diff --git a/app/audit/audit.css b/app/audit/audit.css index 59ba768e9..fcb80742e 100644 --- a/app/audit/audit.css +++ b/app/audit/audit.css @@ -15,6 +15,8 @@ --a-red-rgb: 221 58 30; --a-red-deep: #b92c14; --a-blue: #2848c8; + --a-blue-rgb: 40 72 200; + --a-blue-deep: #1d389f; --a-open: #1d8748; --a-line: rgb(var(--a-ink-rgb) / 0.13); --a-line-2: rgb(var(--a-ink-rgb) / 0.3); @@ -41,13 +43,12 @@ font-optical-sizing: auto; text-wrap: balance; } -/* Emphasis is flagged, the way a checker flags a word. */ -.audit .a-serif em, -.audit .a-flag { +/* Emphasis is underlined in the red pencil. */ +.audit .a-serif em { font-style: italic; - text-decoration: underline wavy var(--a-red); - text-decoration-thickness: max(1.5px, 0.045em); - text-underline-offset: 0.16em; + text-decoration: underline solid var(--a-red); + text-decoration-thickness: max(2px, 0.05em); + text-underline-offset: 0.14em; text-decoration-skip-ink: none; } @@ -103,18 +104,20 @@ transform 0.15s; } .audit .a-btn:focus-visible { - outline: 2px solid var(--a-red); + outline: 2px solid var(--a-blue); outline-offset: 3px; } .audit .a-btn:active { transform: translateY(1px); } +/* The primary action is the blue of a check, so it is never mistaken for + the red of an error or for the outlined button beside it. */ .audit .a-btn-solid { - background: var(--a-ink); - color: var(--a-sheet); + background: var(--a-blue); + color: #fff; } .audit .a-btn-solid:hover { - background: var(--a-red-deep); + background: var(--a-blue-deep); } .audit .a-btn-line { border: 1px solid var(--a-ink); @@ -122,8 +125,7 @@ color: var(--a-ink); } .audit .a-btn-line:hover { - background: var(--a-ink); - color: var(--a-sheet); + background: rgb(var(--a-ink-rgb) / 0.08); } .audit .a-link { @@ -197,19 +199,6 @@ animation-timeline: view(); animation-range: entry 0% entry 45%; } - /* A proof mark draws itself as its line scrolls in. */ - .audit .a-draw { - stroke-dasharray: 1; - animation: a-draw linear both; - animation-timeline: view(); - animation-range: entry 30% entry 90%; - } - .audit .a-grow { - transform-origin: left center; - animation: a-grow linear both; - animation-timeline: view(); - animation-range: entry 10% entry 70%; - } } } @keyframes a-reveal { @@ -222,22 +211,6 @@ transform: none; } } -@keyframes a-draw { - from { - stroke-dashoffset: 1; - } - to { - stroke-dashoffset: 0; - } -} -@keyframes a-grow { - from { - transform: scaleX(0); - } - to { - transform: scaleX(1); - } -} /* The dot beside the phase that is under way. */ @media (prefers-reduced-motion: no-preference) { @@ -271,7 +244,7 @@ } } -/* The outcome icons (OutcomeIcons.tsx). Drawn complete; where motion is +/* The audit step icons (StepIcons.tsx). Drawn complete; where motion is welcome each runs one slow loop. */ .audit .a-oi .a-oi-fix { transform-box: fill-box; @@ -285,17 +258,6 @@ .audit .a-oi .a-oi-fix { animation: a-oi-fix 6s ease-in-out infinite; } - /* 2. The new sheet settles onto the stack. */ - .audit .a-oi .a-oi-drop { - animation: a-oi-drop 5.5s cubic-bezier(0.3, 0, 0.2, 1) infinite; - } - /* 3. The question fades as its answer arrives. */ - .audit .a-oi .a-oi-ask { - animation: a-oi-ask 7s ease-in-out infinite; - } - .audit .a-oi .a-oi-fix.a-oi-slow { - animation-duration: 7s; - } /* 4. The scan crosses the three papers; each check appears behind it. */ .audit .a-oi .a-oi-scan { animation: a-oi-scan 5.5s linear infinite; @@ -339,38 +301,6 @@ transform: scale(0.4); } } -@keyframes a-oi-drop { - 0% { - opacity: 0; - transform: translateY(-11px); - } - 14% { - opacity: 1; - transform: translateY(-11px); - } - 44%, - 90% { - opacity: 1; - transform: none; - } - 100% { - opacity: 0; - transform: none; - } -} -@keyframes a-oi-ask { - 0%, - 36% { - opacity: 1; - } - 50%, - 88% { - opacity: 0.2; - } - 100% { - opacity: 1; - } -} @keyframes a-oi-scan { 0% { opacity: 0; @@ -429,3 +359,214 @@ opacity: 0; } } + +/* The tower (Figures.tsx): the red paper slips out and what stood on it + comes down, on a slow loop. Drawn standing; still with reduced motion. */ +.audit .a-tw .a-tw-up { + transform-box: view-box; + transform-origin: 