019 min
Read a partial log
Ask for the camp opening time from a covered strip.
Ask: “Can you answer from the words you can actually see?”
Listen for: “No; the time sentence is covered.”
Brightlab
brightlab.Made for curious mindsbrightlab.Made for curious mindsBrightlab
Year 6 · Context
What happens when the fact you need falls out of view?
Go to the investigation ↓
Investigate with Pip the possumA context-budget briefing with a highlighted evidence sentence and a conflict note.
What must survive when a briefing becomes shorter?
A system can only use information available to its current computation. This context simulator keeps a limited suffix of a fictional expedition log. A deterministic answer rule extracts a fact only if its complete sentence is visible. Increasing the window can restore a fact, but it does not resolve contradictions or ensure a real language model will use it correctly. The lab performs text slicing and rule-based extraction, not LLM inference.
Before this lesson: Identify a fact and a condition in a short instruction. Useful earlier investigations: A sentence at the crossroads.
A system can only use information available to its current computation. This context simulator keeps a limited suffix of a fictional expedition log. A deterministic answer rule extracts a fact only if its complete sentence is visible. Increasing the window can restore a fact, but it does not resolve contradictions or ensure a real language model will use it correctly. The lab performs text slicing and rule-based extraction, not LLM inference.
Identify a fact and a condition in a short instruction. Useful earlier investigations: A sentence at the crossroads.
Print the fictional log on a long strip. Cover its beginning with paper to leave only the last words visible. Mark that the lab counts whitespace-separated words, unlike production tokenisers.
ACARA V9 AI curriculum connection ↗ · Technologies ↗
A complete teaching sequence · 70 minutes
Ask for the camp opening time from a covered strip.
Ask: “Can you answer from the words you can actually see?”
Listen for: “No; the time sentence is covered.”
Predict the minimum window that contains the complete fact sentence.
Ask: “Is half the sentence enough to support the answer?”
Listen for: “It may leave out the important condition.”
Change the word budget and inspect included/excluded text and extraction result. Keep the log and question fixed.
Ask: “Which exact words returned when the answer became available?”
Listen for: “The sentence containing the opening time.”
Use the contradiction edge case with two visible times and dates.
Ask: “Did a larger window solve the disagreement by itself?”
Listen for: “No; we need to interpret which update applies.”
Write a concise, dated summary from supplied facts and place essential information near the query. Preserve conditions and mark unresolved conflicts.
Ask: “What could a too-short summary accidentally remove?”
Listen for: “The date or exception that makes the fact accurate.”
Collect window size, visible evidence and a reasoned unknown response.
Ask: “Does visible information guarantee correct use by an LLM?”
Listen for: “No, availability is only one requirement.”
A generator remembers everything.
A larger window can include conflicting updates, so more context does not automatically settle which fact applies.
Create a concise dated briefing preserving the required fact, qualifier and any unresolved contradiction.
Ask learners to quote the actually visible supporting sentence rather than a remembered hidden sentence. Check that summaries retain dates and conditions.
| Criterion | Beginning | Secure | Extending |
|---|---|---|---|
| Context reasoning | Assumes all past text is available | Identifies included and excluded evidence | Explains complete-sentence and contradiction limits |
| Briefing design | Copies an unsupported answer | Preserves the fact within budget | Keeps qualifiers and explicitly handles conflict |
Use a six-sentence paper log and count words together. Allow highlighting rather than lengthy written explanation.
Compare suffix truncation with a relevance-based selection strategy and test which question each loses.
Use only the supplied fictional log. Never paste private school correspondence or student records into the journal.
Shared device? Turn remembering off. A project file lets you continue on another device.
Your browser is the laboratory
Reduce the context budget with text and question fixed. Predict when the complete supporting sentence disappears.
The log begins “At camp the gate opens at 8 am on dry days.” Later sentences describe a walk, battery, map and clouds. The initial suffix window holds only 25 words. Only complete visible fact phrases can be extracted.
The experiment opens after your prediction.
There is no penalty for being surprised.
Apple Silicon · PyTorch MPS
Evaluate complete-fact availability at thousands of synthetic positions and twenty context budgets. Change fact length to reveal boundary failures; availability alone does not establish correct model use.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download Mac Studio investigation ↓unzip y6-context-mac-pathway.zip -d y6-context-mac cd y6-context-mac bash setup-mac.sh source .venv/bin/activate python experiment.py --device mps --output results marimo edit notebook.py
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
NVIDIA DGX Spark · PyTorch CUDA
Evaluate complete-fact availability at thousands of synthetic positions and twenty context budgets. Change fact length to reveal boundary failures; availability alone does not establish correct model use.
Students extend the experiment in teams, documenting parameters, outputs and limitations.
The bundle contains lesson-specific working code, a configuration file, a reactive notebook, a deterministic CPU check and hardware setup instructions. Acceleration is reported only after a tensor operation and result read-back succeed.
Download DGX Spark investigation ↓unzip y6-context-dgx-pathway.zip -d y6-context-dgx cd y6-context-dgx bash run-dgx.sh # Open the localhost notebook URL printed by the container.
CPU and available-device execution status is recorded in the downloaded README and validation report. DGX Spark execution requires that hardware; static validation alone does not establish GPU compatibility or performance. The browser lesson remains fully available without this extension.
A context-budget briefing with a highlighted evidence sentence and a conflict note.
Download the editable handout →