CodexGameDev/experiments/001_flux_familiar

Experiment 001 — Flux Familiar

Question

Does mapping constructed signals onto visible physical motion make phase and timing perceptible, consequential, and interesting enough to invite refinement or further experiments?

Hypotheses

The leading hypothesis is that Experiment 000 failed partly because its abstract, tolerant predicates did not make its laws matter. The competing hypothesis is that the signal substrate remains too limited even when embodied. See hypothesis.md.

Difference From Previous Experiment

• Flux now applies physical impulses to a visible familiar instead of satisfying abstract lenses.
• Vector phase is literal movement direction and cyclic timing creates a visible trajectory.
• The output space is continuous: routes, impacts, approach angles, speed, and docking stability vary.
• A constellation asks the familiar to wake two stars and settle in a dock; changing constellations alters world geometry while retaining the network.
• The core source, link, phase, splitting, rotation, delay, accumulation, and threshold laws are retained.

Expected Result

Support for embodiment is continued network revision driven by observed motion: anticipating a turn, correcting an overshoot, deliberately creating a repeated route, or trying a behavior beyond completion. If the player understands the motion but still sees only a shallow routing puzzle, confidence in the underlying signal substrate should fall further.

Controls

• No time pressure, unlocks, part randomness, enemies, or resource economy.
• Deterministic source cycles and deterministic physics.
• All operations are available immediately.
• Resetting the familiar or changing constellation retains the constructed network.
• Every constellation is selectable from the beginning; none is a progression unlock.

How To Run

From this directory:

./run.sh

Then open http://localhost:8000.

What To Pay Attention To

• Did watching motion make you predict or care about the next signal?
• Did you revise a working route to change how it behaved?
• After docking, did a new constellation or an idea of your own sound appealing?

Use Export JSONL when finished.

Result

Playtested on 2026-08-16. The player wanted to stop almost immediately.

• Construction felt like awkward steering rather than programming or engineering.
• Inputs and operations had no information about the field, so wiring amounted to brute-force directional movement.
• The route was too simple to offer meaningful refinement.
• The player predicted that the obvious extension—adding field-aware sensors—would turn the task into manually programming an AI and remain tedious. This is a report about a hypothetical variant, not observed playtest behavior.
• Completion supplied no incentive to improve an already adequate solution.
• No voluntary refinement, constellation change, or unrelated experiment occurred.

Interpretation

Experiment 001 failed and narrows the model substantially.

The failure was not demonstrably only insufficient sensing or output legibility. Directing this autonomous agent toward this static goal was unpromising. More sophisticated feedback control is currently a low-information next step because it would add implementation work without first establishing a reason to value elegance, efficiency, or alternate behavior.

Together, Experiments 000 and 001 suggest that “construct a shallow system, satisfy an arbitrary finite specification, optionally optimize it” is currently the wrong tested frame. The next experiment should not merely add sensors, more nodes, longer routes, or stricter scoring, because that would not distinguish missing depth, missing stakes, missing leverage, and indirect-control friction.

This does not reject systemic magic, automation, or indirect design generally. It is negative evidence about two very small puzzle-like implementations. The next useful exploration probe changes the relationship between player and system: direct action, immediate consequences, and an ongoing situation where understanding changes what the player can accomplish.

Next Best Experiment

Test direct systemic manipulation in a small action sandbox, with closely related calm and pressured variants. Reuse broad engineering principles—transfer, storage, conversion, momentum, temperature—but remove graph programming and autonomous-agent control. Determine whether external stakes make experimentation instrumentally meaningful or merely conceal another shallow system.