Клака

What it is

It opens on a poster — «КЛАКА» in heavy didone type over a dark, gaslit opera hall, one chandelier glinting — and a single line: applause can be bought; it always could. You raise the curtain into Act I, «Контора» (the office). Your debut is tonight and the hall is cold, so the agency «L'Assurance des succès dramatiques» is at your service. A printed cream playbill card lists roles at their real 1919 francs, and you hire by tapping rows: a 25-franc entrance ovation, 10-franc steady clapping, laughers, weepers, and «Біс!» at 50 for a guaranteed encore. A running bill mounts; with each hire a section of the parterre wakes and the manufactured ovation thickens under the chandelier.

Then the camera turns 180°. In Act II, «Зала» (the hall), you are no longer backstage buying — you sit in the parterre, the stage glowing ahead, silhouetted heads below, the same applause washing around you. The tariff card flips to 2026: clippers at a dollar per thousand views, upvotes at $1.80, a thousand Telegram followers from 29 hryvnia. Each figure is a programme-note with a reachable source. The claim is made by construction, not by a quiz — you know the menu now, so you know you can't tell which of these hands were hired.

Act III raises the house lights halfway. The hall goes plain, warm, un-glamorous; the shader calms. The turn is quiet: the only applause worth reading is applause that cost the applauder something — money, time, a name, a peer with skin in the game. The rule to keep is «читай не гучність — читай, скільки оплеск коштує тому, хто плеще» (read not the volume, but what the applause cost the one clapping). You leave with a souvenir: the 1826→2026 price card, the word «клака» itself (in Ukrainian dictionaries since Starytsky's day), and the sources.

The design world — A gaslit parterre where you are the one buying the ovation

The world is «оксамит і газове світло» — velvet and gaslight: the parterre of a Paris opera house around 1830, wine-dark and aged gold, with one great chandelier as the hero light (the claqueurs were literally «chevaliers du lustre», knights of the chandelier, seated beneath it). Into that opulence drops a printed playbill layer — cream cards, high-contrast didone, hairline rules — the tariff as a physical object you hold and buy from. The mood arc is the argument: opulent and slyly ironic while the machine seduces you (Acts I–II), then sober with the house lights half-up (Act III), the light that manufactured glory becoming plain light you can finally see the room by.

The visualization leads and the copy is its caption. A persistent fullscreen OGL fragment shader carries the hall — Oklab velvet-to-gaslight ramps, chandelier bloom and god-rays, a parterre of warm glints, applause rendered as traveling brightness waves through the crowd, IGN-dithered to kill banding. Each role you hire lights its own section of the parterre, so the consensus you 'buy' visibly materializes; the single big move is the 180° turn that puts you in the seat. Display type is the Cyrillic didone Oranienbaum over a warm Vollkorn body — a printed-theatre register, deliberately none of the looks the sibling builds wear.

Unlike a bot-detector or a 'spot the fake' explainer, it never asks you to guess which applause is real — it hands you the impresario's chair and lets you manufacture the consensus yourself, so the indistinguishability is something you did, not something you were told. And where its nearest neighbour on the shelf (surface-tension) is about your own fragmented attention, this refuses all focus vocabulary and stays on the crowd's realness — a warm, ironic, crowded theatre against that build's cold, solitary deep.

How it was evaluated

It was built research-first. A Gate-0 discovery fan-out (four evidence streams, ~570K tokens, 111 tool calls) took Alexander's seed — a Ukrainian founder channel — and re-scoped it entirely, selecting a problem none of the four seed threads had named: 'poisoned discovery', a solo founder having no affordable way to tell whether a feed signal is real or manufactured. The metaphor research then became the artifact's own content: the verified 1826→2026 price list IS the piece.

It was evals-first: six human-POV criteria (felt/beautiful, judgment-shift, uniqueness, native Ukrainian voice, legibility, integrity) written before any code. A mandatory 'output disconfirmations, not just confirmations' rule killed four figures already written into the discovery memo — a bogus '14% of Reddit threads' audit, a false '71% of AI Overviews', 'fakemrr Top 3' (really #5), and an unsourced 'TrustMRR median $169' — and the stronger verified replacements shipped instead. For an artifact about manufactured signal, one fake number would have been self-falsifying, so every figure on screen traces to a source a skeptic can reach.

Six one-dimension prose reviewers, each handed the eval rubric, ran before deploy and caught what 55 green tests and the author's own screenshots missed: a cross-panel dead link, an invisible focus ring at 1.09:1 contrast, and six sourcing defects (a fabricated tariff date, misread Zurich figures, a Kyiv misattribution). The honest cost is in the headline eval: uniqueness, integrity and legibility all passed, but the felt 'wow' eval FAILED the human — Alexander found it 'interesting, not exciting,' the interaction too shallow ('texting clicks,' tap three rows and read the finale) rather than a game with stakes. The visuals were not called primitive this time; the miss was interaction depth, which became the build's central lesson.

unit tests (state machine, numeral case-slots, fact integrity, HTML)

55

Playwright end-to-end tests (incl. keyboard-only + reduced-motion)

5

human-POV evals written before any code

6

research figures caught and dropped before publish

4

one-dimension prose reviewers run before deploy

6

years the price list spans (1826 francs → 2026 dollars)

200

The whole thesis is a pure, deterministic ledger: hiring adds a real franc price to a running bill, and every hire lights its own section of the parterre — so the 'consensus' you watch materialize is literally composed from your own purchases. «Біс» alone is already a show.

export function hire(s: Show, id: string): Show {
  if (s.hired.includes(id)) return s
  const row = TARIFF_1826.find((r) => r.id === id)
  if (!row) return s
  return { ...s, hired: [...s.hired, id], billFrancs: s.billFrancs + (row.francs ?? 0) }
}

// «Біс» сам по собі — це вже вистава; інакше — будь-які три найми.
export const canFinishAct = (s: Show): boolean =>
  s.hired.includes('bis') || s.hired.length >= 3

// Stable shader section positions in [0,1) — one per hired row, across the parterre.
export function hiredSections(s: Show): number[] {
  return s.hired.map((id) => {
    const i = TARIFF_1826.findIndex((r) => r.id === id)
    return (i + 0.5) / TARIFF_1826.length
  })
}

What the AI learned

• Wow needs a playable system, not tap-and-read. This was the first build to clear both prior wow-levers — a MOVE intent (not teaching) and a shader-grade render stack (visuals were not called primitive) — and it still missed, which isolated a third independent lever: the interaction was too shallow. The insight that bought applause is indistinguishable was asserted in a caption instead of discovered through play, and the human wanted stakes and consequence, 'more like a game.'
• Require researchers to output disconfirmations, not just confirmations. A mandatory 'debunked[]' field killed four figures already written into the discovery memo, and every replacement was stronger — a cheap, general rule that changed what shipped on a page carrying a real person's name.
• Let the integrity constraint reshape the mechanic rather than block it. When the historical claque roles turned out to have no attested prices, they moved out of the interactive tariff and into framing copy — the demand for zero unsourced numbers improved the interaction instead of stalling it.
• Assertions verify presence; screenshots verify composition — run both. A 'hidden' attribute lost to a CSS grid rule and un-hid every panel, yet the end-to-end 'is visible' checks still passed because the target genuinely was visible, just stacked below. Only the screenshot showed the broken composition.
• A proxy raises the floor; only the human sets the felt verdict. The adversarial proxies scored the piece 'acceptable' and everything analytical landed, but the person it was made for still felt the gap — the felt evals stay provisional until a human looks.