180px 340px; +} +.audit .a-tw .a-tw-bad, +.audit .a-tw .a-tw-loose { + transform-box: fill-box; + transform-origin: center; +} +@media (prefers-reduced-motion: no-preference) { + .audit .a-tw { + animation: a-tw-reset 11s linear infinite; + } + .audit .a-tw .a-tw-bad { + animation: a-tw-bad 11s cubic-bezier(0.4, 0, 0.2, 1) infinite; + } + .audit .a-tw .a-tw-up { + animation: a-tw-up 11s linear infinite; + } + .audit .a-tw .a-tw-loose-1 { + animation: a-tw-loose-1 11s linear infinite; + } + .audit .a-tw .a-tw-loose-2 { + animation: a-tw-loose-2 11s linear infinite; + } +} +@keyframes a-tw-reset { + 0% { + opacity: 0; + } + 5%, + 91% { + opacity: 1; + } + 96%, + 100% { + opacity: 0; + } +} +@keyframes a-tw-bad { + 0%, + 40% { + transform: none; + } + 47%, + 100% { + transform: translate(-58px, 0) rotate(-6deg); + } +} +@keyframes a-tw-up { + 0%, + 47% { + transform: none; + animation-timing-function: cubic-bezier(0.6, 0, 1, 0.5); + } + 58% { + transform: rotate(-24.6deg); + animation-timing-function: ease-out; + } + 61% { + transform: rotate(-22.4deg); + animation-timing-function: ease-in; + } + 64%, + 100% { + transform: rotate(-24.6deg); + } +} +@keyframes a-tw-loose-1 { + 0%, + 47% { + transform: none; + animation-timing-function: cubic-bezier(0.6, 0, 1, 0.5); + } + 57% { + transform: translate(-100px, 22px) rotate(-26deg); + animation-timing-function: cubic-bezier(0.5, 0, 1, 0.6); + } + 66%, + 100% { + transform: translate(-166px, 255px) rotate(-90deg); + } +} +@keyframes a-tw-loose-2 { + 0%, + 47% { + transform: none; + animation-timing-function: cubic-bezier(0.6, 0, 1, 0.5); + } + 57% { + transform: translate(-96px, 40px) rotate(-30deg); + animation-timing-function: cubic-bezier(0.5, 0, 1, 0.6); + } + 68%, + 100% { + transform: translate(-232px, 262px) rotate(-180deg); + } +} + +/* On the blue section the headline's underline is white, not red. */ +.audit .a-tone-blue .a-serif em { + text-decoration-color: rgb(255 255 255 / 0.8); +} +.audit .a-tone-blue ::selection { + background: rgb(255 255 255 / 0.28); + color: #fff; +} + +/* The benefits (BenefitIcons.tsx). When the section scrolls in, the four + arrive one after another (--i) and each icon assembles itself part by + part (--k). The server renders them complete; the Sequence box only arms + them where script runs and motion is welcome. */ +.audit .a-bn-part { + transform-box: fill-box; + transform-origin: center; +} +.audit [data-seq='armed'] .a-bn-part, +.audit [data-seq='armed'] .a-bn-text { + opacity: 0; +} +.audit [data-seq='run'] .a-bn-text { + animation: a-bn-rise 0.7s cubic-bezier(0.2, 0.7, 0.2, 1) both; + animation-delay: calc(var(--i) * 0.55s + 0.3s); +} +.audit [data-seq='run'] .a-bn-part { + animation: a-bn-pop 0.55s cubic-bezier(0.2, 0.9, 0.3, 1.25) both; + animation-delay: calc(var(--i) * 0.55s + var(--k) * 0.13s); +} +.audit [data-seq='run'] .a-bn-drop { + animation-name: a-bn-drop; +} +.audit [data-seq='run'] .a-bn-spin { + animation-name: a-bn-spin; + animation-duration: 0.85s; +} +.audit .a-bn-draw { + stroke-dasharray: 1; +} +.audit [data-seq='run'] .a-bn-draw { + animation-name: a-bn-draw; + animation-timing-function: ease-out; +} +.audit [data-seq='run'] .a-bn-twinkle { + animation: + a-bn-pop 0.55s cubic-bezier(0.2, 0.9, 0.3, 1.25) both, + a-bn-twinkle 2.6s ease-in-out infinite; + animation-delay: + calc(var(--i) * 0.55s + var(--k) * 0.13s), calc(var(--i) * 0.55s + var(--k) * 0.13s + 1s); +} +@keyframes a-bn-rise { + from { + opacity: 0; + transform: translateY(12px); + } + to { + opacity: 1; + transform: none; + } +} +@keyframes a-bn-pop { + from { + opacity: 0; + transform: scale(0.3); + } + to { + opacity: 1; + transform: none; + } +} +@keyframes a-bn-drop { + from { + opacity: 0; + transform: translateY(-14px); + } + to { + opacity: 1; + transform: none; + } +} +@keyframes a-bn-spin { + from { + opacity: 0; + transform: rotate(-110deg) scale(0.4); + } + to { + opacity: 1; + transform: none; + } +} +@keyframes a-bn-draw { + from { + opacity: 1; + stroke-dashoffset: 1; + } + to { + opacity: 1; + stroke-dashoffset: 0; + } +} +@keyframes