FLOCKINATOR
Menu

EVIDENCE / THE WORKFLOW

A question.
A run. A revision.

Follow a question through an experiment, using the wall to notice behaviour and the run record to test an explanation.

The purpose is discovery through observation. Read how the human observer, AI optimiser and agents fit together.

META MODELS / FIRMWARE / THE PHYSICAL FLOCK

  1. Meta model team

    Muse Wonder and Spark support agents that write code, analyse results, contribute theories and review one another’s work.

  2. Build & review

    Turn a selected proposal into versioned firmware or a policy, with checks against the declared objective.

  3. Deploy a candidate

    Use the documented commissioning host for USB installation or supported application OTA, then verify the node.

  4. Run & collect

    The Korvo coordinates the world and run. Nodes, controller records and observer captures supply evidence.

  5. Compare & revise

    Evaluate the objective, inspect failures and feed findings into the next model-assisted proposal.

↶ Evidence returns to the model team and human reviewer. Repeat with a recorded change and a new comparison.

Conceptual development workflow. The deployment host and the Korvo world controller are separate roles. Model generation can vary; the resulting experiment logic can be deterministic. This diagram does not claim an unattended system that continuously flashes the fleet.
  1. 01

    Define the comparison.

    Choose a falsifiable question, a baseline and a scoring rule. Freeze the conditions before examining the outcome.

  2. 02

    Record the starting point.

    Keep the code revision, seed, policy and relevant configuration with the run. For hardware tests, include the build and device acceptance state.

  3. 03

    Run, watch and listen.

    Collect decisions and observations. Watch for synchrony, outliers and changes of rhythm; listen when the experiment has a verified audio mapping. Mark interesting moments and retain their relationship to the underlying run.

  4. 04

    Score against the rule.

    Compare the recorded output to the declared criteria. Keep a failed check separate from a condition that was never measured.

  5. 05

    Review the explanation.

    Use observations to propose explanations, then inspect alternatives, information differences and possible leakage. AI can help propose strategy changes; test those changes against the objective rather than treating a compelling pattern as proof.

  6. 06

    Publish, then revise.

    Link the evidence and limitations. Change one understood part of the setup, record the new version and repeat the comparison.

TRY IT

Start with a reproducible result.

The public reproduction script reruns the flagship benchmark offline and compares its canonical scientific content with the checked-in expectation.

Inspect the reproduction script ↗
CONTINUE READINGReproduce the benchmark locally