An open experiment in machine learning from experience

Intelligence should compound. Not start over.

Today's AI can solve a problem and forget how it solved it. Project 507 is exploring an open-source substrate that learns from execution, tests what works, and turns successful experience into reusable software.

Open development. Inspectable evidence. Real-world outcomes.
01 / THE EXPERIMENT

Can a system keep learning while it operates?

We're working toward sustained autonomous operation with observable outcomes, recoverable failures, and an inspectable history of what the system learned.

Uptime is one measure of continuity—not proof of intelligence. The more important question is whether experience makes future execution better.

Read the project documentation
research question / evidence
01  # Can the system reach a real outcome?
02  discover candidate activities
03  execute in context
04  verify downstream consumer reach
05  preserve trace and evidence
06  test reuse in a new context
07  # Compile only when justified
08  investigate drift and failures
02 / THE ARCHITECTURE

From an attempted action to a reusable capability.

Project 507 treats execution as something to learn from. Activities are state transitions. Shapes express semantic intent. Downstream consumers test whether the result is actually useful.

Successful pathways can be preserved with their evidence, reused in context, and eventually compiled into inspectable executable resolvers.

Explore the architecture
⌕
DiscoverFind candidate activities for an intent
▷
ExecuteAct within a real context
◎
ObserveTrace and verify outcomes
∿
LearnUpdate what is useful, and when
⬡
CompileTurn reliable traces into resolvers
↺ Reuse, test in new contexts, and fall back to discovery when a resolver fails.
01 / OUTCOMES

Reality over predictions.

A plausible answer isn't enough. The system seeks independent evidence that an intended state was reached.

02 / EXPERIENCE

Reuse what works.

Preserve execution pathways and their evidence, then turn reliable behavior into reusable software instead of repeating the same reasoning.

03 / ADAPTATION

Learn when things break.

When a compiled resolver stops working, return to discovery and experimentation to find a viable alternative.

03 / DISTRIBUTED BY DESIGN

Learning belongs to everyone.

An experience learned in one environment should be a hypothesis another environment can test. Project 507 explores sharing reusable structures, execution evidence, and learned pathways without requiring participants to surrender underlying content or blindly trust someone else's conclusions.

Privacy, contextual differences, and the possibility of misleading evidence remain open engineering problems.

a federated learning principle
# A borrowed lesson is not a fact.

experience = receive(structure, evidence)
context = inspect(local_environment)

if test(experience, context):
    update_local_priors()
else:
    preserve_disagreement()
    investigate_context()

# The model can change.
# The experience should remain.
04 / CURRENT FOCUS

What we're investigating.

≡

Execution traces

Preserving actions, context, outcomes, and evidence.

▥

Evaluation

Measuring downstream success and uncertainty.

⬡

Resolver compilation

Testing when experience can become reliable software.

↺

Cross-task reuse

Understanding when learned behavior transfers and when it fails.

These are research directions, not claims that every capability is implemented or independently validated.

05 / PARTICIPATE

Help build what comes next.

The architecture is ambitious, and its most important claims need reproducible evidence. Engineers, researchers, operators, and skeptics are welcome to inspect the code, challenge assumptions, contribute resolvers, and help design experiments.

$ git clone -b dev https://github.com/project-507-seed/substrate.git
$ cd substrate

# Follow the repository's current setup instructions.
# Run an experiment. Inspect a trace. Open an issue.

Setup commands beyond cloning are intentionally omitted until verified against the repository.