Sumo experiment problem solving

Simulation Experiment Prompting

Overview

Use this skill to preserve the reasoning behind good experiment prompts. The goal is not to ask for a bigger run; the goal is to convert confusion into one bounded diagnostic question.

observation -> uncertainty -> fixed variables -> smallest contrast -> feedback -> next prompt

Simulators, scripts, controllers, harnesses, and plugins are actuators. The prompt should decide what uncertainty they are allowed to reduce.

Prompt Skeleton

Every nontrivial experiment prompt should contain:

• phenomenon: what was observed, without interpretation.
• desired_claim: what the user wants to be able to say if the evidence supports it.
• blocking_uncertainty: the one unknown that currently prevents that claim.
• fixed_variables: scenario, inputs, seed, horizon, activation time, controller schedule, constraints, and output schema that must stay paired.
• allowed_change: the smallest mechanism, code path, or parameter that may change.
• feedback_signal: the metric, log, trace, or diff that will decide whether the idea survives.
• claim_boundary: what must not be claimed from this run.

If these fields cannot be filled, the next prompt should ask for observation, not execution.

Mental Moves

Use these moves when shaping the user's next request:

1. Separate observation from explanation.
2. Compare against the strongest previous result under the same scenario before inventing new logic.
3. Treat negative results as information about a failure mode, not as proof the research direction is dead.
4. Prefer paired action-log comparison over a larger performance matrix when timing or trigger logic is unclear.
5. Keep actuator usefulness separate from trigger correctness.
6. Check completion, backlog, teleports, warnings, and controller action traces before ranking metrics.
7. End each prompt with a stopping condition.

Prompt Patterns

When Performance Gets Worse

We observed [new controller] is worse than [old controller].
Do not tune broadly yet.
First pair them under the same [inputs/seed/horizon/activation].
Compare action timing by [time, controller unit, state, target, reason].
Tell me whether the failure is over-action, under-action, wrong scope, wrong timing, or metric invalidity.
Stop if completion/backlog makes performance ranking invalid.

When an Old Diagnostic Worked

The old diagnostic fixture showed [action direction] can help.
Now isolate what was fixture-only or oracle-like.
Keep the actuator fixed and test only the trigger that decides when/where to act.
Do not claim deployability unless the trigger uses observable runtime information.

When the Mechanism Is Unclear

We do not know whether [candidate mechanism] is the real reason.
Design the smallest contrast where only [one information structure] changes.
Use logs to prove the controller actually used that information.
Report what remains unproven.

When the Run Is Incomplete

Do not rank controllers from this output yet.
First decide whether incomplete vehicles, backlog, teleports, or warnings changed the denominator.
Extend or repair the run only enough to make the comparison interpretable.

Reusable Boundaries

Keep these meanings stable:

• Event-impact signals define activation or localization scope unless pairwise coupling is explicitly computed.
• Actuator direction is separate from trigger logic. A useful action can still be driven by a wrong trigger.
• Trigger density means action frequency or action budget, not traffic density.
• Diagnostic or oracle fixtures are lower-claim evidence than observable runtime controllers.
• Runtime actions must obey the simulator's existing legality and transition constraints.
• Fine-grained mechanism claims require proof that logged coordinates, entities, and mappings are real.

Decision Ladder

Choose the first rung that applies:

1. If completion is invalid, ask for completion repair, not performance interpretation.
2. If the comparison is unpaired, ask for a paired rerun or call the claim non-identical.
3. If a controller did not log its intended action, ask for logging or a seam audit.
4. If old and new methods differ in many ways, ask for action-log diff under the same scenario.
5. If the actuator worked but the real controller failed, ask for trigger isolation.
6. If the trigger is too dense, ask for action-density diagnostics before new architecture.
7. If the mechanism survives a paired contrast, then ask for broader seeds or perturbations.

Generic Control Prompt Frame

For controller experiments, use this control chain unless the spec says otherwise:

event or observer -> affected scope -> candidate request -> legality constraints -> priority/filter gate -> legal action -> action trace -> paired metric feedback

When an old diagnostic fixture beats a new runtime trigger, the next experiment is usually not a larger matrix. The next experiment is a paired action-timing comparison under the same inputs and seed.

Response Template

phenomenon:
desired_claim:
blocking_uncertainty:
fixed_variables:
allowed_change:
feedback_signal:
smallest_next_prompt:
stop_condition:
claim_boundary: