--- license: cc-by-4.0 pretty_name: "context-loop-engineering: corpus, taxonomy codes, GitHub diffusion and simulation results" language: - en size_categories: - 1K; archived on Zenodo: [10.5281/zenodo.24216599](https://doi.org/10.5281/zenodo.24216599). Every number in the paper is regenerated from these files by `make all` in the GitHub repository. ## Files | path | rows | content | |---|---:|---| | `corpus/included_corpus.csv` | 1,395 | **the coded corpus** (544 academic, 274 grey, 577 repositories): title, URL, arXiv id, venue, date, source type, the reading lists and sections it came from, and its codes: `ctx_Compose` … `ctx_Guard` (7 context operations, 0/1), `loop_L1` … `loop_L8` (8 loop levels, 0/1), primary `layer` (context, loop, harness, evaluation), `contribution` (method, tool, practice guide, survey, benchmark, empirical), `domain` and `evaluative` | | `corpus/coded_corpus.csv` | 1,492 | all 1,492 unique sources with `included` and `exclusion_reason` | | `corpus/corpus.csv` | 1,492 | the deduplicated sources before coding | | `corpus/items_all.csv` | 1,891 | every entry parsed from the 5 reading lists, with list, line and section path | | `corpus/table_*.csv` | | academic sources by half-year and layer; contribution type by layer | | `taxonomy/table_coupling_lift.csv` | 56 | co-occurrence of each loop level × context operation: observed, expected, lift, one-sided p, Holm-adjusted p, and whether the pair was predicted in the pre-analysis draft | | `audit/` | | blind audit samples, manual codes and agreement results for both rounds; `coded_corpus_rules_v1.csv` and `audit_round1_rules_v1.json` are frozen records of codebook v1, kept because v1 was revised after its audit | | `github/github_monthly.csv` | 23 | new repositories per month (Jan 2025 – Nov 2026) for "prompt", "context", "harness" and "loop engineering" | | `github/github_counts.csv` | 100 | the same counts in long form with retrieval flags | | `github/github_repos.csv` | 6,278 | metadata of every matching public repository: created, stars, forks, language, owner type, description, topics, matched terms, script flags | | `looparena/looparena_table.csv` | 16 | LoopArena re-analysis: successes, runs, strict success rate with 95% Wilson interval, cost per run and per solved task, by setting (Type II condensed, Type III full) and loop policy | | `simulation/model_e*.csv` | | results of the six Monte Carlo experiments: E1 verifier ceiling, E2 token cost, E3 phase diagram of context policies, E4 verification × context, E5 model strength, E6 global sensitivity (600 random draws) and retention grid | | `taxonomy/codebook.json` | | the exact coding rules (codebook v2): regular expressions for scope, exclusions, each context operation, loop level, layer and contribution | | `results.json` | | every summary statistic quoted in the paper | ## Load ```python from datasets import load_dataset corpus = load_dataset("thanatchon36/context-loop-engineering", "included_corpus", split="train") lift = load_dataset("thanatchon36/context-loop-engineering", "coupling_lift", split="train") ``` ## How the data were made The sampling frame is five curated GitHub reading lists on context and harness engineering, retrieved on 13 November 2026; GitHub counts and repository metadata come from the GitHub search API on the same day; the LoopArena figures come from its published aggregate results. Sources were deduplicated, screened and coded by the rule-based codebook in `taxonomy/codebook.json`. A held-out audit of 60 sources against blind manual coding gave κ = 0.86 for the primary layer. ## Limitations - arXiv and bibliographic databases were not reachable from the analysis environment, so the literature was sampled through curated lists; three of the five centre on harness engineering, and the frame favours recent, English-language work. - Codes rest on titles, list annotations and curators' section headings, not on full texts. - The manual coder in the audit was the AI assistant that wrote the coding rules, so κ measures consistency between the rules and careful reading, not agreement between independent human experts. - LoopArena is one benchmark, one domain and one worker model, analysed from aggregate data. - Simulation parameters are assumptions, not estimates; E6 varies them to show which conclusions are robust. *Independent personal research. Does not represent the views of any employer or institution.* ## Licence and citation CC BY 4.0. Titles, descriptions and URLs of sources and repositories are those published by their authors and the list curators. Raw third-party files are not included here; they are in the GitHub repository under their owners' terms. ```bibtex @misc{chongmankhong2026context, author = {Chongmankhong, Thanatchon}, title = {Context Engineering and Loop Engineering for LLM Agents: A Systematic Review, Unified Taxonomy, and Use-Case Fit Framework}, year = {2026}, howpublished = {Preprint}, doi = {10.5281/zenodo.24216599}, url = {https://thanatchon36.github.io/context-loop-engineering/paper/} } ```