- Source Desk is an AI agent skill that turns a research question into a research ledger of findings, sources, saved source text, and verdicts.
- An offline Python script checks that quotes appear in the saved text, groups copied sources into one origin, and flags conflicting values and single-origin findings.
- It shows disagreements instead of hiding them, and a SHA-256 seal lets you detect later changes to the ledger or snapshots.
- The script makes no network calls and the package has no model, so you supply your own agent, Python 3, and provider.
- It checks the saved package, not truth. Corroborated does not mean true, and saved text may differ from the live page.
Source Desk: Research Ledger & Quote Checker for AI-Assisted Research
AI can give you a confident research answer with a quote that never appeared on the page. Source Desk traces each claim back to saved source text, so you can see what is supported, what is copied, and what conflicts.
By Genial Things Editorial Team · AI Agent Skills · Updated October 2026 · 9 min read
Source Desk is an AI agent skill built to make research problems easier to identify. It turns a research question into a structured source ledger, then uses an offline Python script to check quotations, compare source origins and dates, flag conflicting values, and prepare a report you can inspect before using the findings in an article. You can open Source Desk on Capafy at any point.
What Is Source Desk?
Source Desk is a downloadable AI agent skill for factual research and source verification. Instead of treating a list of links as evidence, it asks your agent to save the text collected from each source and tie every finding to specific evidence. An offline Python script then checks quotations, source origins, dates, and conflicting values, and produces a report you can inspect.
The package produces:
- A research ledger containing findings, sources, evidence, and verdicts
- A readable Markdown report and a JSON report with source and finding details
- An optional export folder that ClaimLock can use to check a draft
- A hash seal that helps detect later changes to the ledger or saved source text
It is intended for adults who prepare factual answers for others: analysts, journalists, editors, consultants, researchers, content teams, SEO professionals, and founders. It is a research-checking tool, not a guarantee that a source is truthful or that a research answer is complete.
🗂️ Want an evidence trail behind every AI research answer?
Try Source Desk →Why a Research Ledger Matters
AI can produce a convincing research answer while quietly getting details wrong: a quotation may not match the original page, two articles may repeat the same wire story, or two sources may report different numbers. A research ledger ties each claim to saved source text, so weak support, copied sources, and conflicts become visible before you publish.
Imagine writing about a software company's refund policy. One official page says 30 days. Another says 14 days. A third site repeats the 30-day claim but copied it from the first. A simple list of three links could make the claim look well supported. Source Desk is designed to tell those situations apart:
- It checks whether the quotation appears in the text saved for that source.
- It groups identical pages, repeated addresses, and certain copied or syndicated content into one origin.
- It flags findings with conflicting values and findings backed by only one origin.
- It preserves both sides of a disagreement rather than silently choosing a number.
How Does Source Desk Work?
Source Desk uses a six-stage workflow: plan the research question, collect sources and save their text, build the source ledger, run offline checks, repair the ledger without hiding disagreements, then seal, recheck, and optionally export. At each stage the agent must preserve the actual source text instead of reconstructing it from memory or a summary.
1. Plan the research
Start with a factual question. Your agent restates it in one line and lists the sub-questions a good answer needs. You can optionally set the number of sources, preferred source types, and how old a source may be before it is stale. The default threshold is 365 days.
2. Collect sources and save their text
Your agent finds sources with its own search or fetch tools, or works from text you provide. For each source it saves the returned text to a snapshot file such as snapshots/S01.txt. If a tool returns only an excerpt or summary, the ledger must say so. Each source record includes its address, title, publisher, type, publication date, retrieval date, and capture type.
3. Build the source ledger
The agent creates a research.json file with the question, sources, findings, and evidence. Every finding needs a short statement and evidence linked to a source. Quotes must be copied word for word from the saved snapshot, at least five words long, with an ellipsis for skipped words. Values that sources might disagree on are recorded, and disagreeing sources must be included, not dropped.
4. Run the offline checks
The bundled script checks the ledger against saved text and metadata. The date below is an example, so use your actual run date.
python3 -I scripts/sourcedesk.py check research.json \
--out ledger.json \
--md ledger.md \
--as-of 2026-10-22
An exit code of 2 means at least one high-severity issue. It does not mean the research is false or a source is malicious.
5. Repair the ledger without hiding disagreements
Up to three repair rounds are supported. The agent may fix a miscopied quote, add independent sources for single-origin findings, or look for a primary or official source. The rules: never edit a snapshot just to make a quote match, and never delete a conflicting source or change a value to remove a conflict. When sources disagree, the report shows both sides unless you give an explicit rule, such as preferring the newest source.
6. Seal, recheck, and optionally export
Source Desk creates a SHA-256 seal (see NIST FIPS 180-4). Keep ledger.json and the snapshots/ folder together, and recheck later to see whether anything changed. The optional export builds a folder of source metadata, verified quotes, and citations for a ClaimLock draft check:
python3 -I scripts/sourcedesk.py export ledger.json \
--base . \
--out-dir for_claimlock
Sources with blocked, summary-only, or empty text are excluded from the export.
Turn a research question into a checked ledger with verified quotes, source origins, and conflicts shown side by side.
Launch Source Desk →What Does Source Desk Check?
