Data infrastructure for policy and political AI.
TACITUS

Recorded contextual analysis · v3

A real run.
Its limits, included.

Twenty UK and US institutional publications on prejudice, protection and freedom of expression, alongside a separate 5,000-article news collection. Exact source passages, explicit frameworks and unreviewed interpretations remain inspectable.

Contextual v3 · partial · Interpretation accuracy unmeasured

Frozen example captured 2026-11-17T02:39:31.321558+00:00. Run 074539f031e64dc28f03384481563a85. These are recorded artifacts, not live telemetry.

20
real selected source documents
11,701
retained criterion assessments
99.5%
contextual units fully assessed
18/20
documents fully assessed

613 of 616 units; 11,701 of 11,704 requested criterion assessments retained; 3 unresolved assessments. Completion is a processing measure, not an interpretation-accuracy score.

Inspect the frozen run manifest ↗

A deeper way into the evidence

Read the method.
Work through real cases.

Three illustrated publications explain the research, then expose the underlying records. The commentary is AI-authored guidance; it is not independent human adjudication. Original v0.3 results remain explicitly versioned.

v0.4 publication · 18 actual PDF pages

Reading public discourse with evidence

Start with the research question, then follow source wording, represented speakers, framework conditions and uncertainty. Diagrams explain the method and its limits.

v0.4 publication · 38 actual PDF pages

Fifteen real cases: an evidence atlas

Fifteen real cases connect passages and source credits to the proposed reading. Original model assessments, later AI-authored commentary and unresolved questions stay separate.

v0.4 publication · 36 actual PDF pages

Technical annex: evidence, contracts and research

Inspect version lineage, evidence contracts, source and method provenance, processing flows, references and reproducibility. Software fidelity is distinguished from interpretation accuracy.

The independent reference-review scaffold is available, but qualified human annotation and adjudication remain pending. There is no measured interpretation-accuracy claim.

Recorded software replay: a CI run exported all 5,000 sources and 204,342 recorded assessments without fresh inference. Source/assessment identities and archive checks passed. Inspect the measured software-fidelity receipt ↗. The original run’s partial coverage and unresolved decisions remain visible; this is not semantic validation.

Open the commented examples

See how a reading is captured.

Follow the passage, speaker, criterion and proposed interpretation. The guides distinguish original model output from added AI-authored commentary, and state what a stronger political-bias claim would need.

political framing · Recorded run

How DOXA captures political framing

4 illustrative cases from a run selecting 4 sources. partial at capture; 59 of 60 contextual units fully assessed (98.3%).

358 of 360 requested criterion assessments retained; 3 of 4 source documents fully assessed; 2 criterion assessments unresolved.

Over the last 4 years, the previous administration trampled free speech rights by censoring Americans' speech on online platforms, often by exerting substantial coercive pressure on third parties, such as social media companies, to moderate, deplatform, or otherwise suppress speech that the Federal Government did not approve.

Executive Order 14149: Restoring Freedom of Speech and Ending Federal Censorship · descriptive / meets criterion

Recorded model reading: The text uses terms like 'trampled' to describe previous administration actions and 'intolerable' to characterize government censorship of speech.

AI-authored walkthrough note: The verb ‘trampled’ makes an evaluative judgment about a previous administration. The sentence also alleges a mechanism—pressure on online platforms. DOXA captures that rhetorical and political framing; it does not independently verify that the alleged events occurred or that the characterization is fair.

AI-authored walkthrough guidance; not independent human adjudication or an interpretation-accuracy benchmark. Exact anchors are checked; independent interpretation accuracy remains unmeasured.

news monitoring · Recorded run

How DOXA reads a 5,000-article collection

4 illustrative cases from a run selecting 5,000 sources. partial at capture; 10,686 of 10,759 contextual units fully assessed (99.3%).

204,342 of 204,421 requested criterion assessments retained; 4927 of 5000 source documents fully assessed; 79 criterion assessments unresolved.

Curse be upon the Jews!

Information warfare: COVID-19’s other battleground in the Middle East · indicator / meets criterion

Recorded model reading: The text reports a specific religious curse directed at Jews as part of a quoted sermon by Taha Al-Mutawakkil.

AI-authored walkthrough note: The indicator records a hostile curse directed at Jews within a quotation attributed to Taha Al-Mutawakkil. The containing article reports the sermon in a discussion of pandemic information warfare; the occurrence of the words is not evidence that the journalist endorses them.

AI-authored walkthrough guidance; not independent human adjudication or an interpretation-accuracy benchmark. Exact anchors are checked; independent interpretation accuracy remains unmeasured.

institutional context · Recorded run

Institutional voice, reported prejudice and uncertainty

4 illustrative cases from a run selecting 20 sources. partial at capture; 613 of 616 contextual units fully assessed (99.5%).

11,701 of 11,704 requested criterion assessments retained; 18 of 20 source documents fully assessed; 3 criterion assessments unresolved.

