Context X-Ray
Inspect the context in your own request. Local analysis, with no raw prompt upload.
Inspect a requestCONTEXT & MEMORY / EXPERIMENT 001
The instruction that disappeared.
The refund policy is in the conversation. Is it in the next request?
45-second walkthrough Free browser labBy Krish Adavi
Browser simulation
The policy is there. A support assistant has a refund policy and a customer question. This short conversation fits the input budget.
SUPPORT ASSISTANT / REQUEST ASSEMBLYIllustrative units
Customer
My annual plan renewed yesterday.
Can I get a refund?
Available context200 / 240
System instruction
24Refund policy
72Example policy: annual renewals can be refunded within 7 days.
Earlier message 1
40Earlier message 2
40Latest question
24The policy fits, for now
What survives depends on how the application assembles its request.
Evidence & method: what this experiment does and does not show
This is a deterministic browser simulation, version 1.0, authored September 27, 2026. There are no recorded model responses or live model calls. The walkthrough uses fixed sample messages and illustrative context units, not provider-tokenizer counts.
The application always reserves the system instruction (24 units) and latest question (24). The refund policy uses 72 units; each earlier message uses 40. Unprotected blocks are removed oldest first until the input fits. Reserving the policy changes which blocks can be removed. A request is rejected if its protected blocks cannot fit.
This is one possible application policy, not automatic behavior of every model or provider. The budget covers input only; production systems must also allow for output and provider-specific overhead. Including evidence does not guarantee a correct answer. Validate actual behavior with representative cases and evaluations.
Lab calculations stay in your browser. The download contains the example settings and derived results. Standard site analytics may record experiment interactions, but no personal prompt or document is requested by this lab.
Take it to your own system.
DriftCheck
Turn a behavior you depend on into a regression check for your next change.
Run your first checkStack Check
Review provider changes that could affect the AI stack you are building on.
Check your stack
The policy is there.
A support assistant has a refund policy and a customer question. This short conversation fits the input budget.