BILGE, Missions, and Classroom Scenarios — AfetAkademi BKT/Piaget Series Part 3 | Sibel Kaçar
ai-ed
BILGE, Missions, and Classroom Scenarios — AfetAkademi BKT/Piaget Series Part 3
March 14, 2026
In Part 1 I explained why I built a 13-model BKT architecture instead of the standard single-model approach. In Part 2 I described how I mapped Piaget's cognitive development stages onto that architecture's parameters. But that was still the "engine in the background."
In this post I describe how I turned that engine into the BILGE mentor character on the child's screen, the mission system, and classroom-oriented scenarios.
When designing AfetAkademi's Unity game prototype, I wanted children not only to find the right answers but also to feel safe.
Disaster themes—earthquake, flood, fire—naturally raise anxiety. Seeing only an avatar and red warnings on screen can easily slip from "game" into fear, especially for 8–10 year olds.
So in AfetAkademi the player has BILGEbeside them: a mentor who appears from time to time and speaks, but does not give instant answers to every question.
I designed BILGE to do three things at once:
Be a companion: Reduce the "you're on your own" feeling. The child hears not only their avatar but a guiding voice.
Offer Socratic mentoring: Instead of stating the solution, ask short questions that open thinking.
Be the model's voice: Translate what the BKT model in the background sees into language the child can understand.
For example, when the BKT model sees three wrong answers in a row on a skill, I only "know" something like "P(L) dropped." BILGE's job is to turn that cold number into a sentence like:
> "This part seems a bit tricky. Want to look at a small example first, then try again?"
So BILGE is both an emotional buffer and the interface of the learning engine.
The mission system: breaking the scenario into learning units
In AfetAkademi's game scenes there is a larger disaster story in the background: earthquake at home, fire drill at school, flood warning in the neighbourhood.
Leaving that story as one big chunk was both pedagogically weak and made the data hard to interpret. So I added a **mission system** on top.
Conceptually, each mission (`Mission`) contains three things:
- A name and short description (e.g. *"Pre-earthquake preparedness"*),
- A set of objectives — steps the player must complete,
-Error types and success criteria tied to those objectives.
Take the "home scenario." In this scene we want the child to make good decisions with family members before and after an earthquake. If I had left it as a single "Complete the mission" sentence, neither the game design nor the learning analytics would work.
Instead I split the scene into smaller pieces:
Mission 1:Find risky objects at home before the earthquake.
Mission 2:Move to a safe spot during the earthquake and drop-cover-hold.
Mission 3:After the earthquake, follow the family meeting plan and go to the assembly area.
Each mission has its own objectives. For example, in the post-earthquake mission one objective looks like:
> "Reach the designated assembly area within the allowed time after the earthquake ends."
This objective talks to several systems in the game:
Timer: When the earthquake effect ends, a countdown starts in the background.
Location tracking: Whether the player enters a zone tagged "AssemblyArea" or "AssemblyPoint" is tracked.
Error system: If time runs out and the player still hasn't reached the assembly area, this is recorded as a specific error type for both scoring and learning analytics.
Thanks to the mission system I give the player a clear "what should I do now?" frame and can cleanly tag which skill was tested when for BKT.
BKT + BILGE: What does BILGE say, and when?
In the previous post I showed how I set BKT parameters for three age groups aligned with Piaget. The critical question here is:
> "When the BKT model says this child is struggling on a given skill, what does BILGE actually do?"
I'll keep it simple.
I defined certain key moments for each mission:
- When the learner gives several wrong answers in a row,
- When time passes but they never try a critical step,
- Or when I see an unexpected error type (e.g. opening a window during a fire).
At those moments the BKT model updates the P(L) curve for the relevant skill, and I collapse that update into three bands:
"Probably doesn't know",
"On the edge",
"Probably knows".
BILGE's dialogues were written for these three bands. Example 1 — "Probably doesn't know"
Consider the skill of finding a safe spot during an earthquake. The learner keeps going to wrong places: by the window, in the doorway, in front of the cabinet.
