Empirical follow-up to JLC v1 (Zenodo DOI 10. 5281/zenodo. 19681402). Single individual, single mini-PC (Beelink SER8), single ~20/month cloud inference plan, single continuous 1000-turn / 32. 62-hour conversation on 2026-04-23 → 2026-04-24, completed with 0 logged infrastructure errors and 0 retries under the JLC architecture. Cumulative prefill tokens 653, 363 (JLC) vs 86, 626, 579 (equivalent raw long-context baseline, same tokenizer) — a 99. 25% reduction. One Stage 2 verified mono-lingual verbatim recall at Δ=969 turns (user's English at T26 reused verbatim by the model at T995, ~32 h apart, JHB capacity 418 tokens at recall time). Bundle includes: paper (v2. tex + v2. pdf), drafts (EN + KO), figures, full plain-text transcript, sha256-locked source manifest, Stage 1 + Stage 2 provenance, per-turn latency profile, deep-recall analysis (Stage 2 verified n=1 verbatim recall + 3 axis-A retrievability observations at strictly weaker grade), 4-AI cold prior + Claude informed follow-up dated 2026-04-25 (median prior probability the report is accurate as stated = 22. 5%). Implementation code not included by author policy; all headline numbers are independently recomputable from the raw transcript with tiktoken cl100kbase alone. v2. 1 update (2026-04-30) Non-breaking additive version. All v2. 0 files (paper PDF, evidenceₚack zip, metadata) are carried over from v2. 0 without modification. v2. 1 adds one raw artifact: tokens. jsonl (1. 71 MB, 1022 records) — the per-turn raw token ledger exported unedited from the runtime token tracer. The file includes the main 1000-turn run (turnₗoggerₚath: jhb/default/turns. jsonl) and 22 preceding dry-run turns (multidomainₖo₂0260423005748: 20 turns; multidomainₑn₂0260423042025: 2 turns) carried over without modification — i. e. no cherry-picking before publication. Each line is a JSON object with turn identity, system/user/assistant token and character counts, JHB state transitions, encoder I/O, per-turn compression ratios, retrieval call statistics, totals, USD cost breakdown, cache state, per-stage latency, and any errors or retries. The paper's §5. 2 and §10 reproducibility claims are already satisfied by the aggregated ledgers in v2. 0's JLCᵥ2ₑvidenceₚack₂026-04-25. zip; this v2. 1 ledger is a transparency bonus and brings v2 into symmetry with the companion 5-tier record at 10. 5281/zenodo. 19681402. Companion record (5-tier ablation, 100 turns x 7 runs): 10. 5281/zenodo. 19681402. Paper PDF: unchanged from v2. 0. SHA256 (tokens. jsonl) = c61989acf9557b548ecfe8df4508d4dabb9f607982feacf944e0ca47a08ffa5c
Kim et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: