Source Desk

Keep and check the evidence behind a research answer.

A research answer is only as good as its sources. Source Desk makes your AI agent keep a ledger: every source it used, the exact quote behind every finding, and an offline script that checks the ledger instead of trusting it.

What the script checks

• Quotes: each quote must appear word for word in the text your agent saved (case, spacing, quote marks and dashes ignored), with the line number
• Copies: syndicated or copied articles, the same address with tracking tags, and the same publisher count as one origin, so two outlets repeating one report don't look like two sources
• Conflicts: when sources give different numbers or values, both sides are shown with dates; Source Desk does not pick a winner
• Single-origin findings: flagged, with lower severity when the source is primary or official
• Old and undated sources: including "currently" statements backed only by old pages
• Weak support: findings backed only by forums, blogs or videos
• Bad captures: blocked pages, login walls, stubs, and summaries saved instead of page text
• Numbers: a number in your statement that is not in its quote

What you get

1. A ledger: each finding with a verdict (corroborated, single-origin, conflict, unverified, unsupported), each source with date, address and SHA-256
2. A readable report with an issues table, conflicts side by side and a not-measured list
3. A seal hash and a recheck command, so later edits to a snapshot or to the ledger show up
4. An optional export for ClaimLock: a sources folder, verified quotes and numbered citations

Included: a clean sample and a deliberately flawed sample (copied article, conflicting numbers, a blocked page, a summary, an invented quote) with their expected reports.

How to use it

1. Unzip the package into your agent's skills or working folder
2. Give your agent a research question (and, optionally, how many sources you want and how old is too old)
3. The agent saves each source's text, writes the ledger and runs the checker
4. Review the findings table, the conflicts and the issues; repair and re-check

Where it runs and what you need

Source Desk is a download package. You unzip it and use it in your own environment. It does not run through Capafy's cloud, and no AI model is included in the price.

You provide:

• An AI agent or assistant that can read a SKILL.md file, read and write files in a working folder, and run shell commands. The instructions are written for a Claude-style agent; other agents have not been tested.
• Search or fetch tools, if you want your agent to find sources on the web itself. Without them, paste the source text.
• Python 3 (standard library only, nothing to install; developed on Python 3.13, other 3.x versions not tested).
• Your own model or provider account. Any model and tool usage fees are charged by those providers, not by this package.

Data sharing

Nothing is sent to Capafy or to the author by this package. The script runs on your computer, makes no network calls and imports no network libraries. Searching and fetching pages is done by your agent's own tools, which may send your queries to third parties under their terms. Your question and the pages your agent reads go to whichever providers you use. Don't put passwords or keys in a question or a saved source.

Honest limits

• It checks that quotes appear in the text your agent saved. It does not show that a source is true.
• It cannot confirm that a saved snapshot matches the live page; open the address for anything that matters.
• It does not check that a paraphrase keeps the meaning of its quote, or that the search was complete.
• Independence is judged by domain, publisher and copied text only; shared owners or shared underlying data are not detected.
• Numbers written as 5k, 1.2M or in words are not compared.
• Not legal, medical, financial or professional advice. For adults (18+).