After several wrong attempts the model pulls P(L) down for this skill. In that case BILGE does not state the correct spot; instead BILGE asks a Socratic-style question like:
> "Think about things that could break or fall during shaking. Is there a place away from those where you could protect your head while you drop and hold?"
This both reminds them of the concept and leaves the decision to the child.
Example 2 — "On the edge"
Sometimes the learner finds the right spot but forgets **drop-cover-hold**. In the model this is a different step in the same skill set; P(L) is not fully confident.
Then BILGE gives a more focused hint:
> "You found a safe spot. Now, do you remember how to protect your body? Where do you hold when you drop?"
Again, BILGE doesn't say "Do drop-cover-hold"; BILGE wants the child to produce the answer.
Example 3 — "Probably knows"
Some learners replay the game after completing the mission several times without errors. The BKT model has pinned P(L) high. Here BILGE's tone shifts:
> "You've really learned this part. We could move on to a harder scenario if you like."
That line is when the adaptive mission system kicks in—opening the door to more complex, multi-step scenarios.
Use in the classroom and research context
Right now the AfetAkademi prototype in the Unity project is mainly a tool I use for **thesis and research sessions**. So it's not yet a packaged product where "the teacher opens it and runs it minute-by-minute in class"; it's more of an **experimental environment** to understand which mission steps different age groups struggle with.
A typical session goes like this:
Short intro**
I explain the context of the scene to the child or small group: "There's an earthquake at home, then you choose what to do with your family." I also clarify in my head which steps I'll observe in particular.
Playing the mission
As the participant plays, BILGE steps in from time to time—sometimes asking more questions, sometimes just confirming. In the background the BKT model tracks in particular:
- What the player does when the earthquake ends,
- Whether they reach the assembly area within the defined time,
- What warnings they get when they insist on staying in the wrong place.
Post-session review
After the game is over I look at the anonymous session data to see which steps had more errors and which BILGE phrases helped. That information then shapes both the game's own mission design and the teacher dashboard and class summaries I'm planning on the web side.
So it's more accurate to read the flows I described in this post as, for today, a **"research prototype used in the thesis"** and, in the medium term, a **design sketch** to be adapted for classroom use.
Data, ethics, and child privacy
However strong BILGE, missions, and classroom scenarios are, every system that works with children's data faces the same questions:
- "Who do these data belong to?"
- "Who can see what, for how long?"
- "What happens if a child wants to leave this game?"
From the start, in designing both the game prototype and the web tools that talk to it, I've stuck to this principle:
"Never collect more data than needed to support learning."
So in both the research data I collect in the game prototype and the class-code / alias model I designed for afetakademi.com.tr, I apply the same idea:
- No student name, school ID, or other personal identifier.
- Only a class code and a system-assigned alias when needed.
- The aim on the teacher or researcher side is not to label individual children but to think at the group level via outcome-based summaries.
The data used on the research side are fully anonymised. The model I built with BKT and Piaget is not for tagging individual children but for understanding which kinds of missions and hints work for which age groups.
This is also an important criterion in inclusion-focused programmes like EşitBiz: increasing access without adding extra risk or stigma for any learner.
Next step: From BILGE to the report card
In this post I described how I brought the 13-model BKT and Piaget age groups to life in the BILGE mentor and mission system for use in the prototype and eventually in the classroom.
In the next part of the series the focus will shift a bit more to the teacher:
- How did I design the class summary screen?
- Which metrics actually help, and which are just "nice graphs"?
- How will the progress report connect AfetAkademi's data to real classroom decisions?
If you're curious how such tools could work in your classroom, you can follow updates on the Unity game prototype on sibelkacar.com under AfetAkademi and Research, and explore the disaster education web platform afetakademi.com.tr. The two are different legs of the same vision: one a research game, the other a scaled web app for classroom use.
This post is the third part of the Building AfetAkademi series. For technical detail see the first two posts; for application and classroom scenarios this post; for ethics and data architecture, the upcoming fourth.