WP44: incident I-6, appendix writer brief
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
5 changed files +27 −5
added
agents/briefs/WRITER_APPENDICES_AEH.md
+20 −0
@@ -0,0 +1,20 @@ | |||||
| 1 | +# Brief — Appendices A (taxonomy), E (data dictionary), H (reproducibility) — files manuscript/appendices/appA_taxonomy.tex, appE_data_dictionary.tex, appH_reproducibility.tex | ||||
| 2 | + | ||||
| 3 | +Follow agents/briefs/WRITER_COMMON.md. These appendices are reference material: precise, tabular, no padding. | ||||
| 4 | +- Appendix A — The ontology: read taxonomy/ONTOLOGY.md and taxonomy/ontology.yaml (do not paste the YAML). Give: the 8 areas and | ||||
| 5 | + 40 leaves with definitions (one table, longtable), the parent/child edges, the signal types and the assignment rule with the | ||||
| 6 | + threshold, per-category counts of mapped OpenAlex topics / DBLP streams / keyword patterns (from the ontology; compute with a | ||||
| 7 | + short Python snippet), the list of unmapped CS topics with reasons, and the known ambiguities (ONTOLOGY.md §2). Mention that | ||||
| 8 | + the validation sample results are reported in Part IX (do not invent them). | ||||
| 9 | +- Appendix E — Data dictionary: read ingestion/build_corpus.py (schemas S), ingestion/openalex_snapshot.py (slim schema), | ||||
| 10 | + systematic_review/define_corpus.py (corpus_works added columns: in_dblp, in_acl, arxiv_id_linked, arxiv_created, arxiv_primary, | ||||
| 11 | + e1–e7 flags, xpac_kept_dblp, cs_core, cs_penumbra, eligible, period, first_date, first_month, date_precision, duplicate_of, | ||||
| 12 | + preprint_work_id, version_relationship), the linkage tables, bibliometrics/03 (work_citation_windows), knowledge_graph/01 | ||||
| 13 | + (edge tables), taxonomy/classify.py outputs (work_categories, work_primary), embeddings/embed_corpus.py outputs. One longtable | ||||
| 14 | + per table family: column, type, definition, source field (OpenAlex path or derivation). Be exact; derive from the code. | ||||
| 15 | +- Appendix H — Reproducibility: read reproducibility/RUNBOOK.md, reproducibility/environment_M3U96b.txt, data/raw/README.md, | ||||
| 16 | + reports/INCIDENTS.md, protocol/amendments.csv. Give the step table (commands, inputs, outputs, measured times), the environment | ||||
| 17 | + (versions of the key packages from the lock file), the provenance mechanisms (manifests, SHA-256 ledgers, MD5 of DBLP, OAI raw | ||||
| 18 | + pages, API ledger), the incident log and what it implies for reproduction, the licence of each source, and what is and is not | ||||
| 19 | + redistributed. Keep numbers to those present in those files (file counts, sizes, times) — they are provenance facts, not corpus results. | ||||
| 20 | +Target: A ≈ 10–14 pages, E ≈ 10–14 pages, H ≈ 6–8 pages. No new bib entries needed except possibly software citations (verify). | ||||
modified
manuscript/references_part02.bib
+2 −2
@@ -1755,10 +1755,10 @@ | |||||
| 1755 | note = {verified: crossref 2026-11-03} | 1755 | note = {verified: crossref 2026-11-03} | ||
| 1756 | } | 1756 | } | ||
| 1757 | 1757 | ||||
| 1758 | −@article{allamanis2019survey, | 1758 | +@article{allamanis2018survey, | ||
| 1759 | title = {{A} {Survey} of {Machine} {Learning} for {Big} {Code} and {Naturalness}}, | 1759 | title = {{A} {Survey} of {Machine} {Learning} for {Big} {Code} and {Naturalness}}, | ||
| 1760 | author = {Allamanis, Miltiadis and Barr, Earl T. and Devanbu, Premkumar and Sutton, Charles}, | 1760 | author = {Allamanis, Miltiadis and Barr, Earl T. and Devanbu, Premkumar and Sutton, Charles}, | ||
| 1761 | − year = {2019}, | 1761 | + year = {2018}, | ||
| 1762 | journal = {ACM Computing Surveys}, | 1762 | journal = {ACM Computing Surveys}, | ||
| 1763 | volume = {51}, | 1763 | volume = {51}, | ||
| 1764 | number = {4}, | 1764 | number = {4}, | ||
modified
manuscript/references_verification_part02.csv
+1 −1
@@ -163,7 +163,7 @@ demoura2015lean,10.1007/978-3-319-21401-6_26,crossref,The Lean Theorem Prover (S | |||||
