search, run meta_agentic_tree_search_v0 or follow the MATS guide.
For enterprise project selection, run alpha_agi_business_3_v1 or follow the
Business 3 guide. Its bounded default requires only Python and exports seven evidence files.
For useful work on your own records, continue with the native mission walkthrough. It produces a signed, reviewable journal using research, allocation, schedule, forecast or isolated code missions.
Understand the launcher
| Command | Behavior |
|---|---|
list / list --json |
Enumerates all 26 entries and their execution modes |
show NAME |
Shows the command, expected result, limitations and guide without launching |
check NAME / check NAME --json |
Checks declared module and bundled-file presence without importing optional backends, downloading or creating run state |
run NAME --output-dir PATH |
Checks prerequisites, then launches the documented command with its declared settings |
check returns 0 when declared prerequisites are present, and 2 when a prerequisite is missing
or the entry has no standalone CLI. It does not exercise model weights, GPU support, services or paid
accounts. Follow the entry's guide for those integrations. The launcher never installs packages.
Some examples need additional modules such as Rich or Click beyond the base operator environment.
The 16 finite catalog examples use explicit offline defaults. Their child processes disable configured OpenAI/Anthropic keys, remote Neo4j/PostgreSQL storage and tracing; tree search uses its bundled synthetic workflow model and reproducible private random stream. Your shell credentials are unchanged. The release tests run all 15 from the wheel with Python network calls blocked, including Python 3.11/3.12 environments with the full backend dependencies and inherited database settings. This verifies those launch contracts; the launcher is not a network or code sandbox. Advanced standalone commands and optional services retain their separately documented behavior.
Output defaults to demo-runs/NAME/. Relative directories resolve from the directory where you launch.
Paths containing spaces are supported. The child does not import Python modules from its output
directory. Stop services with Ctrl+C. Copy the run directory to retain results; choose a new output
directory for a fresh run. Lineage examples append to existing history. Model caches are separate.
Follow the full Ascension lifecycle
| Vision | Runnable path | What the evidence establishes |
|---|---|---|
| 🎖️ Insight → hidden α | Insight Atlas, native research/forecast/allocation, Insight demos | Reproducible analysis of supplied or constructed inputs; predictive superiority requires external evaluation |
| 🌱 Nova-Seeds → foresight genome + FusionPlan | Ascension Lab; NovaSeed.sol; native ascension-compile and ascension-check |
Encrypted recovery, ERC-721 lineage and exact committed goal/metric/bounty jobs |
| 🔮 MARK → risk-gated funding | Protocol Desk; AscensionRiskOracle.sol and AscensionMark.sol |
Local-EVM admission, expiring reviews, bonding-curve funding and recovery paths |
| 👑 Sovereign → an enterprise plan | Plan-restricted treasury and indexed Merkle jobs | Once-only job routing, budget and deadline checks |
| 👾 Agents → reputation-weighted execution | Native missions; AscensionJobMarket.sol |
Staked *.alpha.agent.agi.eth fixture identities; price/time/reputation auction rules and evidence-bound delivery |
| ✅ Validators → payment or failure | Protocol Desk settlement; *.alpha.club.agi.eth fixture identities |
Two-of-three exact-result decisions, refunds, explicit slashing and a 1% payout burn in $AGIALPHA |
| 🧬 Experience → successor enterprises | Proof Bloom, Compounding Lab, successor Nova-Seeds | Reviewable retained evidence and unapproved successors; measured bounded transfer rather than assumed general improvement |
The factory guide maps these mechanisms to implementation. The protocol guide reproduces the actual local-EVM lifecycle, including negative cases and the connection to native signed delivery. The contract reference is undeployed; browser simulations and fixture ENS identities do not establish a live global marketplace. All original industry integrations remain available with their stated research and deployment boundaries.
Resolve a failed launch
- Missing module: activate the intended environment, run
check NAME, then install the optional requirements from that entry's guide. Do not substitute a training stub for native execution. - Missing bundled file: reinstall the same release into a new environment. The catalog refuses an incomplete sample installation before starting the demo.
- Output path is a file or is not writable: choose a new writable directory. Existing results are retained; the launcher reports the path error without a traceback.
- A service exits or a port is busy: read the printed guide and service output. Stop the prior instance or use that service's documented port setting.
- A model is unavailable offline: use its documented, checksum-verified model download first.
- An unexpected result remains: retain the output, version, command and error for diagnosis;
omit credentials.
runpreserves the child exit status and prints the guide on failure.
The validation matrix identifies the catalog, installed-wheel, native CPU, browser, protocol and publication gates. Release readiness defines the verified deployment scope. These guides supplement every preserved original flowchart and research narrative.