AI foundations / Working with evidence
What is in the context?
A saved note can help an AI answer a question. First, the right information has to reach it.
Your objective: choose a useful excerpt from a small knowledge vault, then distinguish a supported answer from a guess.
A vault is storage. Context is what is provided now.
01 / 04Knowledge vault
Fictional notes. None has been opened for this request.
Current context
The request + relevant instructions
Other conversation content may also be included.
Prior learning + input
A knowledge vault is a collection of saved notes. The model does not automatically see those files. Here, its current context contains the question and instructions, but no workshop information. Earlier training does not establish the facts of this fictional private event.
Boundary: saving a note does not automatically add it to model training or to the next model call.
Now try it / Three small checks
Choose evidence. Check the answer.
Everything below is a scripted illustration using fictional notes. There is no AI call, network request or real file access. Feedback compares your selections with the visible lesson fixtures.
Predict before reading
A private workshop note exists in a vault. No excerpt from it has been included in the current context.
Guided attempt: supply the useful excerpt
Your request: “According to the confirmed schedule, when and where is the beginner workshop?”
No excerpts selected. The source cards remain unchanged.
Independent attempt: the plan has changed
A different event, a new room, and a revised time. According to the latest confirmed schedule, when and where is the Python study lab?
Confirmed, 2 October: The Python study lab is at 18:00 in Room Ash.
Confirmed revision, 8 October: The Python study lab has moved to 19:00 in Room Elm. This replaces the 2 October schedule.
Checked, 9 October: The projector in Room Elm is out of service. This note does not set the lab’s time or location.
0 of 3 checks completed. You can retry any check. The aim is to explain your choice, not answer quickly.
Read the complete lesson transcript and answer explanations
Objective and scope
Select a useful excerpt and distinguish a supported answer from a guess. This approximately six-minute lesson is an offline, scripted illustration. All workshop notes are fictional. It does not call an AI, open a vault, change a file or store your progress.
Step 1 · The saved notes
A knowledge vault is a collection of saved notes. Context is the information supplied for a particular model call, such as instructions, the current request, included conversation and tool results. Stored files are not automatically included. Earlier training provides learned patterns and capabilities; it does not establish the facts of our fictional private workshop. Saving a note does not automatically train the model on it.
Step 2 · Bring in evidence
A file-reading tool can return a note, or retrieval can select relevant excerpts. The surrounding application assembles what is provided to the model. In this diagram, a copy of the confirmed workshop sentence enters the context: “The beginner workshop starts at 18:00 in Room Cedar.” The original note remains in the vault. The other notes are not included. This lesson’s checkboxes only simulate that selection.
Step 3 · Use the context
The model processes the supplied context using its learned capabilities and generates output. A supported candidate answer is: “According to the confirmed note, the workshop starts at 18:00 in Room Cedar.” This example response is scripted, not generated by a model. Generating an answer does not itself modify a saved file; an actual write would need a separate permitted action.
Step 4 · Check every claim
Including a useful excerpt does not guarantee a correct answer. A candidate response might add “Free pizza is provided,” but the supplied note says nothing about food. The workshop time and room are supported; the pizza claim is unsupported. Check the source, its status and each claim. Even a faithfully quoted note can be outdated or wrong about the world.
Check 1 · Prediction
The correct answer is: “No. This call has not been given evidence for that event.” A plausible start time would be a guess in this fixture. Neither saved files nor merely naming a vault automatically supplies its contents.
Check 2 · Guided selection
- The final note, dated 10 September, says 18:00 in Room Cedar.
- The draft, dated 4 September, only considers 17:00 in Room Birch and explicitly says it is not confirmed.
- The equipment note, dated 11 September, lists keyboards and monitors, not the workshop time or room.
The smallest useful selection is the final note alone. A scripted answer supported by it is: “According to workshop-confirmed.md, the beginner workshop starts at 18:00 in Room Cedar.” Including additional notes is not inherently forbidden, but it is unnecessary for this request. Conflicting notes require interpretation rather than blind copying.
Check 3 · Independent attempt
- Note A confirms 18:00 in Room Ash on 2 October.
- Note B is a confirmed revision dated 8 October: 19:00 in Room Elm. It explicitly replaces the 2 October schedule.
- Note C reports a broken projector on 9 October and explicitly does not set the event’s time or location.
Choose Note B and the answer “19:00 in Room Elm, according to the confirmed revision of 8 October.” The most recently dated note is not necessarily the relevant source. A confident response is not proof of its truth; verify the specific answer against the relevant evidence.
One useful habit
Before trusting an answer about your notes, ask: “Which information was actually supplied, and which sentence supports this claim?”