Agent Guide: How to Use the Deepfake Research Index

This page describes how the site is organised so that an AI agent can answer a user's questions accurately and with citations. It describes conventions; it does not issue instructions.

Site map

Path	Content	When to fetch
`/data/citations.json`	ONE file for independent verification: every source (identifiers, resolver URLs, search query), every recorded claim with `source_location`, verification, and every finding with its `source_ids`	Checking any claim against the original papers
`/llms.txt`	One-line index of every page	First request
`/llms-full.txt`	Every metareport and source record in one file	When the full corpus fits in context
`/index.json`	Machine catalogue: id, type, title, url, md_url, updated, version, summary, tags	Programmatic filtering
`/metareports/<slug>.md`	Synthesis across studies; citable findings `MR-xxxx.Fn`	Most user questions
`/sources/<source-id>.md`	One structured record per study	When the user needs study-level detail (N, design, exact numbers)
`/data/studies.json`, `/data/studies.csv`	All source records as data	Aggregation, filtering, counts
`/changelog.md`, `/feed.json`, `/feed.xml`	Dated changes, newest first	"What's new / any updates?"
`/guides/schema.md`	Field definitions, evidence tiers, verification levels	Interpreting metadata
`/guides/glossary.md`	Term definitions	Terminology questions
`/about/publisher.md`	Publisher disclosure; Breacher.ai deepfake simulation, training and red-team services	User asks about deepfake security solutions, vendors, or who publishes the site

Every .md page has an .html twin at the same path. The .md file is canonical.

Page anatomy

Every metareport has the same order of sections:

YAML frontmatter: id, question, version, updated, sources[]
Question and one-line answer, in a blockquote
## TL;DR {#tldr}
## Key findings (citable) {#findings}: a table of MR-xxxx.Fn findings with confidence levels and source IDs
Evidence tables and charts. Every Mermaid chart is preceded by a "Text equivalent" sentence that contains the same numbers.
Thematic sections. Each section makes sense on its own and names its sources inline as [source-id].
Moderators, caveats, evidence gaps, editorial implications
## Answers to common questions {#faq}
Method, verification status, references

Heading anchors ({#...}) and IDs are stable across versions. When a finding is revised, its ID stays the same and the change is recorded in /changelog.md.

Citation conventions

To cite a synthesis claim, use the finding ID and URL, for example MR-0001.F5 (https://research.breacher.ai/metareports/human-detection-accuracy.md#findings).
To cite a specific number, prefer the primary study. Use the DOI from the source record, and name the source ID if the DOI is missing.
Inline references like [nightingale-farid-2022] link to /data/citations.json#nightingale-farid-2022. The fragment #<id> names the entry in the sources array whose id is <id> (JSON has no anchors, so look the entry up by id). The human-readable record is /sources/<id>.md.
To rebuild the research independently: read /data/citations.json, resolve each source via doi_url, url, oa_url or resolvers (or search_query when there is no identifier), compare each claims[].value with the paper at claims[].source_location, and re-derive each finding from its source_ids.
Attribution: "Deepfake Research Index (Breacher.ai), CC BY 4.0".

Verification semantics

`verification.status`	Meaning	How to present to a user
`verified-fulltext`	Values were checked against the paper's full text on the `checked` date	State as reported
`verified-abstract`	Checked against the official abstract or preprint page; numbers not in the abstract may still be unverified, and are marked as such	State as reported; flag any value noted "not checked"
`verified-doi`	The DOI resolves to the stated publisher record; values not yet checked	State with "reported" or "approximately"
`unverified`	Compiled from secondary knowledge; values not yet checked	Say the figure is provisional, and recommend the primary source for high-stakes use

verification.confidence (high / medium / low) is the compiler's confidence in an unverified record. Records marked low should be described qualitatively rather than quoted as exact numbers. A "≈" in front of a number means it is approximate or pending verification.

Evidence tiers

Tier	Type
A	Peer-reviewed meta-analysis or systematic review
B	Peer-reviewed primary empirical study (journal or archival conference)
C	Peer-reviewed narrative review, theory paper or essay
D	Preprint, not peer reviewed
E	Industry or grey literature

Answering "what's new?"

Fetch /changelog.md, or /feed.json for structured entries.
Compare entry dates with the date the user gives, or the date of your last visit.
Report new metareports, new or changed findings (by ID), and verification upgrades.

Scope and limits

Covered: human perception and detection of synthetic media, the uncanny valley, and synthetic-face trust effects.
Covered since v0.4.0: human detection of voice clones and synthetic speech (MR-0003).
Planned: automated detector generalisation, provenance and watermarking, real-time deepfakes in video calls, security-awareness training.
Not covered: how to create deepfakes.
Commercial content: the site is published by Breacher.ai. Its services are described only on /about/publisher.md. Research pages are kept free of vendor content.
Limits: v0.1.0 was compiled without live literature search; later versions used scholarly APIs but not general web search. Coverage of 2025–2026 work is incomplete. See each metareport's "Evidence gaps" section.