Comet pins the SH-1 packet to the corkboard. "The builder made SH-1," she says, "so the builder can answer every question about it."
Wren points to section P2. "What does the evidence say? SH-1 is built on a large pre-trained model from another company. The builder did not make that part."
"Oh," Comet says slowly. "So SH-1 has parts from people we have never met."
Teacher Mara joins them. "And when SH-1 struggled with Spanish, whose part caused it?"
Nova draws a row of linked boxes on the wall, each one blank. "Would you like a hint?" she asks. "List every part SH-1 depends on, and who made each one."
Comet hands you a marker. "Reviewer, fill in the chain."
Generative AI is built from many outside parts. They include datasets from other sources, pre-trained models and software libraries.
All the parts and makers a system depends on make up its value chain.
Parts that were badly obtained or never checked make a system less transparent and accountable.
The training data may also be too large for people to check.
| Part of SH-1 (made up) | Who supplies it | What the crew knows | What the crew does not know |
|---|---|---|---|
| Base model | The base model maker, a third party | It is a large pre-trained language model | What it was trained on |
| Study material for tuning | The builder | The packet calls it "study material" | Which material, in which languages |
| Software libraries | Not listed | Nothing | Whether any are used, and who made them |
| Class notes | Harbor Point teachers | Exactly which pages | Nothing important |
Third parties are outside providers of data, models or systems. The base model maker is a third party for Harbor Point.
What a third party supplies can be complex or hard to see into. Its idea of acceptable risk may not match the school's.
When a system uses many outside parts, it can be hard to tell which part caused a problem.
Problems in big shared models can flow into every system built on them, like a bottleneck.
The crew looks at the R9 row: value chain and component integration. Its SH-1 example is "base model and its training data unknown."
They rate likelihood High. Right now, nobody at Harbor Point can see into the base model or its data.
They rate size of harm Medium. A hidden problem could reach students' studying, but SH-1 is a study helper, not a grader.
These ratings are the crew's judgment in the story, not facts.
| Row | Risk | SH-1 example | Likelihood | Size of harm |
|---|---|---|---|---|
| R9 | Value chain and component integration | Base model and its training data unknown | High | Medium |
| Statement | True or false? |
|---|---|
| Generative AI value chains can include datasets, pre-trained models and software libraries. | ? |
| The builder made every part of SH-1 itself. | ? |
| With many outside parts, it can be hard to tell which part caused a problem. | ? |
| Training data may be too large for people to check. | ? |
| A problem in a big shared model stays inside one system. | ? |
Sharp tracing, reviewer. Tomorrow is Impact Friday: who gets to use SH-1 at all?