Models, context, and incorrect answers
Identify how missing information can cause an incorrect answer, even from a capable model.
Published by TaigaHow we write
What you will learn
- Separate model capability from access to current facts.
- Recognize when a missing constraint changes an otherwise plausible answer.
- Ask for evidence you can inspect.
Separate capability from available information
A language model uses learned patterns and the information supplied during a task. Modern models can do substantial reasoning and useful software work. They can also produce a detailed answer that depends on an incorrect assumption.
“Frontier model” describes a changing level of capability. The term does not establish that a model has read your repository. It does not show that the model knows your dependency versions or unwritten business rules. The task environment must supply these facts.
An agent with suitable tools can retrieve information. A chat without access cannot inspect the repository. When an answer appears incorrect, ask two questions. Can the model solve this problem with the correct information? Did the model receive that information?
A different model might help with the first problem. A missing policy or a dependency check can resolve the second problem.
Define the context for this task
Context is the information available for the current response. It includes instructions, supplied files, relevant conversation, and tool results. Products select and retain this information in different ways. They can also summarize earlier content.
Do not assume that a model reads every file in an uploaded folder. Do not assume that an early instruction remains available throughout a long session. Ask the tool to identify the files and instructions it used.
More context does not always improve an answer. A current architecture decision can help more than unrelated source files. An obsolete migration guide can cause an incorrect answer because the guide appears authoritative.
Consider a fictional account settings feature. Supply the route, authorization middleware, relevant data model, and an existing test. Add a specific constraint: “A member can change their display name. A member cannot change their organization role.” The model now has an explicit rule to preserve.
Check the claims in an explanation
An answer might state that an endpoint is safe because middleware checks ownership. Verify each part of this claim.
- Check that the endpoint uses the specified middleware.
- Check that the middleware verifies ownership, not only authentication.
- Identify the source of the user identity.
- Do a negative test as a different user.
A repository reference identifies where to look. The reference does not establish that the explanation agrees with the code.
Use the same method for an API recommendation. Generated code can call a method that your installed package does not export. Check the package version and official documentation before you replace dependencies. An unsupported assumption can otherwise cause an unnecessary migration.
Convert uncertainty into a check
“Be accurate” is not a verification plan. Identify the assumption, required evidence, and consequence of an incorrect result.
For example: “We have not verified tenant isolation for this endpoint. Inspect the request handler. Add a test where a user from another tenant requests the same record.” This instruction gives the agent a specific investigation and an observable result.
For implementation questions, examine the actual system version. A document can describe intended behavior. Code inspection and tests help establish current behavior. If they disagree, record the difference until an owner resolves it. Do not silently select the more convenient answer.
Managers can use this method without reading each code change. Ask which assumptions the team checked. Identify the assumptions that remain open and their owners. This information supports a release decision more directly than the model name.
Do the exercise
Select a small function that you understand. Use code without sensitive information. 1. Ask an approved AI tool to explain the function. 2. Supply the caller and one failed test. 3. Ask the tool to revise its explanation. 4. Record the changed claim and the evidence that changed it. 5. Record any uncertainty that remains.
Download worksheet (Markdown)Check your understanding
Sources & further reading
Related reading from Taiga
Clearing this selection deletes all progress saved in this browser.
Progress stays in this browser. No account, no tracking.