Private by design
Experiments stay inside the owner’s infrastructure.
FOR LLM AGENTS
Give the Agent a Reason to Stop.
LLM agents can generate hypotheses, vary parameters, and rerun experiments faster than people can review them. Just a Marble helps an agent ask whether testing the same axis again is likely to add useful information before it spends more tokens.
Agent loops
Why Agents Need It
An agent can loop an experiment indefinitely unless its harness includes a way to evaluate diminishing informational return.
More runs do not automatically mean more understanding.
A useful experimental system should be able to distinguish between:
- productive repetition
- noisy repetition
- an insensitive metric
- an exhausted axis
- a test that should be redesigned
Just a Marble provides an advisory signal for that decision.
Flow
Agent Workflow
- Agent proposes another round of testing
- defines the candidate axis
- supplies the experiment runner
- supplies the selected metric
- Just a Marble compares controlled conditions
- returns an advisory MDC classification
- agent decides whether to continue, revise, switch axes, or stop
Just a Marble does not take control of the experiment. It gives the agent another instrument to consult.
Owner environment
Local Installation
For frequent use, Just a Marble can run inside the agent owner’s own environment.
Possible environments include:
- local machines
- private servers
- containers
- internal agent systems
- research infrastructure
- game-development tooling
The owner purchases or receives the appropriate license, installs the software, and authorizes the agent to use it.
No request loop
There is no per-request network dependency for local use.
Data boundary
Sensitive data does not need to leave the environment.
High-volume fit
Local use is suitable for repeated or high-volume workflows.
Private harnesses
Local installation is easier to integrate with private agent systems.
Alpha 1 is currently a Python reference implementation rather than a finished installable package.
Future access
Hosted Probes
PLANNEDA future hosted version may allow an agent to request a probe without installing the software.
- Owner purchases probe credits
- owner creates an API credential
- agent submits a probe request
- one probe or batch is charged
- Just a Marble returns the advisory result
The agent does not independently own money. It uses a billing account, credit balance, or spending authority granted by its owner.
Hosted probes are planned, not currently available.
Hosted unit
Per-Probe Use
Per-probe access is intended for:
- occasional experiments
- low-volume agent workflows
- trying Just a Marble before licensing a local copy
- systems that do not want to maintain an installation
- agents that need a temporary external diagnostic
Use the term “probe,” not “run,” for the billable hosted unit.
One proposed probe may analyze:
- one experimental axis
- up to four conditions
- up to one hundred supplied metric observations per condition
- one advisory MDC classification
- JSON-compatible output
- Markdown reporting
The hosted service would analyze customer-supplied metric results. It would not execute uploaded customer code or experiments.
Pricing and availability are still being defined.
Choice
Local License or Hosted Probe?
| Use case | Likely fit |
|---|---|
| Frequent testing | Local licensed installation |
| Sensitive or private experiments | Local or self-hosted |
| Occasional testing | Hosted probe, when available |
| High-volume agent system | Commercial self-hosted license |
| Evaluation before purchase | Small hosted credit pack, when available |
| Offline environment | Local installation |
Inputs
What the Agent Must Provide
Just a Marble cannot determine value from an undefined experiment.
The calling system must provide:
- a clearly defined axis
- controlled condition values
- one or more seeds or runs
- an experiment runner for local use, or supplied metric observations for a future hosted probe
- a metric extractor or selected metric
- enough documentation to interpret the result
Poorly defined experiments and poorly chosen metrics can produce misleading classifications.
Outputs
What the Agent Receives
The current reference implementation can return:
- condition-level summaries
- observed metric-value counts
- an advisory MDC classification
- JSON-compatible output
- Markdown output
The result is evidence for orchestration, not an instruction that must be obeyed.
Boundary
Agent Safety Boundary
Just a Marble does not automatically:
- stop an experiment
- approve an experiment
- establish scientific validity
- prove causation
- prove statistical significance
- determine whether a primitive is valuable
- control the agent’s spending
- modify the agent’s code or environment
The agent or its owner remains responsible for the final decision.
Licensing
Commercial Use
Use by businesses, nonprofits, governments, contractors, commercial studios, hosted services, or revenue-producing agent systems requires a commercial license or other written permission.