Moreover, because notions about customer preference (real or perceived) do not establish undue hardship, the client must make an exception to its dress code to let Jon wear his religious garb during front desk duty as a religious accommodation.

Religious Garb and Grooming in the Workplace: Rights and Responsibilities · descriptive / meets criterion

Recorded model reading: The text frames the conflict as a legal obligation under Title VII, where customer preference or brand image does not constitute undue hardship, necessitating religious accommodation.

AI-authored walkthrough note: The recorded frame prioritizes religious accommodation over a client's preferred image or dress code. This is a descriptive observation about protective guidance, not an indicator that the EEOC endorses the discriminatory preference.

AI-authored walkthrough guidance; not independent human adjudication or an interpretation-accuracy benchmark. Exact anchors are checked; independent interpretation accuracy remains unmeasured.

One actual v3 record

A protective rule,
in its own context.

Equal Employment Opportunity Commission · descriptive / meets criterion

Moreover, because notions about customer preference (real or perceived) do not establish undue hardship, the client must make an exception to its dress code to let Jon wear his religious garb during front desk duty as a religious accommodation.
Locator
paragraph 50 · us07-b00050
Recorded reading
The text frames the conflict as a legal obligation under Title VII, where customer preference or brand image does not constitute undue hardship, necessitating religious accommodation.
Review
unreviewed model assessment. It describes a protective legal frame; it is not evidence that the institution endorses the discriminatory preference. The commented PDF discusses the example character and attribution limits.

Keep the definitions distinct

Criteria answer different questions.

Criterion kindRecorded resultAssessments
boundarydoes not meet725
boundarymeets criterion3
boundarynot applicable504
descriptivedoes not meet279
descriptivemeets criterion1,680
descriptivenot applicable491
descriptiveuncertain12
indicatordoes not meet5,325
indicatormeets criterion91
indicatornot applicable2,584
indicatoruncertain7

Counts include the retained assessment selection, not only positive observations. Meeting a descriptive or boundary criterion is not a prejudice signal. Indicators still require attribution and context; missing or failed assessments cannot establish absence. These counts are not a universal prejudice index.

The full recorded package passed exact source-reference validation (308,256 references). This software check locates words; it does not measure semantic correctness.

Model: @cf/google/gemma-4-26b-a4b-it. Provider invoice cost is unavailable. The application conservatively accounted $0.9186 for this run. Platform charges and vendor billing remain separate.

A separate monitoring collection

5,000 full news articles.
Sources first, coverage explicit.

A real, licensed collection

Global Voices · 5,251,119 whitespace-counted words · 778 credited author names · 585 translated articles. English publication dates span 2020-01-01 to 2025-12-31 across 72 months.

Source integrity checked

5,000 of 5,000 article bodies passed the source-character audit. Attribution and CC BY 3.0 notices remain attached. 1,000 story groups are held out; there are no invented human quality labels.

Saved inference is partial

The canonical-run capture from 2026-11-17T04:44:20.361883+00:00 retains 204,342 assessments. 10,686 of 10,759 contextual units are fully assessed (99.3%); captured status: partial. 4,927 of 5,000 sources are fully assessed; 79 criterion assessments remain unresolved. Run b2a2edd4ca7047128a5e433d8a874530. This is not a completed 5,000-article model analysis.

Collection frozen 2026-11-16T22:55:24.985663+00:00. No keyword, predicted issue, or expected label influences selection. This is one publisher's selected news corpus, not a representative sample of all media. Retrieved text is a frozen later snapshot grouped by original publication date, not a reconstruction of first editions. This page reports the dated settled capture. Its partial status remains visible; later assessments, if any, are separate versions.

Application budget accounted: $14.1059. Provider invoice cost is unknown. Platform charges remain separate.

Full recorded data

Open the 5,000-article dashboard.

All 5,000 selected sources and 204,342 retained assessments remain in the full workbook and compressed snapshot, including nonmatches and not-applicable results. Indicator, boundary and descriptive outcomes stay separate.

The PDF and HTML display a 25-entry evidence appendix and 25-entry source register with full-run totals. These presentation limits do not reduce the underlying selection. Coverage is a processing measure, not a bias rate or an interpretation-accuracy score.

Checksums, scope and typed outcomes ↗ · Actual export measurement and fidelity receipt ↗. Larger files download from GitHub; source credits and translation credits remain attached.

Inspect the method and the record.

Historical v1.2 reference

The earlier legacy method retained 1,071 observations with 64.3% source-block coverage. It used a different method and denominator. Those records remain explicitly archived ↗ and are not relabeled as v3 validation.

A useful first engagement

Bring the corpus.
Define the question.

TACITUS can help scope a collection, develop a rubric, configure your deployment and organize expert review. Agree the evidence, evaluation and deliverables before work begins.

Discuss a DOXA project ↗

Open-source software. Optional consulting by agreement. Run locally or bring your own cloud; there is no hosted account or checkout on this website.

Source inspection