| 163 | demoura2021lean4,10.1007/978-3-030-79876-5_37,crossref,The Lean 4 Theorem Prover and Programming Language,exact, | 163 | demoura2021lean4,10.1007/978-3-030-79876-5_37,crossref,The Lean 4 Theorem Prover and Programming Language,exact, | ||
| 164 | mathlib2020,10.1145/3372885.3373824,crossref,The lean mathematical library,exact, | 164 | mathlib2020,10.1145/3372885.3373824,crossref,The lean mathematical library,exact, | ||
| 165 | hales2017kepler,10.1017/fmp.2017.1,crossref,A FORMAL PROOF OF THE KEPLER CONJECTURE,exact, | 165 | hales2017kepler,10.1017/fmp.2017.1,crossref,A FORMAL PROOF OF THE KEPLER CONJECTURE,exact, | ||
| 166 | −allamanis2019survey,10.1145/3212695,crossref,A Survey of Machine Learning for Big Code and Naturalness,exact,key renamed from allamanis2018survey to match year of record | 166 | +allamanis2018survey,10.1145/3212695,crossref,A Survey of Machine Learning for Big Code and Naturalness,exact,"year set to 2018 (CSUR 51(4), July 2018; OpenAlex publication_date 2018-07-31)" | ||
| 167 | bernstein2017pqc,10.1038/nature23461,crossref,Post-quantum cryptography,exact, | 167 | bernstein2017pqc,10.1038/nature23461,crossref,Post-quantum cryptography,exact, | ||
| 168 | shor1997algorithms,10.1137/S0097539795293172,crossref,Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer,exact, | 168 | shor1997algorithms,10.1137/S0097539795293172,crossref,Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer,exact, | ||
| 169 | nist2024fips203,10.6028/NIST.FIPS.203,crossref,Module-lattice-based key-encapsulation mechanism standard,exact, | 169 | nist2024fips203,10.6028/NIST.FIPS.203,crossref,Module-lattice-based key-encapsulation mechanism standard,exact, | ||
modified
manuscript/sections/part02_evolution.tex
+2 −2
@@ -96,7 +96,7 @@ The third is the \emph{open-source framework as a shared instrument}. TensorFlow | |||||
| 96 | 96 | ||||
| 97 | The fourth is the \emph{entry of industry as a producer rather than a consumer of research}. The laboratories of Google, DeepMind, Facebook, Microsoft and Baidu published the decisive results of the period in the academic venues and on arXiv, and their access to data and compute began to separate them from universities. \citet{ahmed2020dedemocratization} documented the growing presence of large technology firms and elite universities at the major AI conferences and the relative decline of other institutions; the later and broader analysis of \citet{ahmed2023industry} found that industry produced an increasing share of the largest models, the most-cited papers and the leading benchmarks. Whether that concentration intensified after 2022 is hypothesis H6 of the protocol. | 97 | The fourth is the \emph{entry of industry as a producer rather than a consumer of research}. The laboratories of Google, DeepMind, Facebook, Microsoft and Baidu published the decisive results of the period in the academic venues and on arXiv, and their access to data and compute began to separate them from universities. \citet{ahmed2020dedemocratization} documented the growing presence of large technology firms and elite universities at the major AI conferences and the relative decline of other institutions; the later and broader analysis of \citet{ahmed2023industry} found that industry produced an increasing share of the largest models, the most-cited papers and the leading benchmarks. Whether that concentration intensified after 2022 is hypothesis H6 of the protocol. | ||
| 98 | 98 | ||||
