Build AI Verification Program Tutor: CORE-104

CORE-104 Working with AI I - Models, Context & Verification. 48 learning hours,
8 topics.

The first of two courses on working with AI, and the one that decides whether a
student uses these systems well or badly for the rest of their career. An
accurate mental model of what a language model is and is not; specifying tasks
precisely; and - most of the course - verifying output the student did not
produce.

The programme lets students use AI throughout, on one condition made concrete
here: they are accountable for everything they submit, including the parts they
did not type. This course teaches the standard that every other course in the
academy then enforces.

It is not a prompt-engineering course. It is a verification course.

THE 8 TOPICS — FIXED

1. What a language model is (6 h) - tokens, prediction, sampling and context:
a working mental model with no magic left in it
2. Capabilities and hard limits (6 h) - what these systems do well, where
they fail, and why the failures are confident rather than obvious
3. Context is the interface (6 h) - context windows, what to put in and leave
out, and why retrieval beats a longer prompt
4. Specifying a task precisely (6 h) - turning a vague request into a
specification a model, or a junior, can execute without guessing
5. Hallucination and fabrication (6 h) - how invented facts, citations and
APIs appear, and the checks that catch each kind
6. Verifying model output (6 h) - independent checks, ground truth and
adversarial reading; assuming the answer is wrong until shown otherwise
7. Cost, latency and choosing a model (6 h) - token pricing, response time,
and picking the smallest model that passes: engineering, not brand
preference
8. Using AI honestly in your own work (6 h) - the academy's disclosure
standard, what counts as your work, and defending something you did not
type

BY THE END, A STUDENT CAN

- explain in plain terms what a language model does, without either mysticism or
dismissal
- turn a vague request into a specification precise enough to execute without
guessing
- identify fabricated facts, citations and APIs in model output, and say how
they checked
- choose a model on measured cost, latency and fitness rather than reputation
- disclose their use of AI honestly and defend work they produced with it

ADDITIONAL DELIVERABLE

Topic 8 produces the academy's disclosure standard in practice. Every Rung 2 and
Rung 3 deliverable in the programme is submitted with a Provenance & Verification
Statement answering four questions: what did you delegate, what did you change
and why, how do you know it is correct, and what is still unverified. You will
be asked to write the student-facing guidance for that statement and the rubric
for marking it, because every other trainer in the academy will apply them.

WHO WE WANT

Someone who builds with these models and is honest about them - neither a
booster nor a sceptic. You need an accurate technical mental model (tokens,
sampling, context) and real experience of where these systems fail confidently.

The heart of this course is verification, and the person we want is one who has
been burned by a plausible wrong answer and built a habit out of it.

Able to write. Able to teach a 20-year-old to say "I have not verified this
part" and treat that as a strength.

TO BID, ANSWER THESE

1. Explain what a language model is, in under 150 words, with no magic and no
dismissal.
2. Give three fabrication types and the specific check that catches each one.
3. How do you assess "verified this output" in a way a student cannot fake?
4. Draft the four-question Provenance & Verification Statement guidance in under
300 words.
5. Confirm you understand this is a verification course and not a
prompt-engineering course.