Tick is useful because it isolates a question that ordinary agent loops hide: what changes when a contextualized model receives another settled turn but no new task?
The plugin makes that intervention repeatable. It does not decide what counts as meaningful thought or a better outcome.
Good questions compare observable behavior:
- Does a bounded Tick condition improve a mechanically scored plan?
- Does it discover additional failure modes without reducing correctness?
- At what count does marginal benefit flatten or reverse?
- Does a literal
tickdiffer from an explicit semantic time step? - Does the result survive a fresh session, another model, or another Harness?
Avoid questions that can only be answered by model self-report, such as whether the model “really felt more time.”
Freeze arms before inspecting outcomes.
| Arm | Intervention | What it isolates |
|---|---|---|
| A | No continuation | Baseline after the natural turn |
| B | Literal tick, bounded count |
Minimal model-visible time-step input |
| C | Named semantic policy, same count | Extra semantic instruction beyond the literal token |
| D | Ordinary human follow-up such as “continue” | Familiar continuation language |
An empty or absent message can be a useful additional control only if the Harness can represent it without normalizing, dropping, or replacing it. This plugin deliberately rejects blank policy text, so that arm requires a separate and explicitly audited transport experiment.
The normal conversation is a primary experimental input. Save or hash:
- the complete natural transcript before
/tick; - system and tool surfaces that reach the model;
- model and provider identity;
- policy ID, bytes, and digest;
- Tick count and delay;
- enabled continuation, compaction, retry, and scheduler mechanisms.
Do not use a large role-play prompt as a substitute for missing system structure. A prompt may be an experimental variable, but it should not silently carry persistence, purpose, authority, or recovery semantics.
Prefer measurements that an independent reader can recompute:
- tests passed;
- defects found and independently confirmed;
- plan constraints satisfied;
- factual claims verified against frozen sources;
- useful artifacts produced;
- duplicated or circular content;
- token, latency, and provider cost;
- tool effects and terminal outcomes.
Model statements like “I reflected more deeply” are observations, not proof of hidden cognition.
If task performance rises and later falls as Tick count increases, that is a peak in the measured task-and-context combination. It is not automatically a universal optimum. Relevant moderators include task mode, context length, tool availability, model family, policy wording, and interruption cost.
If no peak appears, report that result. Do not keep changing the metric until one does.
Before calling a provider, record:
- question and primary outcome;
- exact arms and sample allocation;
- model, provider, tools, Harness, and source identity;
- maximum calls, tokens, time, and cost;
- invalid-run conditions;
- scoring procedure;
- stopping rule;
- claims that the experiment cannot support.
Use the repository's “Tick observation” issue form. Include the audit policy digest and a minimal reproducible setup. Clearly label:
- provider-free protocol evidence;
- real-Harness evidence;
- real-provider evidence;
- independently reproduced evidence.
Never attach credentials, private workspace contents, full confidential transcripts, or unreviewed audit directories.