Round 2 (Codex, 10 items: 6 ok, 2 bug, 2 hardening): outcomes recorded in supplementaryMaterial/punishmentSummary divs (1670s–80s sessions) are now structured on the sessions row as summary_items {id,type,category,subcategory,text,found_in,linked_ids}; decode fails loudly instead of errors=replace (BOM-aware); rs→nearest div1 ancestor; collision-safe duplicate-id suffixing with a warning; dead code removed. Trials/OA rows unchanged (verified identical on a 20-session sample). Round 1 findings (12): punishments attached via defendantPunishment joins file-wide and shown in both the defendant's trial and the trial holding the text (`found_in`); <lb/> joins get a space before an upper-case continuation; person `name` excludes nested <persName> and falls back to given+surname; `charges.unresolved_ids` keeps targets that resolve to nothing; persons enriched (occupation/place/aliases) identically in trials, sessions and ordinarys_accounts; `xml` decoded with the declared encoding (no errors='…
Parser fixes after review: cross-trial punishments, <lb/> spacing, nested names, unresolved charge targets, join enrichment (occupation/place/aliases) on all configs, declared-encoding XML, null person type
Review findings addressed (12): punishments attached via defendantPunishment joins file-wide and shown in both the defendant's trial and the trial holding the text (`found_in`); <lb/> joins get a space before an upper-case continuation; person `name` excludes nested <persName> and falls back to given+surname; `charges.unresolved_ids` keeps targets that resolve to nothing; persons enriched (occupation/place/aliases) identically in trials, sessions and ordinarys_accounts; `xml` decoded with the declared encoding (no errors='ignore'); strict parse with recovery fallback + warnings; deterministic defendant order; persons.type null when the source records no role; empty names null.
Per-trial `<div1>` XML (trials), whole session file (sessions), whole pamphlet file (ordinarys_accounts) — verbatim from OBO XML 7.2. Gives character offsets for every tagged entity, the nameAlias/persNamePlace joins and <xptr> page refs the flattened columns don't model. Cost: trials 245→559 MB, sessions 225→556 MB (zstd). Resharded to 6 files each.
v2: per-trial structured configs from Old Bailey Online XML 7.2
Replaces the session-level `default` config with three configs parsed from the same OBO XML 7.2 (figshare 10.15131/shef.data.4775434.v2 / dhi.ac.uk zip):
- `trials` (default) — 197,751 rows, one per `<div1 type="trialAccount">`: full text, typed persons (defendant/victim/witness/juror/judiciary/unspecified with gender, age, occupation, place), offences (9 categories / 57 subcategories, crime date, place, linked victims), verdicts and punishments with categories, and the `criminalCharge` joins (defendant → offence → verdict); punishments resolved from part-1 `punishmentSummary` divs via `defendantPunishment` joins.
- `sessions` — 2,163 rows, full session text (front matter etc. included) + typed persons.
- `ordinarys_accounts` — 475 rows.
Also: licence corrected to CC BY-NC 4.0 (the source's licence.txt; the 2022 script hard-coded CC BY 4.0), `dataset_info` dropped (schema changed), parser added under `scripts/`. Previous parquet stays at revision e96ea6d. Card rewrite to follow in a …