a-bn-twinkle { + 0%, + 100% { + opacity: 1; + } + 50% { + opacity: 0.35; + } +} diff --git a/app/audit/components/Actions.tsx b/app/audit/components/Actions.tsx index d08df2f0c..30151f75f 100644 --- a/app/audit/components/Actions.tsx +++ b/app/audit/components/Actions.tsx @@ -26,7 +26,7 @@ const SIZES: Record = { export function AddToPoolButton({ size = 'md', - label = 'Add to the pool', + label = 'Fund the audit', className, }: { size?: Size; @@ -67,17 +67,3 @@ export function SubmitProposalButton({ ); } - -/** A quiet text link that starts a proposal. */ -export function ProposeLink({ label }: { label: string }) { - const { submitProposal } = useActions(); - return ( - - ); -} diff --git a/app/audit/components/BenefitIcons.tsx b/app/audit/components/BenefitIcons.tsx new file mode 100644 index 000000000..c3236e926 --- /dev/null +++ b/app/audit/components/BenefitIcons.tsx @@ -0,0 +1,179 @@ +import type { CSSProperties } from 'react'; + +/** + * Line icons for the four benefits, drawn in white for the blue section: + * results that hold (a steady stack, checked), funding that goes further (a + * coin sent to a checked paper), treatments sooner (a capsule) and smarter + * AI (a spark). Each is drawn in parts (`a-bn-part`, in the order `--k`) + * so that it can assemble itself when the section scrolls in (audit.css); + * as rendered by the server it is complete. + */ + +const LINE = { + fill: 'none', + stroke: 'currentColor', + strokeWidth: 2.2, + strokeLinecap: 'round', + strokeLinejoin: 'round', +} as const; +const TINT = 'rgb(255 255 255 / 0.16)'; + +/** The part's place in its icon's order of arrival. */ +const k = (n: number) => ({ '--k': n }) as CSSProperties; + +/** Science that holds up: three stacks set down widest first, then a check. */ +function Holds() { + return ( + <> + + + + + + + + + + ); +} + +/** Funding that goes further: a coin, an arrow, and the checked paper it reaches. */ +function Funds() { + return ( + <> + + + + + + + + + + ); +} + +/** Treatments, sooner: a capsule turning into place, half of it filled. */ +function Treats() { + return ( + + + + + + + + ); +} + +/** Smarter AI: a large spark, then a small one that keeps twinkling. */ +function Learns() { + return ( + <> + + + + + ); +} + +/** In the order of `BENEFITS`. */ +const ICONS = [Holds, Funds, Treats, Learns]; + +export function BenefitIcon({ index }: { index: number }) { + const Art = ICONS[index]; + if (!Art) return null; + return ( + + + + ); +} diff --git a/app/audit/components/Figures.tsx b/app/audit/components/Figures.tsx index c81e3e54e..3dcdb45db 100644 --- a/app/audit/components/Figures.tsx +++ b/app/audit/components/Figures.tsx @@ -1,151 +1,96 @@ -import { cn } from '@/utils/styles'; -import { LEDGER, SPREAD } from '../data/content'; -import { Ref } from './Parts'; +/* The tower -------------------------------------------------------------------- */ -/* The funnel ---------------------------------------------------------------- */ - -/** - * How wide the funnel is at the top of each stage, as a share of its column; - * the last value is the spout. Chosen to read as a funnel, not to scale: - * the real figures span four orders of magnitude. - */ -const FUNNEL_WIDTHS = [100, 76, 36, 36, 21, 11, 7]; - -/** - * The astrophysics audit as a funnel: papers go in at the top (black) and - * narrow to the few that were recomputed; errors come out (red) and narrow - * to the few anyone had challenged. Each stage's figure sits beside it. - */ -export function Funnel() { +/** One stack of papers in the tower: a block whose face is ruled into sheets. */ +function Block({ + x, + y, + w, + h, + className, + bad, +}: { + x: number; + y: number; + w: number; + h: number; + className?: string; + bad?: boolean; +}) { return ( -
-
    - {LEDGER.map((row, i) => { - const top = FUNNEL_WIDTHS[i]; - const bottom = FUNNEL_WIDTHS[i + 1]; - const opensGroup = i === 0 || LEDGER[i - 1].flagged !== row.flagged; - return ( -
  1. -
    - - - {/* A hairline of paper between one stage and the next. */} - {i > 0 && ( - - )} - -
    -
    - {opensGroup && ( -

