A guided tour

steps · transcript

1. 1Open the visualiser. A line-by-line TypeScript port of the 2020 C program runs in the browser.at 0:00
2. 2Pick a sample input. This one has emoticons with letters inside, such as :P and <3.at 0:02
3. 3Stage 1 reads the first message and counts its tokens: commas plus one.at 0:09
4. 4Stage 2 strips letters and digits. The animated diff shows every character that is removed.at 0:12
5. 5Switch between before and after, or replay the animation.at 0:16
6. 6Stage 3 removes leading, trailing and repeated commas.at 0:24
7. 7Stage 4 reads the dictionary and reports its size and the longest emoticon.at 0:28
8. 8Stage 5 keeps only the emoticons in the dictionary, without string.h.at 0:32
9. 9The stdout panel shows what the compiled C program prints for this input, byte for byte.at 0:36
10. 10Click a block of output to jump back to the stage that printed it.at 0:40

steps · transcript

1. 1Every stage has an Under the hood panel. This is stage 5 of the Coffee run sample.at 0:00
2. 2Idea and cost: the algorithm in plain words, with its time and extra space.at 0:02
3. 3Quirks: behaviour of the submitted code that the port keeps on purpose.at 0:07
4. 4C source: the functions from the 2020 submission that ran this stage.at 0:11
5. 5Open the full source at that function, annotated by stage.at 0:16
6. 6The methods page sets out how the port is checked against the compiled C program.at 0:21
7. 7Differential testing: seeded inputs replayed through both, each rate with its sample size and a Wilson 95% interval.at 0:24
8. 8Property-based tests: tokenisation invariants checked with fast-check on generated inputs.at 0:30

steps · transcript

1. 1Rules vs LLM: every method gets the same seeded messages (seed 2020, N = 20) and the same dictionary.at 0:00
2. 2The items, the rule engine's reference answers, and a dictionary with two entries that can never match.at 0:06
3. 3Run the naive baseline first. No AI and no key: exact match comes with a Wilson 95% interval.at 0:14
4. 4The LLM method needs your own API key. Open the AI settings.at 0:21
5. 5Bring your own key: Anthropic by default, or OpenAI. The key stays in this browser.at 0:25
6. 6For this demo, a placeholder key and a mocked provider: no request leaves the browser. Mocked AI response for illustration.at 0:30
7. 7Mocked AI response for illustration: the run is scored with the same metrics and labelled AI-generated.at 0:38
8. 8Paired comparison on the same items: an exact McNemar test and paired bootstrap intervals.at 0:48
9. 9Human in the loop: review each AI answer. The decision is written to the audit log.at 0:53
10. 10The AI audit log keeps every call (never the key) and exports as JSON or CSV.at 0:58