MAKING-OF · SITE 208 · WAVE 10, SECRETS
A histogram that quietly betrays a cipher
THE CONCEPT
AL·KINDĪ is a fictional single-exhibit museum page for the ninth-century Baghdad treatise that invented frequency analysis. Its audience is anyone who has ever been told "this cipher is unbreakable"; its single job is to show statistics working as a weapon — an intercepted dispatch counted apart live on the sheet, in one uninterrupted take, with the failure case (too little material) available on a slider. Nothing on the page is scripted theatre: every bar, every fixing and every readout number is measured from the running ciphertext.
PALETTE — ABBASID ILLUMINATION
Per the wave's ink/line split, the bright gold is reserved for the secret itself — cipher signs,
histogram bars, hairlines. Small annotations get a derived blue-grey ink; body copy gets the ivory
and a dimmed parchment tone (#B4AC94), both comfortably past 4.5:1 on the dark grounds.
TYPE
Amiri — a revival rooted in Naskh typesetting whose Latin carries the reed pen — sets the display, the body, and every glyph on the sheet, laid out in fixed cells so a cipher capital can flip to a plaintext lowercase without the line breathing. Fragment Mono runs the count HUD, the histogram labels and the fixing table: the clerk's hand against the calligrapher's.
TECHNIQUES
Vanilla canvas 2D, no libraries. The signature is the live crack, and it is genuinely derived, not replayed:
- The count. A 1,304-letter English dispatch (all 26 letters present) is enciphered at load with a seeded, fixed-point-free monoalphabetic key. As the stream arrives (a fast ramp to ~456 signs, then 53 signs/s), unigram and word-internal bigram counts accumulate; the histogram draws those counts and nothing else — the QA suite asserts its total equals the processed length at every intercept setting.
- The comparison. The observed spine is ranked against Lewand's published letter frequencies of printed English (Cryptological Mathematics, 2000) — rank-matching seeds the key guess, and a χ² distance between the two ranked spines is shown live in the HUD.
- The refinement. Every 300 ms the solver runs three hill-climbing chains (rank-init, current-best, perturbed-best; 300 swap trials each) over a score combining the Lewand unigram prior with word-habit bigram log-likelihoods counted at load from a 2,087-letter reference folio embedded in the page. A sign is fixed after its reading holds for five consecutive rounds with at least eight examples — and a fixing that the word test later rejects for three straight rounds is revised, marked in verdigris (one such revision occurred in the recorded full run).
- The resolution. Each fixing ripples down the sheet as a 450 ms gold-to-ivory crossfade, staggered 26 ms per line. After convergence the sheet pans slowly through the recovered dispatch.
- The honest failure. The intercept slider truncates the ciphertext and restarts the count. Measured degradation, deterministic per length: at the full 1,304 signs the crack reads 99.8% of the text (24 of 26 symbols; only the dispatch's single j and two zs swap); at 400 signs, 18 fixings, all correct; at 150, 13 of 20 seen symbols and no fixings survive the stability test; at 100, three of twenty — garble. The curve is noisy in between (150 happens to outperform 130) because thin material is a lottery — which is precisely al-Kindī's caveat.
| INTERCEPT | SYMBOLS RIGHT | TEXT RIGHT (WEIGHTED) | FIXED |
|---|---|---|---|
| 1,304 (full) | 24 / 26 | 99.8% | 23, all correct |
| 400 | 23 / 23 seen | 100% | 18, all correct |
| 150 | 13 / 20 seen | 84% | 0 |
| 100 | 3 / 20 seen | 33% | 0 |
State probe: window.__alkindi exposes processed/total counts, live mapping, fixings,
χ², the true key, accuracy metrics and setLength() / finish() drivers, so
the QA harness gates on solver state rather than wall-clock time.
ITERATION LOG
Raised the small-mono annotation ink from #7D7F8E (≈3.8:1 on lapis — the
collection's recurring contrast trap) to #8E93A6 (≈5:1). Fixed a garbled character-code
expression in the histogram's reference labels. The sheet canvas now shrinks to its content, so a
100-sign intercept no longer floats three lines above ten rows of empty vellum. Driving the failure
path re-set the thin-intercept preset: 120 signs happened to land a lucky deterministic solve
(12/20 symbols), so the chip moved to 100 signs, where the measured crack genuinely collapses
(3/20, weighted 33%). Enlarged the section-divider rosette; hid the panel annotations under 560 px.
The tentative reading — the garble that appears when the count runs dry — was too dim to incriminate itself; lifted it from 50% to 62% ivory and brightened the key-ring's tentative glyphs to match, since the failure is the demonstration. Added the cross-spine micro-interaction: hovering a cipher bar or a fixing cell now also lights the verdigris reference bar of its current reading, physically connecting observed count to English habit. Verified mid-work, not just at rest: locked fixings asserted correct at both 400 and 1,304 signs, and the two-frame screenshot hash confirmed the scene lives while resolving.
Chanel rule: removed the hover background wash behind sheet glyphs — the glyph colour change carries the highlight alone. Re-ran the full 27-assert suite headless: exactly one h1 per page, zero console errors (normal and reduced-motion paths), 375 px clean with nav and wordmark inside the viewport, histogram totals equal to processed length at 1,304 / 400 / 150 / 100, reduced motion settling synchronously to the converged sheet, rAF paused on hidden tabs, DPR capped at 2. The 7-second resting state measures 633 signs counted and 20 of 26 fixed — mid-crack, as the thumbnail requires.
HONESTY NOTES
Al-Kindī worked on Arabic; the page says so and demonstrates on English. Historical claims are hedged where sources vary ("around 850, by most accounts"). The reference distribution is quoted to Lewand's published table; the bigram model is counted live from the embedded folio at load, never hard-coded. The recovered dispatch itself describes the method — a second read for anyone who lets the count finish and earns the illuminated frame.