# WP35 — Phase 0 feasibility report

Date: 2026-11-01 (UTC-4). Executor: Claude Code (claude-fable-5-1) acting autonomously under the PI's mandate
(`docs/MANDATE_VERBATIM.txt`). Nothing in this report is inferred: every value below was read from a command output
recorded in `results/logs/execution_log.md`.

## 1. Environment

| Item | Observed |
|---|---|
| Machine | Apple M5 Max laptop, 18 cores, 48 GiB RAM, 1.0 TiB free on the data volume (df) |
| Cluster | MacLustr (29 Macs on LAN) reachable through `mld`/SSH; not required for Phases 0-4, may be used for simulation sweeps (Phase 5) |
| Python | `uv` 0.11.7 manages a project venv at `.venv` (CPython 3.12.13). System python3 is 3.14.4 (not used) |
| Pinned deps (pyproject.toml) | numpy 2.3.4, pandas 2.3.3, polars 1.36.1, scipy 1.16.2, statsmodels 0.14.5, linearmodels 6.1, scikit-learn 1.7.2, lifelines 0.30.0, networkx 3.5, matplotlib 3.10.7, pymc 5.26.1, arviz 0.22.0, pyreadr 0.5.3, openpyxl 3.1.5, pyarrow 21.0.0, pytest 8.4.2 — import check passed |
| LaTeX | TeX Live: latexmk 4.88 (9 March 2026), pdflatex, lualatex, bibtex at /Library/TeX/texbin; UQO template (`_modele-working-paper`) copied into `paper/` |
| Other tools | git, make, R 4.6.1 (no readxl), pandoc, jq, curl, wget, shasum |
| Connectivity | HTTP 200 on seshatdatabank.info, rug.nl (Maddison), d-place.org, github.com, v-dem.net, zenodo.org, api.github.com |
| LLM APIs | No `ANTHROPIC_*` / `OPENAI_*` variables in the shell environment; the `anthropic` MCP server failed to connect this session. **No paid API calls are planned or needed**: the research is executed by the agent itself; literature search uses public web search / Crossref. Any future API use requires PI approval. |
| Existing repository | None for WP35. Series conventions read from `Recherche-UQO/HARMONISATION.md` (numbering: WP34 exists, so this is WP35; new category folder `economie/`). |

## 2. Data sources: actual releases retrieved (raw files preserved, SHA-256 in `data/metadata/source_registry.json`)

| Key | Release actually obtained | Bytes | Licence (as read from source) | Notes from first inspection |
|---|---|---|---|---|
| seshat_polaris2026 | Polaris-2026 snapshot, GitHub `Seshat-Global-History-Databank/build_polaris_dataset` commit 55ca5fbc2b (2026-02-25) | 2,899,093 | MIT (repo LICENSE) | 8 sheets: Variables (173), Polities (626 rows; 42 `seshat_region`; start_year -13600…1921), General (7,239 rows, 516 polities, 21 vars), Social complexity (26,033 rows, 619 polities, 77 vars; `value_from/value_to`, `year_from/year_to`, `is_disputed`, `is_uncertain`), Warfare (17,536 rows, 376 polities, 49 vars), Luxury goods (1,191 rows), Religion (5,466 rows), **CrisisDB – Crisis consequences (166 rows, 150 polities; 23 consequence columns coded P/A/IP/IA/SU)** |
| seshat_equinox2020 | Equinox-2020 spreadsheet (05.2023) from the same repo | 1,720,038 | MIT (repo) / AGPL-3 on Zenodo 6629022 | single long-format sheet (NGA, Polity, Section, Subsection, Variable, Value.From/To, Date.From/To, Fact.Type, notes) |
| seshat_equinox_continuous_polities | Zenodo 8120128 `Timescale.SM_Data.xlsx`; MD5 verified = 57a4f436f8c7e62c80c4a49a67c6fe86 | 70,197 | CC-BY-4.0 | 1,396 rows: NGA, PolID, AbsTime, RelTime, SPC1, Culture.Sequence, Institutions.Sequence |
| maddison_2023_xlsx / _dta | MPD 2023 (dataverse.nl datafiles 421302/421303) | 4,903,804 / 10,892,389 | academic use with citation | `Full data` sheet: countrycode, country, region, year, gdppc, pop |
| dplace_ea_v3_3 | D-PLACE Ethnographic Atlas CLDF v3.3 (2025-12-19), zip of commit 5aa46ee | 7,039,773 | CC BY-NC 4.0 | cldf/{societies,variables,codes,data}.csv + sources.bib; raw/ folder with original tables |
| vdem_v16_rdata / _codebook | V-Dem v16 from the V-Dem Institute's official `vdemdata` R package (commit f4dd26922e, 2026-03-17) | 33,878,255 / 148,633 | V-Dem terms (non-commercial, cite); package GPL-3 | `vdem` data frame 28,092 country-years × 4,618 columns, years 1789–2025, 202 countries; codebook 781 rows |