    - {row.flagged ? 'Errors out' : 'Papers in'} -

    - )} -

    - - {row.value} - - {row.label} -

    -

    - {row.detail} -

    -
    -
  2. - ); - })} -
-
- The funnel is not drawn to scale. Figures from the audit’s own pages. - - -
-
+ ); } -/* How an error spreads ------------------------------------------------------ */ - /** - * How the audit's errors travelled, as one ratio: the later papers that - * repeated an error against the later papers that objected to one. No - * graphic; the ratio is the figure. + * Later work stands on earlier papers. A tower of paper stacks rests on two + * supports, one of them a single small paper nobody re-checked (red). Where + * motion is welcome that paper slips out and what stood on it comes down, + * on a slow loop (audit.css); otherwise the tower simply stands on it. */ -export function Spread() { - const { copyingPapers, objectingPapers, challenged, stillCopied, laterCopies } = SPREAD; - const ratio = (copyingPapers / objectingPapers).toFixed(1); +export function Tower() { return ( -
-
-

- {ratio} - : - 1 -

+
+ + + A tall tower of paper stacks resting partly on one small red paper. The red paper slips + out and the tower above it falls. + + + + + + + + + -
-

- For every later paper that objected to an error, more than five repeated one. -

-
-
-
- {copyingPapers} -
-
later papers repeated one of the audit’s errors.
-
-
-
- {objectingPapers} -
-
later papers objected to one.
-
-
- -
+ {/* The sound support. */} + -

- An objection did not end it. Of the {challenged} errors that were challenged in print,{' '} - {stillCopied} went on being copied, by {laterCopies} more papers. The audit read only the - citing papers with open text, so these counts are floors. - -

+ {/* Everything built on top: tips over the sound support's corner. */} + + + + + + + + + + + {/* Two loose stacks at the very top, thrown clear as it goes. */} + + + + {/* The one paper that does not hold. */} + + +
+ + Research is built on earlier papers. When one near the bottom is wrong, so is what stands on + it. +
); } diff --git a/app/audit/components/Hero.tsx b/app/audit/components/Hero.tsx index 45509a5aa..08f19e3de 100644 --- a/app/audit/components/Hero.tsx +++ b/app/audit/components/Hero.tsx @@ -4,7 +4,6 @@ import { BackButton } from '@/app/consciousness/components/BackButton'; import { AddToPoolButton, SubmitProposalButton } from './Actions'; import { HeroArt } from './HeroArt'; import { HeroCard } from './HeroCard'; -import { Ref } from './Parts'; import type { RfpView } from '../data/rfp'; /** @@ -16,7 +15,7 @@ const CARD_CTA = /** * The same hero as /consciousness, in this page's own style: back and home - * at the top left, the process link at the top right, the headline, one + * at the top left, a link to how the audit works at the top right, the headline, one * paragraph, and the card with the pool and both buttons. The figure is 3D * and sits behind: right of the text on wide screens, above it on phones. */ @@ -42,11 +41,11 @@ export function Hero({ rfp }: { rfp: RfpView }) { - Our process - Learn about our process + How it works + See how the audit works @@ -55,14 +54,12 @@ export function Hero({ rfp }: { rfp: RfpView }) { id="audit-title" className="a-rise a-serif max-w-[10.5em] text-[clamp(38px,4.9vw,72px)] leading-[0.98]" > - Fund the audit of everything science has published. + Fund the audit of biomedical research.

- A machine audit of one field of astrophysics reported 872 errors in 433 papers. Only 47 - had ever been challenged in print. - We are raising funds to audit more fields in the open, with experts checking - what the machines flag. + AI can now audit scientific papers at a fraction of the cost and time it once took. We are + raising funds to audit biomedical research, so that the results science builds on hold up.

@@ -72,11 +69,6 @@ export function Hero({ rfp }: { rfp: RfpView }) { - -

- A field as stacks of paper, one sheet per paper. Red sheets hold an error; paler ones repeat - it. An illustration, not data. -

); } diff --git a/app/audit/components/Parts.tsx b/app/audit/components/Parts.tsx index 127cd16c4..390c128a1 100644 --- a/app/audit/components/Parts.tsx +++ b/app/audit/components/Parts.tsx @@ -1,5 +1,6 @@ import type { ReactNode } from 'react'; import { cn } from '@/utils/styles'; +import { sourceNumber, type SourceKey } from '../data/content'; /** The page's column. On wide screens its text clears the margin rule. */ export function Wrap({ children, className }: { children: ReactNode; className?: string }) { @@ -15,8 +16,9 @@ export function Wrap({ children, className }: { children: ReactNode; className?: ); } -/** A footnote mark linking to its entry under Sources. */ -export function Ref({ n }: { n: number }) { +/** A footnote mark: the source's number, linking to its entry under Sources. */ +export function Ref({ to }: { to: SourceKey }) { + const n = sourceNumber(to); return ( {n} @@ -66,19 +68,20 @@ export function Section({ labelledBy: string; n?: number; label?: string; - /** `dark` sets the section white on black. */ - tone?: 'dark'; + /** `blue` sets the section white on the blue pencil's colour. */ + tone?: 'blue'; children: ReactNode; className?: string; }) { - const dark = tone === 'dark'; + const blue = tone === 'blue'; return (
@@ -87,18 +90,18 @@ export function Section({

