# Red-team report for v0.4.0

> Post-release Viewer note, 15 August 2026: the original 26-configuration Data Studio fan-out repeatedly hit shared conversion locks. The mutable `main` card now exposes only `research_view`; all frozen v0.4.0 CSV, Parquet and release artifacts remain unchanged. Acceptance criterion 8 below reflects that operational correction.

## Verdict

Version 0.4.0 is an exploratory evidence index, not an empirical test of the agency-transfer hypothesis. The red-team found that v0.3.1 was structurally useful but its central pathway field could make inferred control shifts look observed. The release was therefore recoded rather than patched cosmetically.

## Findings and disposition

| Finding | v0.4.0 disposition |
|---|---|
| One pathway label mixed capability, attempted transfer, observed effect and preparedness | Replaced by separate `capability_status`, `control_mechanism_status`, `attempted_transfer_status`, `agency_change_status` and `agency_preservation_status` fields. |
| Researcher inference could be mistaken for an observed transfer | Added `maximum_conclusion`, `observed_evidence`, `researcher_inference`, `falsification_condition` and `comparator`. The release records zero observed agency changes. |
| Four older agency-transfer claims did not cover all core records | Added one case-specific control-shift hypothesis for each missing retrospective record and recoded all eight as `researcher_hypothesis` / `hypothesis_not_tested`. None has direct supporting evidence links. |
| Allegations and denials were often linked as underlying facts | Reclassified claim and evidence scope. The evidence layer now distinguishes `assertion_was_made`, `institutional_action`, `measurement`, `underlying_fact`, `review_scope_statement` and `researcher_inference`. |
| Selected sources were used to imply global absence | Reworded 12 claims as `not_identified_in_declared_review`; removed their 23 purported support links. These statements describe only the declared review set. |
| Numeric confidence fields implied calibration that did not exist | Removed claim-level `confidence`, `coder_confidence_in_label` and case-level `overall_confidence`. No probability or intercoder-reliability statistic is reported. |
| A 138,000 figure mixed accounts and transactions across geographies | Removed the incompatible aggregate observation. Its component facts remain claim-level context and are not summed as one comparable metric. |
| Transparent, authorised or satirical uses could be read as proven manipulation | Retained them as boundary cases. Disclosure alone proves neither harm nor agency preservation. |
| Official results and platform metrics could leak into causal interpretation | Kept all selected observations descriptive unless a recorded design supports a causal estimate. Official election data remain context only. |
| The validator could pass stale audits, a forged Parquet manifest or silent type coercion | The v0.4 validator compares stored and live audits, verifies all 26 CSV/Parquet schemas and hashes, rejects invalid coercion, and has isolated mutation tests. |
| Legacy scripts could overwrite current outputs | The old `audit-data.mjs` entry point is now a read-only compatibility wrapper; the v0.4 auditor retains its historical `audit-v03.mjs` filename, and the obsolete workbook entry point fails before writing. |
| Public version identity and packaged metadata drifted | v0.4 aligns README, CFF, package, datapackage, Parquet manifest, workbook and ZIP; package members and release artifacts receive explicit SHA-256 manifests. |
| Dataset-card onboarding and link health were weak | Added clickable entry points, a 26-table map, English-language metadata, a Spanish guide, pinned Datasets/pandas/Polars recipes and corrected known dead links. Data Studio exposes one orientation view; all tables remain directly loadable. |

## What remains unresolved

- The claim-coded corpus is purposive, mixed-unit and concentrated in four countries. Six sampling frames still lack a recorded systematic negative search. No prevalence inference is supported.
- One primary coder produced the coding. A second model-assisted adversarial review can detect inconsistencies but is not an intercoder-reliability study.
- Source pinpointing improved but remains bounded: 57/124 relations have a verified full-claim locator, 47/124 a partial/premise locator and 20/124 no verified locator because the cited object was inaccessible. Six unsupported edges were removed or recoded as premises, and one mutable platform count was removed for lack of a snapshot. No general source is archived or hashed, and no official-source object is hashed/versioned. Passing QA does not repair those preservation gaps.
- Most behavioural and electoral effects remain unmeasured. Content, reach, institutional response and election results do not identify voter-level agency change.
- Screening leads and empirical model studies are comparison and discovery layers, not included incidents.

## Release acceptance criteria

A publishable v0.4.0 build must satisfy all of the following:

1. deterministic source build and stored audit parity;
2. zero blocking integrity or semantic errors, with all warnings retained;
3. 26 CSV and 26 Parquet resources with matching rows, schema, source hashes and internal Parquet metadata;
4. mutation tests that reject a forged manifest, stale audits and invalid boolean coercion without replacing valid Parquet output;
5. a visually reviewed workbook whose QA sheet reports the same audit result;
6. a safe ZIP whose member hashes match both its internal manifest and the checked-out files;
7. external SHA-256 checksums for the ZIP and workbook;
8. a fixed Hub revision, public downloads for all 26 tables and an anonymous Data Studio check for the `research_view` orientation configuration.

This report records model-assisted red-team work. It is not a peer review, source-by-source adjudication, legal opinion or substitute for independent replication.