Sampling check: every file opened and parsed with pandas/openpyxl/pyreadr/zipfile (see execution log). No coverage
number above is a guess; each is a row count printed by the parser on the raw file.

## 3. Feasibility assessment per phase

- **Historical panel (Seshat)** — feasible. D-candidates: `polity_population`, `polity_territory`,
  `population_of_the_largest_settlement`, `settlement_hierarchy`, `administrative_level`, `military_level`,
  `religious_level`. K-candidates: `full_time_bureaucrat`, `examination_system`, `merit_promotion`,
  `specialized_government_building`, `formal_legal_code`, `judge`, `court`, `professional_lawyer`, writing/records
  (`written_record`, `script`, `lists_tables_and_classification`, `calendar`, `practical_literature`…), communication
  (`courier`, `postal_station`, `general_postal_service`, `road`, `bridge`, `canal`, `port`), money (`token`, `precious_metal`,
  `foreign_coin`, `indigenous_coin`, `paper_currency`), standards (`*_measurement_system`). Temporal resolution is
  *per fact* (year_from/year_to), so the panel must be built on observed dates, not interpolated.
  Outcome Y: CrisisDB crisis consequences (independent coding: decline, collapse, civil_war, uprising, fragmentation,
  conquest, …) covers 150 polities only → **the H2 event-risk test will have a small-N outcome**; polity end/transition
  (`polity_succeeding_entity`, end_year, relationship to preceding entity) is a second, larger but less independent outcome.
- **Ethnographic cross-section (D-PLACE EA)** — feasible for H1-type scale/organization associations with
  Galton controls (language family, region). No time dimension.
- **Modern panel (V-Dem + Maddison)** — feasible as a separate, modern replication panel (country-year); the
  mandate forbids mapping it onto ancient polities and forbids using democracy indices as capacity.
- **Simulations** — feasible locally (networkx + numpy); sweeps can be parallelised on the cluster if needed.
- **Paper** — latexmk + UQO template available.

## 4. Risks identified before pre-registration

1. Small outcome N in CrisisDB (150 polities, with 42 regions) → low power; pre-register a minimum-detectable-effect
   statement and treat the test as confirmatory only for the primary outcome.
2. Overlap between D and K indicators and Seshat's own aggregate complexity characteristics (Turchin et al. 2018 PC1):
   D and K will be built from disjoint indicator sets fixed in the pre-registration.
3. Reverse causality: crisis coding may inform the coding of institutions in the same period → use lagged predictors
   with observed dates and exclude facts dated inside the crisis window.
4. Uncertainty: `value_from/value_to`, `year_from/year_to`, `is_uncertain`, `is_disputed` must be carried through
   (interval midpoints + bounds as sensitivity).
5. Licence: D-PLACE is BY-NC (academic use OK, no redistribution of raw data); V-Dem/Seshat raw files not redistributed in the repo.

Decision: proceed to Phase 1 (literature + pre-registration) **without** computing any D/K-outcome association.