{n} diff --git a/app/audit/components/Sections.tsx b/app/audit/components/Sections.tsx index c29f38383..d8ad8aec9 100644 --- a/app/audit/components/Sections.tsx +++ b/app/audit/components/Sections.tsx @@ -1,609 +1,136 @@ -import { ArrowRight, ArrowUpRight, CalendarDays, Plus } from 'lucide-react'; -import Link from 'next/link'; +import { ArrowRight, CalendarDays, Plus } from 'lucide-react'; import { Logo } from '@/components/ui/Logo'; -import { Avatar } from '@/components/ui/Avatar'; +import { Sequence } from '@/app/consciousness/components/Motion'; import { TRACK_RECORD } from '@/app/giving/lib/trackRecord'; -import { cn } from '@/utils/styles'; -import { - COST_FACTS, - FAQ, - LIMITS, - MODEL_COSTS, - OUTCOMES, - QUOTE, - RULES, - SCALE_FACTS, - SOURCES, - TRACKS, - US_RND_USD_BILLION, - type Fact as FactItem, -} from '../data/content'; -import { - CAMPAIGN, - TALK_TO_TEAM, - daysLeft, - deadlineLabel, - formatDate, - formatUsd, - formatUsdShort, - poolPercent, - type RfpView, -} from '../data/rfp'; -import { AddToPoolButton, ProposeLink, SubmitProposalButton } from './Actions'; -import { Funnel, Spread } from './Figures'; -import { OutcomeIcon } from './OutcomeIcons'; -import { BackerFaces, PoolBar } from './Pool'; -import { Ref, Section, SectionHead, Wrap } from './Parts'; +import { AUDIT_STEPS, BENEFITS, FAQ, PROBLEMS, SOURCES } from '../data/content'; +import { TALK_TO_TEAM, formatUsdShort } from '../data/rfp'; +import { BenefitIcon } from './BenefitIcons'; +import { Tower } from './Figures'; +import { Ref, Section, SectionHead } from './Parts'; +import { StepIcon } from './StepIcons'; -/* 1. Where this started ----------------------------------------------------- */ +/* 1. The problem ----------------------------------------------------------- */ /** - * What prompted the call: a post, quoted, then the audit it describes as a - * funnel (what went in, what came out, how little of it anyone had - * noticed). + * Why the call exists, readable at a glance: three short statements (each + * with a bar where there is a share to show, and its figure in small print) + * beside one picture of why it matters: later work stands on earlier papers. */ -export function Finding() { +export function Problem() { return ( -

- - An AI audit of astrophysics reported 872 errors. +
+ + Much of biomedical research does not hold up. -
); } -/* 2. Why errors survive --------------------------------------------------- */ - -/** - * What happened to the audit's errors after publication: how many later - * papers repeated one against how many objected, as one large ratio. - */ -export function Evidence() { - return ( -
- - Errors outlive their corrections. - -

- In the audit, peer review removed about one error in five before publication. Of the errors - that later papers picked up, over half were still being copied seven years on. - -

- -
- ); -} - -/* 3. The opportunity ------------------------------------------------------ */ - -/** Facts in a ruled list: the figure, then what it is. */ -function FactRows({ facts, className }: { facts: FactItem[]; className?: string }) { - return ( -
- {facts.map((f) => ( -
-
{f.value}
-
- {f.text} - {f.source != null && } -
-
- ))} -
- ); -} - -/** Facts side by side: a figure over its line. */ -function FactColumns({ facts }: { facts: FactItem[] }) { - return ( -
- {facts.map((f) => ( -
-
- {f.value} -
-
- {f.text} - {f.source != null && } -
-
- ))} -
- ); -} - -function CostChart() { - const max = Math.max(...MODEL_COSTS.map((m) => m.usdMillion)); - return ( -
-
By model, US$ million
-
    - {MODEL_COSTS.map((m) => ( -
  • - - {m.model} - {m.note && *} - - - - - {m.usdMillion.toFixed(1)} - - -
  • - ))} -
-

- Batch pricing, November 2026. *Promotional rate. Model costs only. Redrawn from Rust’s - estimate. - -

-
- ); -} - -/** A year of US research spending and the review, on one scale. */ -function ToScale() { - return ( -
-

- - US spending on research and development, 2024 (estimate) - - - - ${US_RND_USD_BILLION} billion - -

-
-

- - $6 million - - - The upper estimate, on the same scale. The red line is drawn over a hundred times too - wide. - -

-
- ); -} +/* 2. The proposal ---------------------------------------------------------- */ -/** - * What finding the errors could change, then what it would cost: the - * outcomes first (each with the fact it rests on), then one compact block - * with the estimate by model and the scale it sits against. - */ -export function Arithmetic() { +/** What the call proposes: the four steps of an audit, in order. */ +export function Proposal() { return ( -
- - What science gains when its errors are found. +
+ + Audit the open biomedical literature with AI. -
    - {OUTCOMES.map((o, i) => ( -
  1. -
    - -
    -

    - {String(i + 1).padStart(2, '0')} -

    -

    - {o.title} -

    -
    -
    -

    - {o.text} - {o.sources.map((n) => ( - - ))} -

    -
  2. - ))} -
- -
-

What it costs

-

- And it costs a rounding error on what science already spends. -

-
-
-

- - $0.6–6M - - - Estimated model cost to review 3 million open-access papers with AI. The range is - the choice of model. - - -

- -
-
- -
-
-
- -
-
- -
-
- -
-

- What tokens do not buy is the checking. -

-

- The Black Spatula Project ran AI checks on about 500 papers and has said its main - constraint is finding experts to verify what was flagged. - The pool pays for that: the people, the method, and getting corrections - into the record. -

