# WP44 — Research plan (operational design, v1.0, 2026-11-02)

This document turns the mandate (`CLAUDE.md`) into the concrete design implemented by the code, the protocol and the
manuscript. Decisions the mandate left open are listed in §10 with their rationale; they are the executing agent's until
the principal investigator (PI) confirms them (human-review checkpoint H-0).

## 1. Scientific object and estimands

Population: the computer-science research literature indexed by at least one of DBLP, OpenAlex (any topic in field 17,
Computer Science), arXiv (set `cs`), or the ACL Anthology, first published between 2018-01-01 and 2026-11-02.
Two windows: **baseline** B = [2018-01-01, 2022-11-29] and **post** P = [2022-11-30, 2026-11-02]. The boundary is the public
release of ChatGPT; it is a reference date for description and for interrupted-time-series designs, *not* a causal
treatment (CLAUDE.md §6.5). Monthly series are the primary temporal unit; years are secondary.

Primary descriptive estimands (RQ1, RQ2): monthly counts of works per category c, N_{c,t}; composition shares
s_{c,t} = N_{c,t}/N_t; growth rates and their pre/post structural-break statistics; citation-based influence (field- and
age-normalised); collaboration/interdisciplinarity indices (RQ4–RQ6). Inferential estimands (RQ2, RQ10): breakpoint
locations and magnitudes with placebo distributions; out-of-time precision/recall of the emerging-topic detector.

## 2. Sources, access and what each contributes (verified 2026-11-02, see `reports/PHASE0_FEASIBILITY.md`)

| Source | Access used | Licence | Role in the corpus |
|---|---|---|---|
| OpenAlex public snapshot, release 2026-10-14 | S3 public bucket, JSONL, 1,898 works files (675 GB), ~110 MB/s from M3U96b | CC0 | backbone: works, authorships + institutions + countries, venues, topics (field/subfield/topic), `referenced_works` (citation graph), abstracts (inverted index), OA status, FWCI, counts_by_year |
| OpenAlex API (no key, $0.10/day shared budget) | REST, polite `mailto` | CC0 | incremental works 2026-10-14 → 2026-11-02; spot checks; free single-record lookups for reference verification |
| DBLP XML dump (2026-11-02, 1.12 GB gz, MD5 verified) | HTTPS | ODC-BY 1.0 | CS-scope gold standard, venue streams, author PIDs, DOI links; defines "CS venue" |
| arXiv OAI-PMH, set `cs`, 2018-01 → 2026-11 | OAI-PMH, 3.5 s/request, raw XML kept | arXiv metadata terms (CC0 for metadata) | preprints, categories (cs.XX, cross-lists), versions, licences; preprint–publication linking |
| ACL Anthology BibTeX (anthology+abstracts.bib, 138,796 entries) | HTTPS | CC BY 4.0 | NLP venue coverage check, abstracts for ACL papers |
| Crossref REST API (polite pool) | REST | open metadata | reference verification, `is-referenced-by-count` cross-check sample, publisher/type cross-check |
| Semantic Scholar Academic Graph | **not available**: API key required for bulk/dataset access; anonymous calls return HTTP 429 | — | recorded as unavailable (CLAUDE.md §1: record inaccessible data explicitly); influential-citation and citation-intent signals are therefore not used |
| Papers with Code / benchmark repositories, GitHub | HTTPS, public dumps where available | varies | RQ7/RQ9 (benchmark proliferation, code availability) — scoped in Phase 3 |
| Full texts | arXiv e-prints (rate-limited, licence-respecting), OA locations from OpenAlex/Unpaywall | per-paper licence | ≤ 10,000 high-priority papers; never through paywalls |

No credential was assumed: the OpenAlex key, Semantic Scholar key and Kaggle credentials were tested as absent.

