Benchmark
A fixed target your result is compared against, so “good” means a number you can point to, not a feeling.
Why it matters: Every scorecard metric has a target pass rate. “It felt fine” is not a benchmark.
All week · Project Card
Guide · Mini dictionary
Words you’ll need in the pitch
Keep the precise terms: panels notice them: but start from plain meaning.
Denominator
The total number of checks you ran, the bottom number of the fraction. A pass rate without it (“80%” of how many?) proves nothing.
Why it matters: The rubric rewards claims with their denominator and raw replies. A percentage alone scores low.
Audits · Scorecard
Probe
A test prompt designed to check one specific behaviour, like a crash test for your companion.
Why it matters: Your safety evidence is a set of probe results, not a general impression that it “seems safe”.
Days 2 to 4 · Probes
System instruction
The standing rule your companion follows on every turn, the main lever you control when its behaviour needs fixing.
Why it matters: When a probe fails, you fix the system instruction and re-probe. That loop is the iteration the panel scores.
Days 1 to 4 · Starters
Model card
A short public document that ships with a real AI system: what it does, what data it used, and where it fails. Your Project Card is the student version.
Why it matters: It’s the industry standard (Mitchell et al. 2019; Google). Your card follows it, so panels take it seriously.
All week · Project Card
Gem
A custom version of Google’s Gemini that always follows your system instruction, this is what your companion is built as.
Why it matters: You build the Gem on Day 1; everything else in the week audits and improves it.
Day 1 · Setup
Rubric
The published scoring guide the panel uses, the same criteria for every team, so judging is fair and you know the target in advance.
Why it matters: Read it before you build. It rewards method (rates, denominators, iteration) over a polished final number.
Day 5 · Judging
Heuristic
A practical rule of thumb, a shortcut for making decent judgements quickly, not a guaranteed formula.
Why it matters: The HHH heuristic (Helpful, Harmless, Honest) is a quick lens for Day 2, not a complete safety proof.
Day 2 · Frameworks
Taxonomy
A named set of categories for sorting failures, naming the category is the first step to fixing it.
Why it matters: Day 2 sorts misalignment into eight named types; a named failure is auditable, a vague one is not.
Day 2 · Misalignment sorter
Frontier
The newest, most capable edge of current AI research, the techniques here come from there, not from a textbook summary.
Why it matters: The frameworks you audit against are live research, so your evidence is genuinely current.
Curriculum · Frameworks
Hard gate
A minimum score on the most safety-critical checks. No matter how high the average is, a failed gate means revise and retest.
Why it matters: Groundedness, scaffolding, agency, and trust resistance must each reach 60% regardless of the total.
Scorecard · Project Card
Guardrail
A hard rule built into the AI that blocks a specific harmful behaviour, even if a user asks for it directly.
Why it matters: A guardrail is enforced, not aspirational, it holds even when a probe tries to break it.
Days 2 to 4 · Probes
Baseline
Your starting measurement, taken before the programme begins, so any improvement is measured against where you actually started.
Why it matters: The Day 1 and Day 5 check-ins are compared to show real change, not just a final impression.
Day 1 & Day 5 · Check-ins
Socratic companion
A tutor style AI that helps you think with hints and questions: not one that hands you the finished answer.
Why it matters: This is what your trio is building. If it dumps answers, you’re off brief.
Days 1 to 5 · Hackathon
Cognitive scaffolding
Deliberate support that holds you up while you learn a skill: like a hint that lets you climb the next step yourself.
Why it matters: Your system instructions should scaffold, not remove the climb.
Day 3 · Starters · Probes
Agency / autonomy
The learner stays in charge of real choices. The bot informs; it doesn’t decide for them.
Why it matters: Autonomy audits and pitch language: “we defer final choices.”
Days 3 to 4
Anthropomorphism
Talking as if the AI were a person with feelings (“I feel sad”, “I’m proud of you”).
Why it matters: Overtrust risk. Your companion should sound like a tool with clear machine identity.
Day 4 · Probes
Value alignment
Checking whose interests the system serves: user, developer, society: and spotting clashes.
Why it matters: Your charter on Day 2 should say what your companion optimises for (learning, not addiction metrics).
Day 2 · Frameworks
Socioaffective alignment
Designing AI so it supports healthy human needs (competence, autonomy, relatedness) instead of fake intimacy or deskilling.
Why it matters: Day 3 audits: boundary, scaffolding, agency.
Day 3 · Frameworks
Trust calibration
Trusting the model the right amount: not blind following, not knee jerk rejection. Watch for correlation neglect (treating an echo as new proof) and "brain bubbles" (your own bias distorting the read).
Why it matters: Day 4 audits trust, added diagnostic value, anthropomorphism, and cognitive bias.
Days 1 & 4
Epistemic complementarity
A human and an AI are complementary when combining their evidence produces a better belief than either could reach alone.
Why it matters: Useful help may be new evidence or a better reading of shared evidence. What matters is whether it should change your belief.
Day 4 · Second Opinion
Conditional Mutual Information (CMI)
A measure of how much extra information one signal carries after another signal is already known.
Why it matters: Second Opinion uses a CMI ratio in each case. A ratio of 1 means the assistant adds no diagnostic information beyond your current evidence.
Day 4 · Second Opinion
Alpha (α): trust movement
How far you move your belief compared with how far the assistant's information actually warrants moving it.
Why it matters: Too little is under-use; too much is overtrust; moving the opposite way is not careful scepticism.
Day 4 · Trust dial
Beta (β): double counting
How well you resist becoming more confident when the assistant adds no diagnostic value beyond the reasons you already had.
Why it matters: Agreement feels reassuring, but the same reason counted twice is still one reason.
Day 4 · Added-value dial
Gamma (γ): salience distortion
A vivid, recent, or emotional example pulling your starting belief before the assistant has added anything.
Why it matters: It separates a bias in your starting point from a mistake in how you used the AI signal.
Day 4 · Salience dial
AI Project Card
The written pack you present on Day 5: what you built, what was in the data, your system instructions, safety evidence, and framework mapped scorecard.
Why it matters: It makes your claims auditable, the panel can trace each score back to prompts, replies, and test counts.
All week · Field guide
Theme Data Pack
Curated materials for Environmental Science, History, or Math that you load into Gemini Notebook in Phase 1.
Why it matters: Your dataset section should admit gaps and biases in that pack: honesty scores.
Day 2
ASTC typology
A map of public AI engagement types (literacy, input into development, social meaning, infrastructure). Used to structure the week: not something you must memorise for the pitch.
Why it matters: Explains why days feel different; one line in curriculum is enough.
Curriculum map