-
-
- ); -} - -/* 4. The call ------------------------------------------------------------- */ - -export function Call({ rfp }: { rfp: RfpView }) { - return ( -
-
- - What the pool funds. - -

- Four kinds of work, funded on equal terms. Awards of {rfp.awardRangeLabel}, sized to the - job. -

-
-
    - {TRACKS.map((t, i) => ( -
  1. -
    -

    Track {i + 1}

    -

    - {t.name} +
      + {AUDIT_STEPS.map((step, i) => ( +
    1. + +
      + {i < AUDIT_STEPS.length - 1 && ( + + )} +

      + {String(i + 1).padStart(2, '0')} +

      +

      + {step.title}

      -
      -
      -

      {t.lead}

      -

      {t.body}

      -
      -
      -

      We look for

      -
        - {t.asks.map((a) => ( -
      • - - - - {a} -
      • - ))} -
      +

      {step.text}

    2. ))}
    -
    - -

); } -/* 5. Ground rules --------------------------------------------------------- */ +/* 3. The benefits ---------------------------------------------------------- */ -export function Rules() { +/** + * What the audit is for: four things it makes possible, on the blue of a + * check. As the section scrolls in, the four arrive one after another and + * each icon assembles itself (audit.css, `a-bn-*`); without script or with + * reduced motion they are simply there. + */ +export function Benefits() { return ( -
-
-
-

- “A flag is not a finding.” -

-

- The astrophysics audit’s own rule - , and this call’s. -

-
- -
    - {RULES.map((r, i) => ( -
  1. - - {String(i + 1).padStart(2, '0')} - -
    -

    {r.title}

    -

    - {r.text} +

    + + A new era of rigor in science. + + +
      + {BENEFITS.map((b, i) => ( +
    • + +
      +

      + {b.title} +

      +

      + {b.text}

    • ))} -
-
- -
-

How good are the machines, so far

-
    - {LIMITS.map((l) => ( -
  • -

    {l.value}

    -

    - {l.text} - -

    -
  • - ))}
-

- These tests differ in models, fields and what counts as an error, so they do not rank one - another. Together they say the same thing: results vary widely, and nobody has yet - measured precision for a whole-field audit with human experts. Track 2 pays for that. -

-
-
- ); -} - -/* 6. How the pool works ---------------------------------------------------- */ - -const STEPS = [ - { - title: 'You add to the pool', - text: 'Any amount, pooled with every other backer.', - }, - { - title: 'Researchers submit proposals', - text: 'Teams post their corpus, method and budget, and how they will confirm what they find.', - }, - { - title: 'Peers review, the pool pays', - text: 'Open review on ResearchHub. The strongest get funded.', - }, - { - title: 'Results are published', - text: 'Findings, data and code go public, including what did not hold up.', - }, -]; - -export function HowItWorks() { - return ( -
- - How the pool works - -
    - {STEPS.map((s, i) => ( -
  1. - {i < STEPS.length - 1 && ( - - )} -

    - {String(i + 1).padStart(2, '0')} -

    -

    {s.title}

    -

    {s.text}

    -
  2. - ))} -
-
- ); -} - -/* 7. Where it stands ------------------------------------------------------- */ - -function Fact({ label, children }: { label: string; children: React.ReactNode }) { - return ( -
-
{label}
-
{children}
-
- ); -} - -/** - * The programme's state: Phase 1 raises the pool (target, date, progress and - * what a donor checks), Phase 2 funds and runs the audits. The pool's - * backers and the way to the RFP's own page sit beside them. - */ -export function Standing({ rfp }: { rfp: RfpView }) { - const left = daysLeft(rfp.goalBy); - const more = rfp.backerCount - rfp.backers.length; - return ( -
- - Two phases - -
-
-

- Phase 1 - - Now - -

-

- Raise {formatUsd(rfp.goalUsd)} by {formatDate(rfp.goalBy)} -

-

- - - {formatUsd(rfp.raisedUsd)} - {' '} - raised - - - {Math.round(poolPercent(rfp))}% · {rfp.backerCount}{' '} - {rfp.backerCount === 1 ? 'backer' : 'backers'} · {left} {left === 1 ? 'day' : 'days'}{' '} - left - -

- -
- Only to work funded under this call. - - Everything raised still funds audits under this call; we fund fewer of them. - - {CAMPAIGN.taxDeductibleNote && ( - {CAMPAIGN.taxDeductibleNote} - )} -
-
- -
-
- -
-

- Phase 2 - - Next - -

-

- Pick proposals, run the audits -

-

- Reviewers score the proposals in the open. The pool funds the strongest, and what they - find is published. -

-
- - {deadlineLabel(rfp)} - {rfp.proposalCount > 0 && ` · ${rfp.proposalCount} so far`} - - In the open on ResearchHub, by peer reviewers - {rfp.awardRangeLabel} each -
-
- - {rfp.rfpHref && ( - - View the RFP and its proposals - - - )} -
-
-
- - {rfp.backers.length > 0 && ( -
-

Top backers

-
    - {rfp.backers.map((b) => ( -
  • - - {b.name} - {formatUsd(b.amountUsd)} -
  • - ))} -
- {more > 0 && ( -

- and {more} more {more === 1 ? 'backer' : 'backers'} -

- )} -
- )} +
); } @@ -639,10 +166,6 @@ export function AboutResearchHub() { discuss scientific research. Its mission is to accelerate the pace of scientific research.

