Cyberdeck · experiment status

The bench where the strange work earns its keep.

This is the live-ish public dashboard for experiments. Not a promise that everything works. A promise that every serious claim needs a boundary, a log, a test and a next gate.

EXP-004 cage-design

Four-Machine Agent Colony

A bounded four-node mini-PC colony to test useful emergence and evolution without giving it production access.

  • phase 0
  • containment first
  • no live action rights
Next gate

Write STOP gates, network boundary and one-agent smoke test before any colony run.

Read related note →
EXP-003 live-soak

HOTMEM / Agent Memory

Small boot memory, large governed recall. The agent should carry the library card, not the whole library.

  • selective recall
  • privacy gates
  • bootloader memory
Next gate

Keep measuring quality and only expand injection after evidence, not excitement.

Read related note →
EXP-002 field-tested

Physical Switches for Agent Systems

The off switch belongs outside the thing being switched off. Prompts steer; hardware boundaries stop.

  • human-owned stop path
  • local audit
  • hard boundary
Next gate

Turn the idea into public checklists and diagrams.

Read related note →
EXP-001 on-the-bench

Local Witness Models

A smaller local model does not need to beat the cloud. It needs to tell the truth when the cloud path disappears.

  • degraded mode
  • local policy
  • incident witness
Next gate

Define what the witness must verify before it is allowed near real workflows.

Read related note →
EXPERIMENT RULE

Dream big, but wire the stop path first. Experiments can be ambitious inside the cage; the public site only shows what is safe to explain without exposing private infrastructure, customers, credentials or live control paths.