CodexGameDev/experiments/003_rig_trials

Experiment 003 — Rig Trials

Question

Is modifying a physical construction enjoyable when it directly changes the body the player pilots, and do qualitatively changing requirements preserve reasoning better than escalating statistics?

Hypotheses

The leading possibility is that Experiments 000 and 001 made construction feel like low-leverage programming, while enjoyed construction games let a design become the player's practical capability. A competing explanation is that modular construction will still collapse into obvious loadout work or a universal rig. See hypothesis.md.

Difference From Previous Experiment

• The player modifies a persistent rig rather than upgrading numeric character statistics.
• The constructed object is controlled directly; it is not an autonomous agent.
• Three complete trials impose different physical requirements without unlocking parts.
• Existing construction persists when switching trials and edits are cheap.
• Physical placement affects connectivity, thruster exhaust, tool exposure, cooling, shielding, mass, and wind response.
• There are no enemies, defeat waves, random drops, or multiplicative progression.

Expected Result

Support would look like revising the rig after observing a physical weakness, retaining a useful subassembly while changing the overall layout, or testing a build idea beyond the minimum trial requirement. Completion by itself is weak evidence.

A universal starter layout, obvious one-part substitutions, or course execution dominating build decisions would count against the intended mechanism.

Controls

• Every part and every trial is available immediately.
• Part behavior is deterministic and displayed.
• The same rig and direct controls are used across all trials.
• Trial completion grants no statistical power.
• Switching trials preserves the build but resets only the field state.
• After a switch, reset, or edit, the field waits until the arena is clicked so construction is not performed under accidental trial pressure.

How To Run

From this directory:

./run.sh

Then open http://localhost:8000.

What To Pay Attention To

• Use the starter rig or change it whenever you want.
• Try any or all trials; switching does not erase the rig.
• Use Save JSONL when finished. When launched through run.sh, it writes directly into the repository-level JSONL/ directory; a browser download remains the fallback if the local save server is unavailable.

Result

The first playtest completed all three trials in about 6:47 but was reported as “pretty boring.” Haul consumed most of the session and involved repeated tractor/cargo manipulation plus two unexpected thermal collapses. Furnace was answered by stacking five sinks around one remaining right drive, then completed in under eight seconds. Gale was completed with four cardinal drives and no ballast. See results/e1fa0f4d-analysis.md.

Validation

• JavaScript syntax checking passes.
• A 1672×976 Chromium screenshot—the viewport used in the preceding playtest log—kept the whole arena visible and all essential builder controls on screen.
• Real browser input placed and removed modules, rotated placement, switched trials, piloted the rig, produced completion/failure events, and emitted revision-tagged JSONL without runtime exceptions.
• The untouched starter rig thermally collapsed around 69% of Furnace. Adding a parallel right-facing drive completed it. Placing a drive directly behind another correctly blocked the inner exhaust instead of increasing thrust.
• Haul rejected a misaligned tractor approach and accepted a corrected approach; the attached fragment immediately increased effective mass. The open-loop validator did not complete the return route because it cannot visually correct momentum around walls.
• Trial fields remain stationary while the player edits. Gale begins only after the arena is clicked.

Interpretation

Controls were not the primary problem. Haul prompted an unintended two-tractor juggling workaround only because the player missed Sinks and misread heat as a movement budget; it was still boring. Furnace reduced to one forward Drive plus five Sinks, Gale reduced to restoring the default rig, and lattice placement felt inconsequential except for fiddly Drive orientation.

The prototype behaved like a spatially slower equipment menu. Changing requirements produced prescribed full loadout swaps rather than meaningful recomposition. This is negative evidence for H06/H17 as implemented, not a general rejection of physical construction.

Next Best Experiment

Pivot to a genuine legible-discovery experiment. Test whether learning and transferring a surprising stable rule has pull independent of combat power, numerical progression, or a cosmetic construction lattice.