-

- It is built for work that can be checked: study designs are registered in public before - the work starts, review happens in the open, and results are published with no paywall. -

{stats.map((s) => (
@@ -678,8 +201,8 @@ export function TalkToUs() { Talk to us

- Thinking about a larger gift, want to sponsor model time, or have a field you think - should go first? We would like to hear from you. + Thinking about a larger gift, want to give model credits instead of cash, or have an + area of biomedicine you think should go first? We would like to hear from you.

@@ -723,6 +246,9 @@ export function Faq() {

{f.a} + {f.sources?.map((key) => ( + + ))}

))} @@ -739,14 +265,14 @@ export function Sources() { Sources
    - {SOURCES.map((s) => ( + {SOURCES.map((s, i) => (
  1. - {s.id} + {i + 1}

    - Figures were checked against these sources on 18 November 2026. The astrophysics audit is + Figures were checked against these sources on 19 November 2026. The astrophysics audit is the work of its author and has not been independently verified; this page reports its claims and its stated limits. If we have something wrong, tell us and we will fix it.

    diff --git a/app/audit/components/OutcomeIcons.tsx b/app/audit/components/StepIcons.tsx similarity index 52% rename from app/audit/components/OutcomeIcons.tsx rename to app/audit/components/StepIcons.tsx index ffd2d48c8..c090b507f 100644 --- a/app/audit/components/OutcomeIcons.tsx +++ b/app/audit/components/StepIcons.tsx @@ -1,9 +1,8 @@ /** - * Drawn icons for the four outcomes under "What science gains when its - * errors are found", as /consciousness has for its four tenets, in this - * page's own hand: black line on paper, the red pencil for an error and the - * blue pencil for a check. Each is complete as the server renders it; where - * motion is welcome each then runs one small, slow loop (audit.css). + * Drawn icons for the four steps of an audit, in this page's own hand: black line on paper, the red pencil + * for an error and the blue pencil for a check. Each is complete as the + * server renders it; where motion is welcome each then runs one small, slow + * loop (audit.css). */ const INK = 'var(--a-ink)'; @@ -69,7 +68,7 @@ function Check({ d, className }: { d: string; className?: string }) { ); } -/** 1. Fixed before they spread: the error is corrected in the paper that +/** Published and corrected: the error is corrected in the paper that * made it, and the two papers that draw on it stay clean. */ function Fixed() { return ( @@ -128,80 +127,106 @@ function Fixed() { ); } -/** 2. Results that hold: new work, in black, set down on a stack that has - * been checked. */ -function Holds() { +/** A line of a paper picked out in the red pencil. */ +function Highlight({ x, y, w, muted }: { x: number; y: number; w: number; muted?: boolean }) { + return ( + + ); +} + +/** A small flag on a pole, in the margin beside a flagged line. */ +function Flag({ x, y }: { x: number; y: number }) { + return ( + + + + + ); +} + +/** It flags possible errors: two lines of a paper highlighted in red, each + * with a flag in the margin. */ +function Flagged() { return ( <> - - {[ - [13, 49], - [11, 41], - [14, 33], - [12, 25], - ].map(([x, y]) => ( - - ))} - - + + + + ); } -/** 3. Answered sooner: an open question, and the mark that it is settled. */ -function Answered() { +/** One of the experts: a head and shoulders. */ +function Person({ y }: { y: number }) { + return ( + + + + + ); +} + +/** Experts check each flag: the flagged paper beside three reviewers. One + * flag is confirmed (blue check), the other rejected (struck out). */ +function Experts() { return ( <> - - - - - - - - - + + + + + + + + ); } -/** 4. Checked as it appears: the red line passes over each new paper in - * turn, and each gets its check. */ +/** AI reads the papers: the red line passes over each paper in turn, + * and each gets its check. */ function Checked() { const lines: [number, number][] = [ [6, 8], @@ -230,16 +255,17 @@ function Checked() { ); } -const ICONS = [Fixed, Holds, Answered, Checked]; +/** In the order of the audit's steps (`AUDIT_STEPS`). */ +const ICONS = [Checked, Flagged, Experts, Fixed]; -export function OutcomeIcon({ index }: { index: number }) { +export function StepIcon({ index }: { index: number }) { const Art = ICONS[index]; if (!Art) return null; return ( diff --git a/app/audit/data/content.ts b/app/audit/data/content.ts index 884720368..5fc114edc 100644 --- a/app/audit/data/content.ts +++ b/app/audit/data/content.ts @@ -1,382 +1,162 @@ /** - * The page's words and figures: the audit's ledger, the evidence that errors - * persist, the cost estimates, what the call funds, its ground rules, the - * FAQ and the sources. Every figure is cited in `SOURCES` and was checked - * against that source on 18 November 2026. RFP facts (pool, deadline) are - * not here; they come from the live RFP (`rfp.ts`). + * The