TERMINAL INTELLIGENCE RESEARCH VISUALIZATION

Recursive self-improvement.

Keep the better starting point. Then begin again.

Measured trajectories of alternating RSR and RPR Three chapters expand from one to four to six benchmarks. Blue trajectories are retained; gray branches are discarded. Later stages begin at retained checkpoints, not discarded branch ends. All y axes are linear and stages equally spaced. 01 / A BETTER STARTING POINT MEASURED TRAJECTORIES · LINEAR SCALES
01 · Focus00:00/01:19
DATA & METHODOLOGY

Measured trajectories. Nothing smoothed away.

Extracted from the first main table of RSI Result.csv, using the stage mapping of plot_rsi_results.py. The repeated summary table is excluded. 128 measured records.

Selected checkpoints ≠ per-suite maxima
RPR-1 step 20 and Long-Reward step 18 are the retained roots. Discarded points may score higher on an individual suite. Final TB4 performance is 19.84%.
Missing values are not imputed
SWR100 is measured only at Base, RSR-1, and RPR-1: 3.00, 5.00, and 7.00%. Its line stops at the last observation.
Independent dual axes
LHTB: blue = completed tasks (0, 0, 2, 4, 4 of 46); light blue = process reward (0.21, 0.29, 0.30, 0.31, 0.35). The two linear axes have different units.
Animation encoding
Diamonds mark RSR; stars mark retained RPR. The final star is 1.3× larger. Base labels are smaller and unbolded. No curve smoothing or missing-value extrapolation is used.

SOURCE CSV SHA-256

This standalone player uses no remote fonts or libraries. Reduced-motion preferences are respected; drag the timeline or step through events for a static reading.