Initial Commit

experiments/002_conversion_breach/results/.gitkeep

@@ -0,0 +1 @@
+

experiments/002_conversion_breach/results/0732452b-analysis.md

@@ -0,0 +1,70 @@
+# Playtest Analysis — Session 0732452b
+
+Date: 2026-08-16
+
+Source log: `JSONL/conversion-breach-0732452b-3ce4-4329-aa69-b121180c61a8.jsonl`
+
+## Observed Behavior
+
+- Total logged duration: 970 seconds (about 16 minutes 10 seconds).
+- Breach events span about 14 minutes 23 seconds across two breach segments.
+- The player reached wave 11 and cleared ten complete waves.
+- 74 constructs were destroyed in Breach.
+- 67 kinetic discharges affected an average of 1.66 constructs; none affected zero targets.
+- 184 tether sessions: 139 extraction and 45 injection.
+- 25 tether sessions lasted under 150 ms, consistent with at least some control interruption or accidental taps; the log cannot prove the cause.
+- 80 thermal-state transitions, including 33 entries into brittle and only two into unstable.
+- Only one unstable release occurred.
+- 69 of 74 destroys came from wall impacts. Five came from construct collisions.
+- 19 player-damage events occurred. Thirteen were in the first logged Breach wave; only six occurred across later waves. Minimum recorded health was 22.
+- Reservoir was below 5 in 42 of 150 Breach snapshots (28%). Mean snapshot reservoir was 47.5, but later capacity upgrades inflate that mean.
+- 24 calibration choices: 19 conversion, four capacity, one transfer.
+- Final calibration values were capacity 240, transfer 35.1, conversion 85.071. Multiplicative conversion scaling became extreme.
+
+## Player Report
+
+- Calm was perceived as a sandbox/tutorial rather than a test and was left quickly for Breach.
+- Breach remained interesting for an estimated 5–10 minutes; the log shows longer engagement.
+- Trackpad palm rejection made keyboard-plus-pointer controls poor during this run. The player subsequently obtained a mouse, so this should not be treated as a central cause in the next comparison.
+- Initial tactic: extract as much heat as possible, then blast.
+- Revised tactic: avoid leaving the final enemy brittle and depleted unless already positioned for an immediate wall blast.
+- Capacity felt unhelpful; transfer already felt fast; most upgrades went to conversion.
+- The player tried to ram constructs after perceiving possible collision damage, but could not understand or safely reproduce it.
+- Extracting through a wall felt too safe.
+- Crowded constructs appeared to phase through walls.
+- Full-screen layout hid instructions and/or part of the field; the player could not confidently scroll. A narrower window appeared to fit better.
+
+## Important Model Mismatch
+
+The implementation multiplies brittle collision damage by 13. Brittle targets do not take less collision damage. The reported difficulty killing a final brittle target is more plausibly a resource-state problem: extraction removed the target's heat, the player reservoir was empty, and no other enemy remained as fuel. Cooling also reduced pursuit speed, which could make arranging an unpowered collision harder.
+
+Do not “correct” the player report as if their experience were invalid. The mismatch is evidence that the causal feedback did not clearly distinguish toughness from lack of available impulse.
+
+Player ramming did not deal construct damage in the implementation. The player nevertheless formed and tested that hypothesis from observed collision behavior. This is positive voluntary-experiment evidence, while the inability to read the result is a feedback/affordance failure.
+
+## What This Supports
+
+- A coupled tradeoff produced real adaptation: extracting heat both funded attacks and changed the remaining target/resource state.
+- Practical stakes and/or wave progression provided incentive absent from Experiments 000 and 001.
+- The player chose upgrades instrumentally around the perceived bottleneck rather than distributing them evenly.
+- Knowledge and power both increased: later waves generally produced few player hits and many were cleared much faster.
+- Visible physical laws generated at least one unrequired hypothesis (ramming).
+
+## What This Does Not Establish
+
+- Pressure was not isolated. Calm lacked continuing waves/progression and looked like a testing room.
+- Direct control was not isolated from action feedback, wave goals, upgrades, escalating power, and enemies.
+- The property system may still be shallow. Wall blasts dominated, brittle was common, unstable behavior was nearly absent, and conversion upgrades overwhelmed other choices.
+- Later safety may reflect learning, exponential power, through-wall extraction, collision bugs, or their combination.
+- Continued play may have been driven by wave progression/power accumulation rather than curiosity about laws.
+
+## Corrective Work Before Further Comparison
+
+- Make Calm a complete wave/calibration variant differing only in pursuit and player damage.
+- Clearly label Calm and Breach as two full experimental variants.
+- Add pointer-only movement and thermal-mode controls; add Space as an alternate discharge input. Retain these as general accessibility/readability improvements, but do not center the next test on trackpad friction because the player now has a mouse.
+- Make the entire field and instructions fit a full-screen viewport, with unmistakable scrolling where needed and resize telemetry.
+- Block tethering through obstacles.
+- Strengthen obstacle collision resolution to prevent clustered phasing.
+- Improve feedback that separates low target heat, low player reservoir, target integrity, and actual impact damage.
+- Preserve the current playtest build's result before using corrected behavior as new evidence.