| 99 | −The propagation pattern of this wave was \emph{methodological} and sequential. Deep learning entered computer vision and speech first, natural language processing second (roughly 2014--2016, with the sequence-to-sequence results), reinforcement learning and robotics third, and then, more slowly and with more resistance, the subfields whose objects are not perceptual: program analysis \citep{allamanis2019survey}, systems (learned index structures and learned scheduling), databases, and scientific computing \citep{raissi2019pinn}. In each case the receiving subfield first treated learning as a tool for a pre-existing problem, and only later, if at all, reorganised its problems around it. That two-stage pattern, tool adoption followed by problem redefinition, is the template against which Part~IX reads the post-2022 movement of methods across the forty categories. | 99 | +The propagation pattern of this wave was \emph{methodological} and sequential. Deep learning entered computer vision and speech first, natural language processing second (roughly 2014--2016, with the sequence-to-sequence results), reinforcement learning and robotics third, and then, more slowly and with more resistance, the subfields whose objects are not perceptual: program analysis \citep{allamanis2018survey}, systems (learned index structures and learned scheduling), databases, and scientific computing \citep{raissi2019pinn}. In each case the receiving subfield first treated learning as a tool for a pre-existing problem, and only later, if at all, reorganised its problems around it. That two-stage pattern, tool adoption followed by problem redefinition, is the template against which Part~IX reads the post-2022 movement of methods across the forty categories. | ||
| 100 | 100 | ||||
| 101 | \subsection{The transformer and pretraining era, 2017--2022} | 101 | \subsection{The transformer and pretraining era, 2017--2022} | ||
| 102 | \label{sec:p2-transformer} | 102 | \label{sec:p2-transformer} | ||
@@ -191,7 +191,7 @@ The programming-languages literature of the baseline is marked by the success of | |||||
| 191 | 191 | ||||
| 192 | Proof assistants themselves reached a new stage. The verified C compiler CompCert \citep{leroy2009compcert} and the verified microkernel seL4 \citep{klein2009sel4} had shown a decade earlier that production software could be proved correct; the Kepler conjecture was formally proved in 2017 after a twelve-year effort \citep{hales2017kepler}; and the Lean theorem prover \citep{demoura2015lean}, re-implemented as a programming language in its fourth version \citep{demoura2021lean4}, acquired a community-maintained mathematical library whose scale and organisation were new in formalised mathematics \citep{mathlib2020}. The relevance to the post-2022 period is direct: Lean and its library became the environment in which language models were later asked to prove theorems, so that formal verification, one of the categories predicted \emph{not} to break in H3, is also one where the contact with AI is most visible in the recent literature. The hypothesis is therefore a sharp one. | 192 | Proof assistants themselves reached a new stage. The verified C compiler CompCert \citep{leroy2009compcert} and the verified microkernel seL4 \citep{klein2009sel4} had shown a decade earlier that production software could be proved correct; the Kepler conjecture was formally proved in 2017 after a twelve-year effort \citep{hales2017kepler}; and the Lean theorem prover \citep{demoura2015lean}, re-implemented as a programming language in its fourth version \citep{demoura2021lean4}, acquired a community-maintained mathematical library whose scale and organisation were new in formalised mathematics \citep{mathlib2020}. The relevance to the post-2022 period is direct: Lean and its library became the environment in which language models were later asked to prove theorems, so that formal verification, one of the categories predicted \emph{not} to break in H3, is also one where the contact with AI is most visible in the recent literature. The hypothesis is therefore a sharp one. | ||
| 193 | 193 | ||||