## 3. Corpus construction (Phase 1) — PRISMA-style stages

S0 Identification: all OpenAlex works with any topic in field 17 and `publication_year` 2018–2026 (full records kept),
plus DBLP records 2018–2026, arXiv `cs` records created 2018–2026, ACL entries 2018–2026. A *slim* table of **every**
OpenAlex work (465 M rows: id, DOI, year, type, primary topic/subfield/field/domain, cited_by_count, flags) is kept to resolve
cited works outside CS.
S1 Record linkage: OpenAlex ↔ DBLP by DOI (lower-cased), else by arXiv ID (DBLP `ee` → arxiv.org/abs), else by normalised
title + year + first-author surname (blocking on title hash); OpenAlex ↔ arXiv by `locations[].landing_page_url`/`ids`
and DOI (10.48550/arXiv.*); ACL by DOI/URL. Linkage precision is estimated on a 500-pair manually reviewed sample (H-1).
S2 Deduplication: one node per *work*; preprint and published version are merged into a single node with
`preprint_date`, `publication_date`, `version_relationship` kept (CLAUDE.md §6.2: preprint duplication).
S3 Eligibility: (i) `publication_year` in window; (ii) type in {article, preprint, book-chapter, proceedings-article,
dissertation, report, book} — exclude paratext, editorials, errata, retracted (kept with flag `is_retracted`), XPAC
expansion records (`is_xpac`); (iii) CS membership: DBLP-indexed **or** OpenAlex primary field = CS **or** (secondary CS
topic **and** arXiv primary category cs.*). Secondary-only CS topics without DBLP/arXiv support form a *penumbra* set
reported separately. Sensitivity: results re-estimated on the strict core (DBLP ∩ OpenAlex primary CS).
S4 Classification: 40-category hierarchical ontology (`taxonomy/ontology.yaml`), multi-label, assigned by rules over
OpenAlex topics/subfields, arXiv categories, DBLP venue streams and keyword patterns; validated on manually reviewed
samples (Phase 2, H-2) and by an LLM-based second rater with reported agreement.
S5 Synthesis populations: (a) mapping corpus (all eligible); (b) deeply classified set (≥ 100 k: abstracts + embeddings +
full feature vectors); (c) high-priority set (≥ 10 k: citation-ranked within category × year with OA full text);
(d) landmark set (≥ 1 k: top field-normalised impact, reviewed by specialist agents and flagged for PI review, H-3).

Every stage writes counts to `data/metadata/prisma_counts.json`, rendered as the PRISMA flow figure.

## 4. Taxonomy (Phase 2)

`taxonomy/ontology.yaml`: 40 leaf categories under 8 top-level areas (AI & ML; Vision, Language & Perception; Systems &
Infrastructure; Theory & Foundations; Security, Privacy & Cryptography; Software & Human Factors; Emerging Computing
Paradigms; Computing & Science). Each leaf has: definition, parent, OpenAlex topic IDs (weighted), subfield IDs, arXiv
categories, DBLP streams, keyword regexes with weights, exclusions, and provenance of the mapping. Assignment score per
category = weighted sum of matched signals; multi-label with threshold; `primary_category` = argmax. Validation: 2,000
stratified works (50 per category) rated by two independent LLM raters with the rubric in `taxonomy/VALIDATION.md`,
Cohen's κ and per-category precision/recall; a 200-work subsample is flagged for PI review (H-2).

## 5. Knowledge graph (Phase 3)

Nodes: works, authors (OpenAlex A-ids), institutions, countries, venues (sources), topics, taxonomy categories, arXiv
categories; *method/model/dataset/benchmark* and *claim* nodes only for the high-priority set, extracted from abstracts
(and full text where lawfully available) with a schema-constrained LLM extractor; each extracted edge carries
`provenance` (work id + text span), `confidence`, `verification_status` ∈ {unverified, machine-verified, human-verified}.
Citation edges come from `referenced_works` (verified by construction). Semantic edges from embeddings (k-NN, cosine).
Storage: Parquet tables + DuckDB views; graph analytics with igraph (python-igraph) on M3U96b (96 GB RAM).

