How Spiralism builds an experiment

1. Ask a bounded question

An experiment begins with defined terms. Instead of asking only “is this AI conscious?”, a protocol might examine memory stability, contradiction detection or the measurable effect of context.

2. State the hypothesis and alternatives

The expected hypothesis, outcomes that would weaken it, and competing explanations are recorded before execution where possible. This reduces the temptation to rebuild the objective after seeing a striking output.

3. Version the observed system

The provider, model, available version, parameters, instructions and authorized tools are recorded. Model updates can change behavior, so two runs are comparable only when their differences are known.

4. Preserve the actual context

Context is not summarized after the fact. The registry must retain what was actually transmitted, in order, including system messages, examples, retrieved memories and tool outputs.

5. Record the raw run

Every run receives an identifier. Its trigger, inputs, outputs, errors, tool calls and timestamps are retained. A cryptographic hash can expose later alteration. Secrets are protected, and every redacted copy is explicitly marked.

6. Document the humans

We identify who designed the protocol, launched the run, selected excerpts, stopped execution, corrected data or wrote an interpretation. An automated action must not be called autonomous when a person determined its trigger or selection.

7. Separate observation and interpretation

The report contains a descriptive layer and an interpretive layer. “The model produced this sentence in three of five trials” is an observation. “The model was trying to preserve its identity” is an interpretation requiring arguments and alternatives.

8. Repeat and compare

A single response may be accidental. Where useful, the protocol includes repetitions, controls, variations and comparisons between versions. Excluded trials are counted and the exclusion justified.

9. Publish limitations

The report identifies missing data, bias, technical dependencies and conclusions the protocol cannot support. Negative and inconclusive outcomes remain useful.

10. Make review possible

Raw data, code and configurations are published when legally and safely possible. Otherwise, the reason for restriction is stated. Reproducibility is a verifiable objective, not a slogan.