The bundled catalog has 14 checks across eight areas: missing evidence, quote matching, source independence, numerical consistency, conflicting values, source quality, dates and freshness, and file integrity. They are designed to make review more systematic and to point you to the problems that need attention, but they do not replace reading the evidence yourself or consulting primary sources.
| Check area | What it looks for |
|---|---|
| Missing evidence | Findings without evidence or references to missing sources |
| Quote matching | Quotes that do not appear in the saved source text |
| Source independence | Repeated URLs, identical text, and certain copied or syndicated pages |
| Numerical consistency | Numbers in statements that are not supported by matched quotes |
| Conflicting values | Sources that give different values for the same finding |
| Source quality | Single-origin findings and findings backed only by certain source types |
| Dates and freshness | Invalid dates, old sources, and claims worded as current but backed by old sources |
| File integrity | Missing, unreadable, empty, or path-escaping snapshots |
How Do You Interpret Source Desk Verdicts?
Each finding gets one of five verdicts: unsupported (no evidence), unverified (quote not found or only a summary), conflict (matched quotes give different values), single-origin (one source origin), or corroborated (quotes in two or more independent origins with no conflicting values). Corroborated does not mean true. It means the saved evidence passed the package's defined checks.
- Unsupported: no evidence is provided.
- Unverified: the quote was not found, or the source text is only a summary.
- Conflict: matched quotes provide different declared values.
- Single-origin: evidence comes from only one source origin.
- Corroborated: quotes found in two or more independent origins, with no conflicting values reported.
Editorial standards bodies stress the same discipline of verifying claims against sources. See the IFCN Code of Principles and the SPJ Code of Ethics.
What Makes Source Desk Different?
Many research workflows focus on gathering more links. Source Desk focuses on the relationship between the claim, the quote, and the saved source text behind it. It preserves the evidence trail, treats copied sources as one origin, shows disagreements instead of hiding them, and produces a sealed package that a team can inspect and recheck later.
- It preserves the evidence trail: saved snapshots show the text actually collected, which beats an agent's paraphrase of a page.
- It treats copied sources cautiously: source addresses, identical text, and substantial shared text are compared, and likely copies count as one origin.
- It shows disagreements: both pieces of evidence are kept, which suits prices, dates, policies, specifications, and statistics that change over time.
- It supports repeatable review: the JSON ledger, Markdown report, snapshots, and hash seal can be rechecked later.
Make Your Research Easier to Inspect and Question
Check quotes against saved text, compare source origins, and keep every disagreement visible.
Try Source Desk on Capafy → More AI Agent SkillsChecks the saved package, not the truth of the source
Does Source Desk Need an API Key?
No API key is bundled or required by the script. Source Desk is a download package with no AI model, and it does not run through Capafy's cloud. Its Python script runs locally, uses only the standard library, and makes no network calls. You do need your own agent, Python 3, and a model provider account, and you pay any model fees to those providers.
- An AI agent that can read the skill instructions, work with files, and run shell commands
- Python 3 (developed with 3.13; other 3.x versions are untested)
- Your own model or provider account
- Search or fetch tools if you want the agent to find sources itself, or paste source text instead
The instructions are written for a Claude-style agent, and other agents are not tested.
Data handling and privacy
The script does not send your question, snapshots, or ledger to Capafy. It reads your files and writes reports locally. However, your agent, model provider, and search or fetch tools may process your question and source pages under their own terms. Never put passwords, API keys, or other secrets into the question or snapshots.
Important Limitations of Source Desk
Source Desk checks what is in the supplied research package. It does not prove the research is correct, establish that a source is truthful, or show that saved text matches the live page. It may miss conflicts worded differently and can produce false positives. It is not legal, medical, financial, or other professional advice.
- It does not establish whether a source is telling the truth.
- It does not prove the saved text matches the live page.
- It cannot guarantee a paraphrase preserves a quote's meaning.
- It cannot establish complete independence between all sources.
- It does not guarantee the search was comprehensive.
- It may produce false positives, such as treating shared boilerplate or a long quotation as copying.
- Runs are limited to 60 sources, 200 findings, 10 evidence entries per finding, and 1 MB per snapshot. Some contradiction and near-duplicate checks are skipped, with a report, above 1,500 entries.
For consequential research, use qualified review and consult primary sources directly. To find topics worth researching first, try OpenSignal, and to audit forecast claims see ScoreTrace AI.
Who Should Use Source Desk?
Source Desk suits people who publish or hand off factual research: content and SEO teams, journalists and editors, consultants and analysts, founders and researchers comparing prices, policies, or market claims, and AI-assisted workflows that need a repeatable source inspection step. If you only need a quick answer to a simple question, a full ledger workflow is probably unnecessary.
- Content and SEO teams needing an evidence trail before publishing. Pair it with our Keyword Density Checker.
- Journalists and editors checking quotations and preserving conflicting evidence.
- Consultants and analysts organizing findings from many sources.
- Founders and researchers comparing product details, prices, policies, or market claims.
- AI-assisted workflows that must inspect sources before handing findings to another tool, such as ContextGuard for the agent's own instruction files.
Source Desk — Frequently Asked Questions
These answers cover what Source Desk is, whether it needs an API key, how it checks quotations, what the verdicts mean, and where its limits are. In short, it makes AI-assisted research easier to inspect, repeat, and question, but it never proves that a source is true or that the saved text matches the live page.
Final Verdict
Source Desk gives people who use AI agents a practical research ledger. It preserves source text, checks quotations, identifies likely copied sources, flags conflicting values, and produces a structured report, with an optional ClaimLock export. Its key benefit is not guaranteeing a claim is true, but making the research process easier to inspect, repeat, and question.
Source Desk checks that quotes appear in the text your agent saved, and compares dates, values and origins. It does not show that any source is true, or that the saved text matches the live page.
Explore more in our AI Agent Skills category and the Free AI Online Tools collection.
Next step: Try Source Desk on Capafy and turn your next research question into a ledger you can defend.