Generation-loss simulator

Watch text degrade across repeated copies, and see why the same settings never produce the same result twice.

Week 1 asked you to count how a copy becomes harder to read, and drew a careful line between three claims: the counting rule is deterministic, your own repeatability is something you can check, and two readers still disagree because resolving a damaged mark is an interpretation. Two small tools below run that same rule on text instead of a scanned card. Several other weeks reuse them too, as an analogy rather than the same data — see each tool's own "used in" list.

Chain simulator
Type anything, then watch it get copied from itself, generation after generation. Shows the compounding effect and the cumulative-error chart.
Character-diff visualizer
Paste two equal-length strings and see every position marked as a match, a misreading or an illegible mark — the counting rule shown inline over the text.
Resampling simulator
Set up a small "jar" of categories and watch what happens over several rounds when each round trusts only its own last guess instead of checking back with reality — an illustrative analogy of week 9's shape, not its dataset or seeds.

Chain simulator

Used in: Week 1: The office that ate itself, Week 6: Reading the damage backwards, Week 7: When ancestry becomes hard to read, Week 9: The curse of recursion

Text to copy

Corruption model

A copier's error either follows the world's own randomness, or it follows a fixed rule — the same distinction week 1's worked answer draws between the algorithm and the world it measures. Switch between them on the same text to see how differently each one degrades it.

Random (rate-based) — each character has an independent chance of damageFixed rule (every Nth character) — no randomness at all

Generations: 8Error rate per generation: 6%

Generation 8 output

Cumulative error, by generation

G0 is your original text, uncopied. G1is one copy of it, G2 a copy of that copy, and so on — each bar is the percentage of characters that now differ from G0.

Which confusable characters this run affected

Week 1's card was built from characters that look like one another — 0O, 1lI, 5S, 8B, 2Z, 6G, 9g. Each bar is how many times a character from that group was misread or went illegible in this run.

Count (final generation vs. original)

A misreading is a character swapped for a different one (like 0 becoming O); anillegible is a character replaced with · because it could no longer be resolved at all.

Misreadings
Illegibles
Total

Character-diff visualizer

Used in: Week 1: The office that ate itself, Week 5: The scriptorium

The chain simulator above always compares your original text against its own generated output. Paste any two equal-length strings here instead — your own original and a chosen generation from above, for instance — to see week 1's counting rule marked inline over the text itself, character by character.

OriginalCopy

Character by character

match misreading illegible

Misreadings
Illegibles
Total

Resampling simulator

Used in: Week 9: The curse of recursion

Week 9's pack uses its own real dataset and a seeded random generator — this tool is a separate, editable illustration of the same idea, not that data or those seeds. Try setting one category really low (say 2-3%), then compare the "fully recursive" mode against "baseline" over several rounds and watch what happens to it.

Every round below draws 200 samples, like drawing 200marbles from a jar. The question this toy asks is simple: does each round check the real jar, or does it just trust the last handful it drew? If a round only trusts the last handful, and that handful happened to be a bit short on one colour, the next round assumes that colour is rarer than it really is — and draws even less of it. Do that enough times and a colour that was always in the jar can vanish from your guesses completely, even though it never left the real jar.

A concrete example: say 5% of your customers really do order decaf coffee. One week, by pure chance, your small sample of orders only has 3% decaf. If next week's ordering guess is built only from that one sample (no re-checking with real customers), you'll stock even less decaf. That makes next week's sample even lower — maybe 1%. Repeat this for a few weeks and your system can end up "deciding" nobody orders decaf, even though 5% of customers always did. This is the same problem behind "model collapse": an AI that keeps training on its own past guesses, instead of real data, can quietly forget the rare things it once knew — try the "Fully recursive" option below to watch it happen.

1. Set up the world

What's in the jar? (write it as label:weight, comma-separated — e.g. "Cats:70,Dogs:25,Birds:5")

2. Pick how each round re-estimatesBaselineEvery round checks the real jar directly, like re-polling actual customers each time. Nothing from a past round ever carries forward, so there is nothing to pile up.Fully recursiveEach round only looks at what the last round drew — like a photocopy of a photocopy. It never checks the real jar again, so any small mistake gets copied forward and can grow.Retained mixEach round takes what the last round drew, then also takes one fresh peek at the real jar. Reality gets a say every time, so it cannot drift as far.

Round 0 always checks the real jar in every mode — the three modes only start behaving differently from round 1 onwards.

Rounds: 5

3. Watch the rarest category

Full distribution, round by round

Think of each row as a jar of 200 marbles for that round — every coloured stripe is one category's share of that round's draws. This is not week 9's own dataset or seeds. Every control here draws a fresh set, so switching mode changes the randomness and the mode at once — to compare two modes on like-for-like luck, switch back and forth a few times rather than reading one pair of runs as the difference.

Arriving from a workshop page? The tool above may already show that week's own example — look for the highlighted section, and change the text freely, nothing here is graded or saved.

See Week 1: The office that ate itself for the graded activity these tools supplement — they do not replace the pack, the worksheet or the worked answer.