experiments/002_conversion_breach/results/bb30ddfc-analysis.md

@@ -0,0 +1,50 @@
+# Playtest Analysis — Session bb30ddfc
+
+Date: 2026-08-16
+
+Source log: `JSONL/conversion-breach-bb30ddfc-8e4f-4e18-bb7e-22235266d9ed.jsonl`
+
+Prototype revision: 2
+
+## Observed Behavior
+
+- Total logged duration was 1,179 seconds, about 19 minutes 39 seconds.
+- Calm lasted about 2 minutes 55 seconds. The player destroyed 12 constructs, cleared wave 1, and defeated six of the seven constructs in wave 2 before switching.
+- Breach then lasted about 16 minutes 44 seconds. The player reached wave 29, cleared 28 waves, and destroyed 195 constructs.
+- The player made 208 kinetic discharges: 25 in Calm and 183 in Breach. Four Breach discharges affected no target.
+- There were 431 tether sessions: 343 extraction and 88 injection. Injection was more common than in the first session.
+- The pointer transfer buttons themselves were toggled only once in each direction; most injection continued to use Shift.
+- There were 69 forced calibration choices. In Breach these were capacity 33, conversion 24, and transfer 8. This is substantially broader than the first session's conversion-heavy 19/24 distribution.
+- Final Breach values were capacity 1,255, transfer 220.247, and conversion 374.144. All three upgrade dimensions became extreme.
+- 176 of 207 total defeats came from wall impacts: all 12 Calm defeats and 164 of 195 Breach defeats. Construct collisions accounted for most of the remaining non-release defeats.
+- There were 16 unstable releases, all in Breach. Twelve occurred in the middle portion of the run, but they did not displace wall impact as the dominant answer.
+- Breach produced 49 player-damage events but no collapse. Damage remained intermittent into late waves.
+- Later Breach waves often took only roughly 10–25 seconds. From wave 24 onward, conversion climbed from 40.565 to 374.144 within about 70 seconds.
+
+## Player Report
+
+- The player could not say much beyond the first playtest and is ready for a different experiment.
+- A mouse made control easier.
+- Easier input led them to use wider conduit and reservoir upgrades more than before.
+- Eventually the default amount of supplied heat was sufficient to blast everything away.
+- Accidental Shift plus right-click could still open the browser context menu and disrupt play.
+- Clicking interface text outside the canvas selected text. The player considers these browser-medium problems rather than central game problems.
+
+## Interpretation
+
+The mouse had a measurable behavioral effect, so the original input complaint was not noise. Upgrade selection diversified and injection was used more often. However, this did not generate a materially new reported experience or a new strategic layer.
+
+The strongest result is not merely that conversion was overtuned. The activity offered one durable answer—acquire any heat and turn it into wall-directed impulse—and every upgrade increased the throughput or reserve of that same answer. Capacity and transfer therefore broadened the numerical build without broadening the decision structure.
+
+The long session should not be read as equivalent to deep interest. Wave progression, repeated calibration interruptions, rapid power accumulation, and a desire to see the system's limit can all sustain behavior after strategic curiosity has ended. The player's explicit “can't really say much more” and readiness to move on are important counterevidence to duration as a fun proxy.
+
+The matched comparison gives weak evidence that threat pressure increases continuation: Calm received almost three minutes and Breach nearly seventeen. It does not show that pressure creates deeper reasoning or greater enjoyment. Familiarity with Breach, action intensity, wave escalation, or simple challenge could explain the difference.
+
+The coupled heat tradeoff mattered early in Experiment 002, but numerical escalation eventually erased it. Once a small amount of heat created overwhelming impulse, target temperature, remaining fuel, mass, position, and reservoir scarcity stopped constraining the answer. This is direct evidence of strategy crystallization inside a short run.
+
+## Design Consequence
+
+Do not make Experiment 003 by adding enemy types, nerfing conversion, extending waves, or offering more numerical upgrades. Those changes would tune the lifetime of the same dominant answer without testing whether another activity is fun.
+
+The next experiment should remove escalating numerical power and test whether the player enjoys adapting a directly controlled constructed artifact to qualitatively changing physical requirements. This preserves instrumental building and direct agency while replacing repeated extermination waves with different problems that should demand different layouts. It also probes whether construction felt tedious in Experiment 001 because it programmed an autonomous agent, or because construction itself lacked useful leverage.
+