Skip to content

Project notice

Meta-Agentic AGI v3 — Curriculum Lab operating guide

The Curriculum Lab generates small integer-program tasks, searches for solutions from examples, and lets a meta-agent choose the next solver configuration. Separate tasks review the final frozen candidate. No account, API key, Docker installation or model download is needed for this workflow.

First useful result

  1. Open the lab and keep A curriculum that can challenge its solver selected.
  2. Read the frozen solver accuracy, improvement over the baseline and counted operation cost.
  3. Move the round slider to inspect the examples, inferred program and predictions. Reveal the reference to compare it with the hypothesis; the solver itself never receives the reference or expected test answers.
  4. Compare candidate configurations and expand their parent/child lineage.
  5. Read all four review gates, then inspect a held-out task. Eligible for review does not mean approved.
  6. Download the six-file evidence ZIP and retain it outside browser storage.

Try When the curriculum misses an entire skill to see a coverage failure and Accuracy has a resource cost to see a candidate held despite high accuracy. A hold is a valid experiment outcome.

Run locally

Install the matching release using Start here, then:

curriculum-lab --list
curriculum-lab --case balanced --output curriculum-runs
curriculum-lab --serve

Open the printed loopback URL (http://127.0.0.1:7863/meta_agentic_agi_v3/ by default). Stop with Ctrl+C. If the port is occupied, use --port 7864. The server exposes only packaged static assets and a health check; it has no upload, credential, shell or execution API. python -m alpha_factory_v1.demos.meta_agentic_agi_v3 is equivalent to curriculum-lab.

The core is standard-library-only. Cloud-related environment variables do not change its behavior.

Customize before evaluating

Setting Allowed values Meaning
Seed Integer 0–4294967295 Private, reproducible xorshift random stream
Rounds 1–12 Number of curriculum/selection iterations
Tasks per round 3–12 Requested unique behaviors; a saturated grammar may yield fewer
Maximum depth 1–3 Maximum operators in a composed hypothesis
Families Unique subset of arithmetic, nonlinear, remainder Training curriculum coverage
Positive examples Boolean Restrict training examples to positive inputs to expose ambiguous hypotheses
Entropy weight 0–1000 Weight for diversity of solved task families
Review policy Accuracy, weakest-family accuracy, baseline gain, operation ceiling Independent eligibility thresholds

The visible controls cover common settings; expand Inputs & reproducibility for complete JSON. Apply edited JSON before running. Every change invalidates existing exports. Save, restore and forget are explicit device-local operations; unrelated browser storage is retained. Saved settings do not restore results.

Choose seeds and policies before examining evaluation results. This public synthetic test set is an educational control, not a secret benchmark. Repeated evaluation-guided tuning requires new independent data.

Retain and verify

The ZIP and native run directory contain:

File Contents
scenario.json Exact accepted settings and stated assumptions
run.json Tasks, solver attempts, complete candidate metrics, lineage and independent review
solver-proposal.json Frozen candidate, input hash and UNAPPROVED status; baseline remains active
jobs.json Unsubmitted Ascension review job with goal, success metric and draft AGIALPHA bounty
review.md Human-readable outcomes, exact comparisons and scope
SHA256SUMS Hashes of the five preceding files 
curriculum-lab --verify path/to/run.json
curriculum-lab --input my-scenario.json --output curriculum-runs

Verification recomputes all evidence and rejects rehashed forgeries. Identical native reruns reuse the identical SHA-256 named directory; altered or incomplete bundles are rejected. Back up complete directories. Checksums do not authenticate a person or prove validator approval.

The Ascension protocol can compile the exported job into a FusionPlan. The draft bounty is 100 $AGIALPHA, expressed as 100000000000000000000 base units. This lab neither posts nor funds the job; separate identity/staking, escrow, validator and settlement contracts govern that lifecycle.

Scope and preserved research

This is an actual finite hypothesis search, configuration evolution and adaptive curriculum. It is not AZR model training, PPO, NSGA-II, a general-purpose autonomous agent, a measured financial-alpha system, or live enterprise commissioning. Its free-energy diagnostic is a cost-minus-entropy proxy, not thermodynamic energy or an ELBO. Counted interpreter operations are a reproducible resource proxy; they are not measured dollars, carbon or latency.

The source README explains the complete algorithm and keeps the original architecture diagram visible. The original research archive, notebook, configuration and visual replay are preserved. Historical claims do not expand the validated scope. The updated notebook runs the finite lab. The original provider experiment and RoyaltyRadar reconciliation remain available with corrected safety/accounting behavior and explicit limits.

Troubleshooting

  • No export: apply pending JSON and run again. Downloads never silently use old settings.
  • Import rejected: restore the original source scenario and correct its validation error. Inputs are limited to 1 MB, bounded UTF-8 JSON, known fields and finite values. Do not patch a run report to force acceptance.
  • Offline reload fails: first load the public lab online, wait for its service worker, then reload. The packaged loopback server works without caching or network access. Other pages or models may need separate downloads.
  • Legacy code evaluation fails: generated Python requires a usable Docker daemon and sandbox image. There is no host-execution fallback. Use the finite lab for a dependency-free local experience.
  • Royalty settlement requested: the example prepares evidence only. Assumed revenue discrepancies are not debt findings; it never sends claims or native-currency transfers masquerading as AGIALPHA payments.

Evaluation probes use −8, −6, 6 and 8, disjoint from training probes −5, −2, 0 and 5. Reference functions are independently sampled from the same finite grammar and may recur across the two sets; this tests a frozen solver configuration on fresh input probes, not transfer to an unseen language or domain.