| 194 | −Software engineering had already begun to absorb learned models before 2022. The survey of \citet{allamanis2019survey} organised a decade of ``big code'' research around the naturalness hypothesis, that source code is as statistically regular as natural language; the CodeXGLUE benchmark collection \citep{lu2021codexglue} standardised evaluation; and the Codex models \citep{chen2021codex} made program synthesis from natural language a product. The software-engineering category is thus expected to show an early and strong AI signal, and its baseline is not an AI-free one. | 194 | +Software engineering had already begun to absorb learned models before 2022. The survey of \citet{allamanis2018survey} organised a decade of ``big code'' research around the naturalness hypothesis, that source code is as statistically regular as natural language; the CodeXGLUE benchmark collection \citep{lu2021codexglue} standardised evaluation; and the Codex models \citep{chen2021codex} made program synthesis from natural language a product. The software-engineering category is thus expected to show an early and strong AI signal, and its baseline is not an AI-free one. | ||
| 195 | 195 | ||||
| 196 | \subsection{Security and cryptography} | 196 | \subsection{Security and cryptography} | ||
| 197 | \label{sec:p2-security} | 197 | \label{sec:p2-security} | ||
modified
reports/INCIDENTS.md
+2 −0
@@ -6,3 +6,5 @@ | |||||
| 6 | | I-2 | 2026-11-03 ~15:50 | DuckDB "Too many open files" reading 1,898 Parquet files per table | linkage step failed | `compact_tables.py` (one Parquet file per table) + `ulimit -n 65536` | | 6 | | I-2 | 2026-11-03 ~15:50 | DuckDB "Too many open files" reading 1,898 Parquet files per table | linkage step failed | `compact_tables.py` (one Parquet file per table) + `ulimit -n 65536` | | ||
| 7 | | I-3 | 2026-11-03 ~16:30 | `multiprocessing.Pool` + pyarrow in spawned workers hangs on macOS (0 % CPU, no progress) for the classifier | classification test stalled 10 min | shard mode: independent processes started by `run_classify_shards.sh` (two nodes) | | 7 | | I-3 | 2026-11-03 ~16:30 | `multiprocessing.Pool` + pyarrow in spawned workers hangs on macOS (0 % CPU, no progress) for the classifier | classification test stalled 10 min | shard mode: independent processes started by `run_classify_shards.sh` (two nodes) | | ||
| 8 | | I-4 | 2026-11-03 ~16:58 | **M3U96b froze** (ping 100 % loss, SSH "host is down") while running the second snapshot pass (8 curl+parse workers), the SPECTER2 embedding job (MPS) and earlier DuckDB jobs concurrently | second pass (≈ 1,400/1,898 files done), the chained corpus v2 / classification, and the M3U96b embedding share interrupted; the three other embedding nodes unaffected | automatic restart expected (`mld harden`); all jobs are resumable from ledgers/shards; rule adopted: never run ingestion, embeddings and classification concurrently on one node | | 8 | | I-4 | 2026-11-03 ~16:58 | **M3U96b froze** (ping 100 % loss, SSH "host is down") while running the second snapshot pass (8 curl+parse workers), the SPECTER2 embedding job (MPS) and earlier DuckDB jobs concurrently | second pass (≈ 1,400/1,898 files done), the chained corpus v2 / classification, and the M3U96b embedding share interrupted; the three other embedding nodes unaffected | automatic restart expected (`mld harden`); all jobs are resumable from ledgers/shards; rule adopted: never run ingestion, embeddings and classification concurrently on one node | | ||
| 9 | +| I-5 | 2026-11-03 17:16 | **M3U96b froze again** two minutes after the second pass was restarted alone (8 download+parse workers) on a freshly rebooted node | pass 2 interrupted a second time; chain and classification watchers lost | pass 2 moved to the idle node M5M36 (10 workers, resumable ledger copied); M3U96b reserved for DuckDB/classification; chain now waits for a flag raised after the pass-2 outputs are copied back; Apple diagnostics show autobugcapture_Performance reports, no kernel panic found in the quick log search | | ||||
| 10 | +| I-6 | 2026-11-03 17:17 | **M4M64a froze a second time** within two minutes of restarting its SPECTER2 embedding share (batch 32); M3U96b rebooted at the same minute while running pass 2 alone | M4M64a embedding share lost again | M4M64a and M3U96b withdrawn from GPU/embedding and bulk-download work; ranges reassigned to M4M64c (58–77 queued) and M3U96a (0–29 queued); pass 2 on M5M36 | | ||||
| 9 | 11 | ||||