The cycle at the board-exam rung

CONCEPT

Pull out your Class 9 project logbook, and look at the six stages you mapped, in order. You know them: Problem Scoping, Data Acquisition, Data Exploration, Modeling, Evaluation, Deployment. Each stage has a defined output. The next stage needs that output before it can begin.

The Class 10 form adds something the stage names do not show: two planned switchbacks, and the condition that triggers each. The first returns from Evaluation to Data Acquisition. This switchback is taken when the model's accuracy cannot improve with the data already collected. The second returns from Modeling to Data Exploration. This one is taken when the chosen features do not explain the pattern.

At Class 9, naming the six stages in order was the required form. At Class 10 the board also tests each switchback: its direction, and the reason it is taken. Quick check: which stage do you return to when accuracy cannot improve? Which stage do you return to when the features do not explain the pattern? If you can answer both without looking, you are ready for the one-mark items.

six stages, in order — and two planned switchbacks
Problem ScopingData AcquisitionData ExplorationModelingEvaluationDeployment
Deployment → Problem Scoping  two switchbacks loop back: Evaluation→Data Acquisition, Modeling→Data Exploration
Class 9 tested the six stages in order; Class 10 also tests each switchback's direction and its trigger condition

↑ Back to top

The second layer — ethical checkpoints

CONCEPT

Look at your logbook again. Under Stage 2, Data Acquisition, did you write "whose data?" Under Stage 6, Deployment, did you write "who acts on this output?" Most Class 9 entries do not have these notes. Those questions were not on the Class 9 syllabus. They are on this one.

A second layer of decisions sits inside the cycle. The six stage names never name it directly: ethical checkpoints, at exactly two stages. At Data Acquisition, the checkpoint asks whose data is being collected, and with what consent. At Deployment, it asks who acts on the output, and whether they are equipped to question it.

Each stage name defines a technical output. It does not say who is affected by what was collected, or who acts on what was deployed. A team can finish all six stages without asking the checkpoint questions. That team has done the technical work — and may have skipped the harder decisions. This is the whole Class 10 increment to the cycle: not a new stage, but a second reading of two stages you already know.

two checkpoints the stage names do not name
DATA ACQUISITION (Stage 2) — whose data is collected, and with what consent?
DEPLOYMENT (Stage 6) — who acts on the output, and are they equipped to question it?
a team can finish all six stages and still skip these — the cycle done technically, its hardest decisions left unasked

↑ Back to top

The three domains — from names to tools

CONCEPT

You hold your phone to your face and it opens. The camera captured pixels, and a Computer Vision step ran on them. That is the whole domain — a gesture you make every day without naming it. CV is the domain whose AI step looks at pixels and returns a label or a location.

Four CV tools are already in daily use. Face unlock matches facial geometry against a stored template. A QR or barcode scanner reads the grid pattern from the image. Medical-image screening flags a region as normal or not. A licence-plate reader finds the plate, then reads its characters. In each case the step takes an image and returns a structured result. That result is a match, a decoded string, a class, or a position.

When CV fails, it fails in one specific way. It returns the wrong label with high confidence and shows no sign of doubt. The phone may let the wrong person in; the screening tool may flag a healthy region. The CV failure signature is confident, wrong, and silent about it — the model never says, "I am not sure."

pixels in, a label out — and how it fails
INPUT — an image: a face, a barcode, an X-ray, a number plateCV STEP — find the pattern in the pixels (the domain's job)OUTPUT — a label or a location: match, decoded string, class, positionFAILURE — wrong label, high confidence, no flag of doubt
the CV failure looks wrong but is never flagged: the door opens for the wrong face, with full confidence
CONCEPT

The subtitles on a video appear while the speaker is still mid-sentence — word by word, no human typist involved. That is a Natural Language Processing step running. Most people have watched it hundreds of times without naming it. NLP is the domain whose AI step reads human language and returns a structured output.

Four NLP tools show up in daily life. Auto-generated subtitles turn speech into text to display. A voice assistant turns speech into intent, then action. A machine-translation tool turns text in one language into text in another. Keyword tags in a news app turn article text into topic labels. In each case the step takes language — spoken or written — and turns it into something a system can act on.

When NLP fails, the output sounds right but means something wrong. A translation can be a correct sentence in the target language and still carry the wrong meaning. This is harder to catch than CV's failure. The CV output looks wrong — the door stays shut when it should open. The NLP output sounds fine. That is the trap: a fluent answer that is quietly incorrect.

language in, structure out — and how it fails
INPUT — human language: speech or textNLP STEP — turn the language into a structured output (the domain's job)OUTPUT — subtitles, an intent, a translation, topic tagsFAILURE — fluent and grammatical, but the meaning is wrong
CV's error looks wrong; NLP's error sounds right — that is what makes it the harder trap to catch
CONCEPT

A streaming app suggested a video. It did not read your mind. Someone built a table of watch history, an AI step read that table, and the output became the thumbnail at the top of your screen.

That step is DS — Data Science, or Decision Systems. It is the domain where the AI reads a table of features and returns one number or one category. A person then acts on what it returns.

Four tools you already know work this way. A recommendation engine turns watch history into your next pick. A weather forecast turns past sensor data into tomorrow's conditions. A student-performance predictor turns marks and attendance into a risk flag. A fraud detector turns transaction data into flag or pass.

When DS fails, the output still looks reasonable. It returns a plausible number built from a flawed table.

Consider the student-performance predictor. It may flag students with low attendance. But low attendance can reflect a part-time job, not weak ability. The table recorded what happened. It did not record why.

This is the quietest failure signature of the three domains. No wrong face, no garbled subtitle — just a believable number, with the wrong assumption buried in the table that produced it.

a table in, a decision out — and how it fails
INPUT — a table of features: marks, attendance, transactions, watch historyDS STEP — compute a number or a category from the rows (the domain's job)OUTPUT — a risk flag, a forecast, a recommendation, a fraud alertFAILURE — a plausible number with a wrong assumption buried in the table
no wrong face, no garbled subtitle — just a believable number and a buried wrong assumption; the quietest failure

↑ Back to top

Real products combine domains

CONCEPT

Think about a voice assistant. A user speaks a command. The device recognises their face, parses the words, and ranks a set of responses — one product, three domain steps running in sequence.

It is easy to assume a real AI product uses one domain. That means assuming it is either a CV product, an NLP product, or a DS product. Most real products combine domains in a single pipeline. The voice assistant uses NLP to read the request. It uses CV to recognise the speaker's face, and DS to pick the best response.

The three domains are technical roles within a system, not walls between products. The useful skill is naming which step belongs to which domain. Image to label is CV. Language to structure is NLP. Table to decision is DS. The skill is not deciding which single domain the whole product "is." Thinking the domains cannot mix is the mistake. Knowing which step is which is the readiness.

one product, three domain steps — each owns one step of the pipeline
Computer Vision
NLP
Data Science / DS
face-unlock step: image → find the face pattern → match / no-match → device unlocks
subtitle step: speech → language to text → caption shown → the reader reads it
recommend step: watch-history table → rank options → next-watch pick → shown on screen
a voice assistant runs all three: it reads the words (NLP), recognises the face (CV), and ranks the replies (DS)
The Project Cycle, Revisited — Grade 10 (417) · projected from the LATTICE via prism_html.py · register: school-g10

↑ Back to top