CodexGameDev/experiments/002_conversion_breach/README.md

Experiment 002 — Conversion Breach

Question

Are stable magical/physical laws interesting when the player manipulates them directly during an ongoing activity, and does external pressure make understanding instrumentally valuable?

Hypotheses

The leading hypotheses are that direct agency is preferable to autonomous programming and that optimization needs practical stakes. The competing explanation is that action merely masks another shallow interaction. See hypothesis.md.

Difference From Previous Experiment

• No graph, wiring, autonomous controller, or static route specification.
• The player directly extracts, stores, injects, and converts energy while moving and aiming.
• Temperature consistently affects speed, brittleness, and instability; momentum, mass, walls, and collisions determine damage.
• Calm and breach variants use the same verbs, entities, and laws. Only threat pressure changes.
• Periodic calibration opportunities offer chosen improvements, not random verbs or permission unlocks.

Expected Result

Support for the new model would look like deliberate exploitation: freezing before a wall impact, overheating a construct before driving it into a group, using one enemy as fuel or a projectile, or choosing upgrades to deepen a discovered tactic. The pressure hypothesis is supported only if breach mode makes these decisions more meaningful without eliminating experimentation.

Controls

• All manipulation verbs exist from the start in both variants.
• The initial field layout and construct properties are deterministic.
• Calm and breach modes share physics; constructs only pursue and injure the player in breach mode.
• Switching variants starts a fresh baseline so calibration upgrades cannot contaminate the comparison; resetting within one variant preserves its calibration.
• No authored spell recipes, equipment drops, or part unlocks.

How To Run

From this directory:

./run.sh

Then open http://localhost:8000.

What To Pay Attention To

• Try both Calm and Breach; which one would you rather continue?
• Did you discover or deliberately exploit an interaction?
• Did improvement choices create a tactic you cared about, or only numbers?

Use Export JSONL when finished.

Result

The first playtest produced the first sustained engagement in the program: about 16 logged minutes, approximately 14 in Breach, ten waves cleared, 74 defeats, tactical adaptation, and strongly directed upgrade choices.

The player changed from indiscriminate heat extraction to conditional cooling based on target position and remaining fuel. They voluntarily tested player ramming. They chose conversion 19 of 24 times, capacity four times, and transfer once.

The test also contained major contamination: an input-device conflict during the first run, viewport failure, through-wall extraction, crowded wall phasing, an unreadable resource deadlock, exponential conversion scaling, and an invalid Calm comparison. The player now has a mouse, so the first issue is historical rather than a priority for the next run. See results/0732452b-analysis.md.

Interpretation

This is positive evidence that a coupled property/resource system with practical stakes can provoke learning and tactical revision. It is not sufficient evidence that action, pressure, or direct control is the root cause.

The most promising observed interaction was not simply “freeze enemies.” Extracting heat simultaneously created ammunition, altered pursuit speed, enabled brittleness, and depleted the remaining world resource. That coupling caused the player to revise when and where they extracted.

However, the dominant outcome was wall impact, conversion upgrades became exponentially overwhelming, and Calm did not share wave goals. The result could be driven by progression, increasing power, conventional action feedback, a safe wall exploit, or the property tradeoff. These remain competing explanations.

Next Best Experiment

Corrective revision 2 gives Calm the same waves, populations, layouts, and calibration as Breach; only pursuit and contact damage differ. It also fixes full-screen containment and sidebar scrolling, blocks transfer and discharge through obstacles, strengthens obstacle collision resolution, clarifies brittle/unstable impact feedback, adds click-to-move and alternate pointer controls, and marks new logs with prototype_revision: 2.

Revision 2 passed JavaScript syntax checking and real-browser layout checks at 1920×1080 and 1365×768. Browser input validation confirmed click-to-move and logged displacement, both pointer transfer modes, variant switching, inspection of an unobstructed target, rejection of a wall-occluded target, and revision-tagged telemetry.

The revision-2 playtest lasted about 19:39: roughly 2:55 in Calm and 16:44 in Breach. Calm received a genuine attempt, while Breach reached wave 29. Input was easier with a mouse and upgrade selection broadened substantially. Nevertheless, wall-directed impulse remained the durable answer, escalating statistics eventually let the default heat supply clear enemies, and the player reported little new to say. See results/bb30ddfc-analysis.md.

Move to an exploratory Experiment 003 rather than further tuning 002. Test a directly piloted constructed artifact under qualitatively changing physical requirements, without numerical power escalation.