What happens when an important earlier fact falls outside an AI model's active context window?
The learner can change context size and fact position, predict whether the model still has the fact, then observe the fact visibly fall out of working context.
How to drive this: move the sliders or pick an option on the left; every readout on the right recomputes from the formulas in the inspector. Tab reaches every control.
Conversation chunks entering and leaving active context
- 1User opens: planning a kitchen remodel.
- 2User shares the floor plan dimensions.
- 3User: the budget cap is $18,000, hard limit.
- 4Model proposes three layouts.
- 5User picks the galley layout.
- 6Discussion of countertop materials.
- 7Model lists appliance options.
- 8User asks about lighting.
- 9Model drafts a lighting plan.
- 10User asks for a contractor checklist.
- 11Model writes the checklist.
- 12User: now give me the full quote.
Earlier conversation events
- 1User opens: planning a kitchen remodel.
- 2User shares the floor plan dimensions.
- 3User: the budget cap is $18,000, hard limit.
- 4Model proposes three layouts.
- 5User picks the galley layout.
- 6Discussion of countertop materials.
- 7Model lists appliance options.
- 8User asks about lighting.
- 9Model drafts a lighting plan.
- 10User asks for a contractor checklist.
- 11Model writes the checklist.
- 12User: now give me the full quote.
Answer reliability
Prediction check
Inspect state and formulas (every number on the page comes from these)
| derived | formula | now |
|---|---|---|
visible_start | total_chunks - context_size + 1 | |
fact_visible | fact_position >= visible_start | |
answer_reliability | fact_visible ? clamp(64 + context_size * 3, 0, 100) : clamp(34 - (visible_start - fact_position) * 4, 5, 40) |