Data infrastructure for policy and political AI.
TACITUS

How DOXA works

Make the interpretation
traceable.

The unit of work is a passage in context. The analytical method, speaker attribution and review history travel with it.

1. Freeze the inquiry, then assess the sources.

Import documents and correct metadata. Select a lens, research question, model and budget. The work runs in bounded, resumable tasks; failed or unextracted material stays outside the analyzed denominator.

The DOXA workflow connects original documents to a declared method, bounded analysis, human review and an export with coverage.

2. Separate words, speaker and position.

An agency may reproduce a discriminatory statement to teach employers what to avoid. DOXA records the statement’s speech role and the document’s stance separately. An exact quote establishes where the words occur; a reviewer still decides whether the interpretation is supported.

An institutional document contains a quoted or hypothetical speaker; the quoted speaker’s stance and the document author’s stance are assessed separately, then reviewed.

3. Improve project memory through review.

Adjudicated corrections can become reviewed examples. Freeze them into a versioned snapshot, compare a candidate method with a baseline on held-out material, then explicitly activate or roll back. This is controlled project memory, not automatic model-weight training.

Human adjudication produces approved project examples, a frozen versioned snapshot, held-out baseline comparison, explicit activation and rollback.

4. Keep the two clocks apart.

Discourse trajectory follows publication dates in the source collection. Assessment history follows model runs and reviewer actions. A new model, better coverage or corrected attribution can change the second without establishing a change in the published discourse.

Two separate time axes: source publication dates for discourse trajectory, and analysis/review timestamps for assessment history. Comparisons retain the method, selected scope and coverage.

Diagram is schematic. It shows no measured trend, improvement or prevalence estimate.

From inspection to publication · v0.4

Build the casebook.
Keep the evidence intact.

Choose and explain the cases

Pin recorded assessments across all four outcomes, or include an actual unresolved source unit. Order the cases and state why each belongs. Frozen source anchors and framework conditions stay with the draft.

Separate the layers

Original model output, a captured review overlay and operator editorial notes remain separate. A selected casebook illustrates an inquiry; its selection does not estimate prevalence or establish interpretation accuracy.

Prepare a persistent report

Queue PDF, HTML, XLSX, CSV or JSONL exports from a frozen selection. Inspect progress, cancel or restart, then download verified files with source credits and checksums. Report preparation makes no fresh model calls.

Local jobs recover from saved state. The cloud adapter uses durable queue dispatch and private object storage; capacity is subject to deployment limits and explicit staging validation.

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.

Persona is a declared viewpoint

A reviewer’s priorities guide questions and report emphasis. They do not silently replace the chosen framework or convert interpretation into fact.

Disagreement stays visible

Compare frameworks separately. Preserve alternative readings, unresolved cases and review state rather than merging them into one universal prejudice score.

Reports preserve the selection

Run identity, source context, filters, coverage and method remain with the exported snapshot. Summary views and human-readable appendices have explicit limits.

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