TABLE OF CONTENTSCAMPAIGN

Competition / task: OpenAI Parameter Golf: build a compressed language-model submission under contest constraints and minimize validation bits per byte.

Public competition context: the competition and pull requests were public OpenAI Parameter Golf PRs. Adapting to public SOTA, comparing against public submissions, and incorporating the best available techniques were part of the competition. All work included in this report is mine, built within that collaborative public field.

Track calibration: the experimental branch was the non-record unified transformer / recursion research lane. It contains architecture probes and recurrence ideas tested for insight; the neural branch is the competitive leaderboard-submission track.

BPB: bits per byte. Lower is better; this report separates BPB model scores from custom-kernel / FA3 timing ratios.

Valid submission: a BPB submission counts only when it follows the three-seed contest standard: artifact under the 16 MB cap, a 10-minute training run on 8xH100, a 10-minute maximum evaluation period on 8xH100, and no test-data leakage.

Invalid scores: BPB values below 1.06 are treated as TTT test-data leakage and excluded from valid competition comparisons.

Excluded / non-comparable: artifacts over 16 MB, runs outside the 8xH100 / 10-minute train / 10-minute eval / three-seed standard, crash rows, loader failures, and kernel ratio measurements are not counted as valid BPB submissions.

Best valid score: 1.0620 VAL_BPB.

Baseline: the tracked Naive Baseline reference was 1.2244 VAL_BPB; the best valid result is a 0.1624 BPB reduction, about 13.3% lower than that baseline.

Improvement: the strongest line was a hybrid neural submission using a larger 10k-vocabulary CaseOps tokenizer with looped GQA transformer mechanics, per-group compression, int7/int6 packaging, LQER, and phased LoRA TTT.

Evidence: all files and logs provided for review or reproduction.