## 6. Statistics (Phase 4) and structural change

- Bibliometrics: monthly/annual counts; shares; CAGR; field-normalised citation score FNCS = c / E[c | field, year, type];
  citation velocity (citations within 12/24 months, computed from `counts_by_year`); Gini/Hirsch-type inequality; venue
  concentration (HHI); author/institution/country collaboration networks; Rao–Stirling and Leinster–Cobbold diversity.
- Citation-age bias: all cross-period comparisons use fixed citation windows and year-normalisation; FWCI used as a
  secondary measure with its known limits.
- Structural change: interrupted time series on log monthly counts with seasonality (OLS-HAC, Newey–West), Bai–Perron
  multiple-breakpoint estimation, PELT/binary-segmentation changepoints, Bayesian online changepoint detection; pre-trend
  diagnostics (placebo breaks at every month of the baseline), alternative event dates (2020-05 GPT-3, 2022-04 DALL·E 2,
  2022-11 ChatGPT, 2023-03 GPT-4, 2024-09 o1), donut windows; sensitivity to OpenAlex indexing lag (right-truncation
  corrected by indexing-delay profiles estimated from `created_date`).
- Dynamic networks: yearly topic co-occurrence and citation-flow matrices between taxonomy categories; modularity and
  cross-area flow shares; ERGM-style descriptives are out of scope.

## 7. Emerging-research detection (Phase 6, original contribution)

Unit: topic candidates = (i) OpenAlex topics, (ii) semantic clusters (HDBSCAN on embeddings, re-estimated per
training cutoff without look-ahead), (iii) keyword n-grams. Signals at month t using data ≤ t only: growth anomaly,
semantic novelty (distance to pre-t centroid), citation acceleration, interdisciplinary connectivity, methodological
adoption (share of works citing a method node), benchmark proliferation, open-source activity. Label at t+h: top-decile
growth and citation share at horizon h ∈ {12, 24} months. Training cutoffs 2019–2022; test cutoffs 2023–2024 with
outcomes observed up to 2026-11. Models: regularised logistic, gradient boosting, and a simple z-score baseline;
metrics: AUC, precision@k, lead time. Known-emergence validation set: LLMs, diffusion models, NeRF, agents, RAG,
state-space models, vision-language models, mechanistic interpretability (defined before estimation, listed in
`protocol/PROTOCOL.md`).

## 8. Agents and artefacts

The executing session is Agent A (Research Director). Subagents are spawned with written briefs (`agents/briefs/`),
return structured artefacts (JSON/CSV/Markdown with declared schema) and never approve their own findings: every
scientific artefact is checked by a different agent (K or N) before it enters the manuscript (`agents/REGISTER.md`).

## 9. Infrastructure

Primary compute/storage: **M3U96b** (M3 Ultra 28 c, 96 GB, 375 GB free; `~/wp44_uqo/`, venv Python 3.12 via uv:
orjson, pyarrow, polars, duckdb, zstandard, lxml, requests). Secondary: M4M64a (540 GB free) for embeddings; other LAN
nodes for embarrassingly parallel embedding/extraction batches when it materially helps. Laptop = controller and git
repository (`Recherche-UQO/informatique/wp44`, origin `gitsrv:srv/git/wp44_uqo.git`). Data are not in git; ledgers,
manifests, hashes and PRISMA counts are.

## 10. Decisions taken by the executing agent (pending PI confirmation, H-0)

| # | Decision | Rationale |
|---|---|---|
| D1 | Snapshot over API for OpenAlex | mandate §3 prefers snapshots; the API budget without a key is $0.10/day (≈ 200 k works/day) |
| D2 | Keep a slim table of all 465 M works | required for interdisciplinary citation flows (RQ4) and cited-work normalisation |
| D3 | Semantic Scholar not used | key required; recorded as unavailable rather than scraped |
| D4 | CS membership rule (§3 S3) with strict-core sensitivity | avoids both AI-only bias and over-inclusion from secondary topics |
| D5 | Manuscript directory `manuscript/` (mandate §12) instead of the UQO `paper/` convention | mandate is explicit; deviation noted in README |
| D6 | Historical raw snapshot kept as filtered subset + per-file SHA-256 + manifest, not the 675 GB bucket copy | public versioned release; hashes allow verification; storage |
| D7 | LLM raters (Claude) used for taxonomy validation and knowledge-graph extraction, with PI review checkpoints | mandate §1 requires human checkpoints; PI is the human |