# Revised Holoscope leadership study — 9 October 2026

Open [the full HTML report](report.html). The searchable table contains **1,202 deduplicated leader–employee–company pairs**, with evidence for both endpoints and a typed explanation of the connection. The closer team/project/specific-unit layer contains **293 pairs**.

The original census of 1,929 people across 107 employers/divisions was checked again against native role titles, departments, research records, legacy packets and saved reporting. The revised census contains **1,990 people** across 108 scope organizations (the original 107 plus a sourced Microsoft parent-level scope); **84** have accepted program or personal-funding records. All 61 added people received the same secondary program/grant check. The scope is exhaustive for the frozen database and enumerated saved files, with targeted public-source verification.

## Results

Leader category is first. “Record” includes recorded EA-linked program participation or listed mentorship, or personal EA-linked funding. “No record” retains unknown history.

| Sample | Record/record | Record/no record | No record/record | No record/no record |
|---|---:|---:|---:|---:|
| All included leadership | 41 | 54 | 125 | 982 |
| Team, project and specific unit | 23 | 33 | 62 | 175 |

In the closer layer, employee-record shares are **41.1%** under record-positive leaders and **26.2%** under leaders with no record. Conversely, leader-record shares are **27.1%** among record-positive employees and **15.9%** among employees with no record. Both complementary percentages and the broad-union comparisons are in the conditional tables.

The report separates explicit reporting, named team leadership, project leadership, shared-unit inference, project advising and indirect department/company chains. It also provides participant/personal-grant-only, role-continuity and closest-leader-per-employee/company sensitivities. There is one source-attributed original hiring decision and one separate recruitment-involvement record. Leadership and retrospective program/funding records support descriptive association; selection claims require hiring authority, pre-selection dates and the relevant applicant pool.

## Reader files

- [Leadership table CSV](data/leadership_table.csv): one row per included pair, leader, subordinate, unit, layer, dates and direct evidence links.
- [Full pair evidence JSON](data/leadership_pairs_final.json) and [CSV](data/leadership_pairs_final.csv): every retained source observation, role and correction.
- [Counts](data/aggregate_counts.csv) and [conditional percentages](data/conditional_percentages.csv): all 15 samples and sensitivities, with denominators.
- [Every employee](data/revised_person_roster.csv) and [employee–company coverage](data/employee_company_coverage.csv): all 1,990 people and 2,067 employee–company units, including 1,258 units with no included leader.
- [Unresolved/excluded table](data/unresolved_leadership_table.csv) and [full candidate evidence](data/leadership_candidates.json): 2,331 candidate pairs, including uncertainty and rejected connections.
- [Canonical evidence graph](data/entity-graph.json) and [source register](data/source_ledger.csv).
- [Independent source audit](data/frontend_structural_audit.json), [independent aggregation audit](evidence/independent_aggregation_audit.json) and [capture provenance](raw/project-sources/capture-manifest.json).
- [Public package checksums](data/package_manifest.json): the files shipped in this revision directory, with byte sizes and SHA-256 hashes. The [original research manifest](data/original_research_manifest.json) preserves the complete local archive’s original 138-file hash ledger; the public manifest records which of those files are supplied here.

## Reproduction and archive scope

The public package provides the report, published data tables and evidence graph, source registers, audits, capture provenance and six Python scripts documenting aggregation, graph export, packaging, rendering and source corrections. The checksum inventory covers these shipped artifacts and excludes its own manifest file. The reader files support independent inspection of the reported counts and relationship evidence.

A full rebuild uses the complete local research archive. Its inputs include the sibling original-study data, the Holoscope snapshot, primary-role additions, review overrides, canonical identity maps, intermediate datasets, raw source captures and report CSS, JavaScript and font assets. The original research manifest retains the hashes and paths for that archive.

From the revision directory in the complete local research archive, run:

```sh
python3 scripts/aggregate_leadership.py
python3 scripts/export_revision_graph.py
python3 scripts/package_study.py
python3 scripts/build_report.py
```

The supplied `scripts/frontend_finalize_audit.py` and `scripts/frontend_independent_audit.py` document the source corrections using those archived inputs. The public capture manifest records capture provenance; the complete local archive retains the corresponding raw source files under `raw/` and `sources/`.

Database access used the full preserved export and repository snapshot. Secondary checks searched 285 captured program texts and 6,926 public funding/project records for each person and recorded alias, followed by review of matches. The revision searched 8,943 saved files and skipped 326 duplicate-content files. Unknown dates, self-reported ORCID employment, open-ended role continuity and AI-aggregated source records retain explicit quality flags.