render-review-rows

render-review-rows

Design Principles

1. The filing is the single source of truth. Every fact traces to source_accession_number + verbatim source_text.
2. Fail closed on source ambiguity. Fail open on extraction incompleteness.
3. Facts before judgments. Proposals are raw events. Classification is Skill 8, not Skill 5.
4. Alex's collection instructions are the extraction spec. Used for methodology and taxonomy. Never as a factual source.
5. master.csv is the review artifact, not the estimation artifact.

Overview

Denormalize all extraction and enrichment artifacts into a 47-column review CSV. Pure mechanical assembly -- no judgment, no classification, no interpretation. If the input artifacts exist, this skill always succeeds.

When To Use

• After Skills 1-8 have completed (orchestrated or independent).
• To rebuild master_rows.csv after manual overrides or re-extraction.
• To rebuild the global Data/views/master.csv after any deal changes.

Input Artifacts

All paths relative to Data/deals/<slug>/:

Artifact	Source Skill
extraction/deal.json	5. extract-events
extraction/actors.jsonl	4. build-party-register
extraction/actors_extended.jsonl	5. extract-events (if new actors minted)
extraction/events.jsonl	5. extract-events
extraction/event_actor_links.jsonl	5. extract-events
enrichment/process_cycles.jsonl	7. segment-processes
enrichment/judgments.jsonl	8. classify-bids-and-boundary
extraction/audit_flags.json	6. audit-and-reconcile (optional)

Output Artifacts

Artifact	Description
master_rows.csv	47-column review CSV (one row per event-actor link)
review/review_status.json	Review status and flags for this deal
review/overrides.csv	Header-only CSV for reviewer corrections
Data/views/master.csv	Global rebuild: all Data/deals/*/master_rows.csv concatenated

Tools

Tool	Purpose
File read	Read all input artifacts from disk
File write	Write master_rows.csv, review_status.json, overrides.csv
File search	Find all Data/deals/*/master_rows.csv for global rebuild

Procedure

1. Read deal.json -> extract Group A (deal header) fields.
2. Read actors.jsonl + actors_extended.jsonl (if exists) -> build actor lookup.
3. Read events.jsonl -> build event list.
4. Read event_actor_links.jsonl -> build link list.
5. Read process_cycles.jsonl -> build cycle/round lookup.
6. Read judgments.jsonl -> build classification lookup.
7. Read audit_flags.json (if exists) -> build review flag inputs.

8. For each event:
   a. Look up event-actor links for this event.
   b. If links exist: emit one row per link (event + actor columns).
   c. If no links: emit one row with blank actor columns.
   d. Populate all 47 columns per the column spec.
   e. Derive bid_note from event_type using the mapping table.
   f. Derive review flags (approximate_date, missing_nda, etc.).

9. Sort rows: deal_slug -> event_date -> event_type priority -> actor_alias.
10. Write master_rows.csv with header + sorted rows.
11. Write review/review_status.json with status, flags, extraction date.
12. Write review/overrides.csv (header only).
13. Rebuild global Data/views/master.csv:
    - Find all Data/deals/*/master_rows.csv files.
    - Validate headers are consistent across all deals.
    - Concatenate all rows sorted by deal_slug.
    - Write to Data/views/master.csv.

Gate

master_rows.csv exists in Data/deals/<slug>/.

Failure Policy

Always succeeds if input artifacts exist. This is pure denormalization -- no judgment calls, no ambiguity. If an input artifact is missing, skip its columns (leave blank) and note in review_status.json.

Required Reading

1. references/column-spec.md -- 47-column specification, sorting rules, bid_note mapping, review flag derivation, global master.csv rebuild