Step 6 · Discussion and evaluation (Criterion D)

The highest-weighted criterion: what your findings mean, and how far they can be trusted.

By the end of this step

Your full draft, for written feedback from your supervisor.

Two things, worth 8 marks

Strand	What it means	Top level looks like
Discussion	A balanced view of how significant your findings are, supported by evidence	Fully supported by evidence; considers other interpretations and the wider computing context
Evaluation	How effective your essay was: method, data, sources, strengths and limitations	Strengths and limitations are explained, not just listed

The IB guide saysEvaluation should run through the essay, not sit in one section at the end. Reflect on dataset choice, parameters and metrics; say whether your findings agree with other studies; make suggestions for further work that are relevant and thought through.

Stated → described → explained

✗ Weak

“A limitation is that I only used one computer.”

✓ Strong

“All runs used one laptop with a 12 MB L3 cache. Because linear probing's advantage at low load factors comes from cache locality, a machine with a smaller cache may show a smaller gap, so the crossover point (α ≈ 0.55) should not be generalised. Probe counts, which don't depend on hardware, support the same trend, so the main conclusion still holds.”

The strong version says what the limitation is, how it affects this result, and how much it matters.

Every limitation answers three questions What is it? · How does it affect this result? · How much does it matter?
Only the first → stated. The first two → described. All three → explained (top band).

Stated, described or explained? Self-check drill

Climb the ladder: how far does each limitation go?

Is the difference real?

Is the gap bigger than the noise? Quick check

A question to ask of every comparison you discuss.

Real life: results that vanished with a different seed In 2018 researchers (Henderson et al., Deep Reinforcement Learning that Matters) showed that published comparisons between reinforcement-learning algorithms could flip just by changing the random seeds. Discussing reliability isn't nit-picking: it's what real computer scientists argue about.

Where to look for strengths and limitations in CS

Area	Questions to ask
Hardware and software	Would another CPU, OS, language or interpreter change the result? Did background processes add noise?
Implementation	Are both techniques implemented equally well? Library vs your own code? Any bugs found and fixed?
Data	Is the dataset representative? Big enough? Biased? Synthetic vs real?
Metrics	Do your metrics really measure what the RQ asks? What did they miss (e.g. memory, energy)?
Range	Would the trend continue beyond the values you tested?
Reliability	How big was the spread between repeats? Are differences larger than the noise?
Sources	Peer-reviewed? Up to date? Vendor benchmarks with an interest in the result?

Further research that's worth writing

✗ Weak

“In future I could test more algorithms and use more data.”

✓ Strong

“Testing Robin Hood hashing, which reduces variance in probe length, would show whether the degradation above α = 0.7 is a feature of open addressing in general or of linear probing in particular.”

Limitations of this study:

1. It was done by me. 2. On a laptop. 3. During a thunderstorm. (Not specific enough: what did the thunderstorm actually do to your results?)

Ready to move on?

I discuss the significance of my findings, not just what they are.
I consider at least one other interpretation of my results.
Strengths and limitations are explained, with their effect on the conclusion.
Evaluation appears throughout, not only in one section at the end.
My suggestions for further research are specific and follow from my findings.