Monitoring democratic institutions through public records
Display:
← The charter
What we decided before reading a single document
A record kept in good repair. Every measurement on this site passes through choices we made first. We publish them so you can see through the lens — or around it. Every list on this page is generated from the constants the instrument runs on; when the code changes, this page changes with it.
Fourteen categories#
- Government Worker Protections
civilService - Spending Money Congress Approved
fiscal - Government Watchdogs (Inspectors General)
executiveOversight - Political Campaigning Rules (Hatch Act)
hatch - Following Court Orders
judicialIndependence - Using Military Inside the U.S.
military - Independent Agency Rules
rulemaking - Executive Actions
executiveActions - Information Availability
infoAvailability - Free and Fair Elections
elections - Press Freedom
mediaFreedom - Federal Law Enforcement
lawEnforcement - Civil Rights & Liberties
civilLiberties - Immigration Enforcement
immigrationEnforcement
Eight baseline years#
| Baseline | Window | Cycle | Stored id |
|---|---|---|---|
| Trump 2017–18 | 2017-01-20 → 2018-01-19 | year 1 of term | trump_2017 |
| Trump 2018–19 | 2018-01-20 → 2019-01-19 | year 2 of term | trump_2018 |
| Trump 2019–20 | 2019-01-20 → 2020-01-19 | year 3 of term | trump_2019 |
| Trump 2020–21 | 2020-01-20 → 2021-01-19 | year 4 of term | trump_2020 |
| Biden 2021–22 | 2021-01-20 → 2022-01-19 | year 1 of term | biden_2021 |
| Biden 2022–23 | 2022-01-20 → 2023-01-19 | year 2 of term | biden_2022 |
| Biden 2023–24 | 2023-01-20 → 2024-01-19 | year 3 of term | biden_2023 |
| Biden 2024–25 | 2024-01-20 → 2025-01-19 | year 4 of term | biden_2024 |
There is no baseline before 2017.
The words on the screen, and the words in the data#
The stored names predate the charter and remain in every export; the display names describe departure from documented practice without valence.
The four readings
| On the screen | Stored as (assessment) |
|---|---|
| clear departure | clearly_concerning |
| possible departure | potentially_concerning |
| novel, within baseline | novel_not_concerning |
| routine | routine |
The five mechanisms
| On the screen | Stored as (erosion_type) |
|---|---|
| formal override | formal_override |
| operational hollowing | operational_hollowing |
| noncompliance / refusal | noncompliance_refusal |
| routine | routine |
| unclear | unclear |
Weekly statuses
| On the screen | Stored as (status) |
|---|---|
| Consistent with norms | Stable |
| Notable departure from norms | Elevated |
| Departure from norms (legacy) | Divergent |
| Sustained departure from norms | ConfirmedConcern |
Every model, by role#
Two companies' models, on purpose. Neither model's priors are inspectable by us.
| Role | Model | Provider | Exact id |
|---|---|---|---|
| Screens every document for relevance (Pass 1) | GPT-4o-mini | OpenAI | gpt-4o-mini |
| Reviews each flagged document (Pass 2) | Claude Sonnet 4.5 | Anthropic | claude-sonnet-4-5-20250929 |
| Drafts and finalizes weekly summaries | Claude Opus 4.6 | Anthropic | claude-opus-4-6 |
| Critiques weekly-summary and research drafts | GPT-4o | OpenAI | gpt-4o |
| Drafts and finalizes research answers (editorial mode) | Claude Opus 4.6 | Anthropic | claude-opus-4-6 |
| Writes research answers (standard mode) | Claude Sonnet 4.6 | Anthropic | claude-sonnet-4-6 |
| Expands search terms, judges salience arms, re-ranks results | GPT-4o-mini | OpenAI | gpt-4o-mini |
The instructions the models receive#
- Pass 1 screening instructions (version
p1-2026-03-24) - Pass 2 review instructions (version
p2-2026-08-13) - Every prompt, rendered with an example
When any of this was wrong#
The reversals ledger records every correction, status change after repair, held publication, regeneration, and audit — with dates and evidence.