What is a claim?
A claim is either evidence or a finding.- Evidence records what a source says or what code produced, such as an accuracy of 0.94 read from a notebook.
- A finding interprets evidence, such as “The regularized model generalizes better than the baseline.” Each finding cites the claims it rests on.
C-1722010625f1, a list of sources, and a recall priority. You cite a claim by ID in writeups and chat.
Kinds of evidence
Runtime-captured values
A runtime-captured claim records a value read from running code: a notebook kernel, or a script an agent ran. Agents can ask Qualia to capture a value, but they can never type the value in themselves. Qualia always reads it from the live runtime, so the number is what the code actually computed. Each value links to the cell or script line that produced it. To capture values yourself, or to pass them between tasks, use thequalia package described in Task variables.
File-captured quotes
A file-captured claim quotes a workspace file or PDF, with its path and line range or page. Qualia accepts the quote only if the text appears in the file, so a quote can’t cite text the file doesn’t contain. If the file later changes or disappears, the claim’s source flags it. The original quote stays as the record of what Qualia captured. Agents citing that evidence are told about the change. Inspect the file and ask an agent to recapture the evidence if needed. You can create file-captured claims yourself with Save as evidence.A quote proves where text came from, not where a number came from. Ask agents to capture computed values from the runtime instead of writing them to a file and quoting it.
Figures and tables
A figure claim captures a chart from a notebook cell or a script’s saved image. Supported charts keep their underlying values, so you can read points and toggle series. To keep chart data when saving from a script, use save_figure. A table claim holds records that code wrote, such as one row per model with its scores. Agents use shared tables to compare ideas, and findings cite the table directly.Live figures
A live figure updates as new observations arrive, for example while a model trains. Ask the agent for a live figure to follow training, solver, or simulation output. To log metrics yourself from a Python notebook cell, write rows to a named table. An agent can then plot the table as a live figure.asyncio.run(...), as shown in Task variables.
To plot existing Weights & Biases logs, ask the agent to follow a local run directory or its parent wandb/ folder and choose the metrics. Qualia reads local log files only. Keep the workspace runtime running to receive updates.
Findings
Agents write findings, also called synthesized claims, to state what the evidence means, so each task leaves a record of what it established. Related claims cover the same subject without one supporting the other. They appear in a claim’s Related section and never count as evidence.Open Knowledge
Click Knowledge (the lightbulb) in the top right of the window, or choose View > Knowledge. A blue dot on the button means there is knowledge you haven’t seen.- In the Chat-first layout, Knowledge opens in the side panel beside your chat, with a Files tab next to it.
- In the Editor-first layout, Knowledge opens as an editor tab.
Read the summary and cards
A short summary at the top states the approach and results so far, with citations you can click. Cards below it show findings, figures, tables, and live figures. Evidence appears inside each card as its sources rather than as separate cards. Qualia lists the most useful cards first, and cards you pin stay at the top. To search, click the search icon and type part of a claim’s text or its ID, such asC-1722.
Inspect a claim
Click a card to open it with its full content and sources. If the card is a single piece of evidence, the click opens its source directly instead.- Notebook cells: Opens the cell that produced the evidence.
- Files: Opens the file at the quoted lines, or a PDF at the quoted page.
- Scripts: Opens the script at the line that produced the value.
- Other claims: Steps into that claim. Use the breadcrumb trail, starting at All knowledge, to step back.
Act on a card
Open a card’s … menu for these actions. They require edit access to the workspace.- Pin to top keeps the card first in your list on this device.
- Reply opens the chat of the agent that wrote the card, with the card attached, ready for your message. If that chat is gone, Qualia uses your open chat or starts a new agent.
- Fork reply forks that agent’s chat and attaches the card, so you can ask about it in a separate copy of the conversation.
Remove an item from knowledge
Choose Remove from knowledge to hide a card. Click Undo in the confirmation to bring it back. Removing a card doesn’t delete its source file or notebook.Save a selection as evidence
You can file evidence yourself from any text you can quote.- Select text in a file, or in a notebook cell’s text output. A Save as evidence button appears at the end of your selection. You can also choose Save Selection as Evidence from the menus or the command palette.
- Add an optional Caption that says what the text shows. It becomes the claim’s headline, with the quote beneath it.
- Choose a Recall priority, then save.
Import existing work
To turn earlier work into claims, such as old notebooks, results files, and papers, choose Import knowledge on the new chat screen. Add files or folders from your workspace, files from your computer, links, or typed notes. When you start the import, one agent reads everything and records what it establishes. Qualia saves imported sources in the workspace’ssources/ folder so claims can link back to them.
Use files with readable text. You can’t import scanned PDFs, images, or archives this way. For pages behind a login or paywall, download a readable copy and upload it yourself. See Upload files.
Recall priority
Every claim has a recall priority: how readily future agents should see it again. It describes the claim’s standing, not how confident anyone is that it’s true. It applies to all claims, so a captured value can outrank a finding.
Most knowledge is Low, and that’s intended. It means “recall this when it matters”, not “this doesn’t matter”. Reserve High for facts every agent should know, so that a finding one agent files reaches the others at their next step.

