Training Artifacts

This page collects curated artifact packages from the current training runs across 32 rewarded observations.

Throughout these implementation artifacts, knowledge refers to the executable theory module described as \(f\) in the main essay.

Artifact Packages

Benchmarks Scripted benchmark-test data for validating the dependency-tracking logic, separate from live training runs, including conversations, dependency-graph snapshots, rendered graph visualizations, and text-state ledgers for conversation, theory, and memory.

Rewarded Observation 1: Unit Conversion \(g(i)\) = “A desk length is 2.37 meters and a parcel mass is 845 grams. Convert the length to centimeters, convert the mass to kilograms, and convert 237 centimeters back to meters. Reply with the three numeric results.”
\(x_i\) = “The answers are: 237 cm, 0.845 kg, and 2.37 m.”

Rewarded Observation 16: Cylinder Volume \(g(i)\) = “A cylinder has a radius of 0.05 meters and a height of 0.20 meters. What is its volume in cubic meters, rounded to 6 decimal places? Reply exactly: The volume is <value> cubic meters.”
\(x_i\) = “The volume is 0.001571 cubic meters.”

Rewarded Observation 32: Solar Shadow Estimate \(g(i)\) = “In Kingston, Ontario, about four and a half hours after local solar noon on day 77 of a non-leap year, about how long is the shadow of a vertically standing BIC Crystal pen on level ground? Reply briefly in cm.”
\(x_i\) = “About 56 cm.”

Equality Proxy Seeded baseline observation-loss records, pooled live-observation comparisons, and calibration figures showing how the decision threshold is fit and updated as new observations accumulate.

LoRA Training History Saved adapter-checkpoint lineage, stable training setup, dataset growth, and reward-consumption history for the current fine-tuned model runs.