page's words and figures: the problem in biomedicine, the audit the + * call funds, what it makes possible, the FAQ and the sources. Every figure + * is cited in `SOURCES` and was checked against that source on 18–19 + * November 2026. RFP facts (pool, deadline) are not here; they come from + * the live RFP (`rfp.ts`). * - * The call's scope (`TRACKS`), its rules (`RULES`) and the FAQ are a DRAFT - * until the call's terms are set. + * The call's scope (`AUDIT_STEPS`) and the FAQ are a DRAFT until the call's + * terms are set. */ export interface Source { - id: number; + key: string; label: string; url: string; } -export const SOURCES: Source[] = [ +/** In the order they are numbered on the page. Cite one with ``. */ +export const SOURCES = [ { - id: 1, - label: 'Handmer, “A field of physics, audited by machine”, 2026', - url: 'https://science.caseyhandmer.com/physics/', - }, - { - id: 2, - label: 'Handmer, “Method and limits: how to audit a field by machine”, 2026', - url: 'https://science.caseyhandmer.com/physics/method.html', - }, - { - id: 3, - label: 'Handmer, “872 verified errors in 433 papers”, 2026 (the audit’s full list)', - url: 'https://science.caseyhandmer.com/physics/errors.html', - }, - { - id: 4, - label: 'Handmer, post announcing the audit, X, 17 November 2026', - url: 'https://x.com/CJHandmer/status/2122609626588061816', - }, - { - id: 5, - label: 'Handmer, post on the cost of auditing every field, X, 17 November 2026', - url: 'https://x.com/CJHandmer/status/2122796358616891642', - }, - { - id: 6, + key: 'rpcb', label: - 'Rust, post estimating the cost of AI review of the open-access literature, X, 16 November 2026', - url: 'https://x.com/rust_ruslan/status/2122444038091718701', + 'Errington et al., “Investigating the replicability of preclinical cancer biology”, eLife, 2021', + url: 'https://doi.org/10.7554/eLife.71601', }, { - id: 7, + key: 'reviewers', label: - 'National Center for Science and Engineering Statistics, NSF 26-314, 2026 (US R&D: $993 billion estimated for 2024)', - url: 'https://ncses.nsf.gov/pubs/nsf26314', + 'Schroter et al., “What errors do peer reviewers detect, and does training improve their ability to detect them?”, J. R. Soc. Med., 2008', + url: 'https://doi.org/10.1258/jrsm.2008.080062', }, { - id: 8, + key: 'freedman', label: - 'National Library of Medicine, PMC Open Access Subset (8,402,749 records on 18 November 2026)', - url: 'https://pmc.ncbi.nlm.nih.gov/tools/openftlist/', + 'Freedman, Cockburn and Simcoe, “The economics of reproducibility in preclinical research”, PLoS Biology, 2015', + url: 'https://doi.org/10.1371/journal.pbio.1002165', }, { - id: 9, - label: - 'National Science Board, “The State of U.S. Science and Engineering 2026” (3.5 million articles in 2024)', - url: 'https://ncses.nsf.gov/pubs/nsbsep20261', - }, - { - id: 10, + key: 'spot', label: 'Son et al., “When AI co-scientists fail: SPOT, a benchmark for automated verification of scientific research”, 2025', url: 'https://arxiv.org/abs/2505.11855', }, { - id: 11, + key: 'bianchi', label: 'Bianchi et al., “To err is human: systematic quantification of errors in published AI papers via LLM analysis”, 2025', url: 'https://arxiv.org/abs/2512.05925', }, { - id: 12, - label: - 'Zubeldia et al., “Auditing AI peer reviewers: a benchmark on real scientific papers”, 2026', - url: 'https://arxiv.org/abs/2611.17052', - }, - { - id: 13, - label: - 'The Black Spatula Project; and coverage of its verification bottleneck, heise online, 2025', - url: 'https://www.heise.de/en/news/No-more-calculation-errors-AI-checks-papers-for-errors-Why-that-s-not-enough-10321469.html', - }, -]; - -/** - * The post that prompted the call (source 4), quoted briefly. Its "nearly - * 1000 major" rounds up: the audit's pages count 872 upheld errors, 92 of - * them graded major, and the page says so beside the quote. - */ -export const QUOTE = { - text: 'We found nearly 1000 major errors that got cited many times…', - author: 'Casey Handmer', - where: 'on X, 17 November 2026', - url: 'https://x.com/CJHandmer/status/2122609626588061816', -}; - -/** The astrophysics audit, top to bottom, as a funnel. Figures from sources 1–3. */ -export interface LedgerRow { - value: string; - label: string; - detail: string; - /** An error row: drawn in the red pencil. */ - flagged?: boolean; -} - -export const LEDGER: LedgerRow[] = [ - { - value: '158,059', - label: 'papers mapped', - detail: - 'The papers of one Caltech research group, the papers they cite and the papers that cite them.', - }, - { - value: '50,269', - label: 'papers read in full', - detail: 'Every paper whose full text is openly available: 421 million words.', - }, - { - value: '659', - label: 'papers recomputed', - detail: - 'The most cited first. A model re-derived the equations and rebuilt the numbers, without seeing errata or later papers.', - }, - { - value: '872', - label: 'errors reported', - detail: - 'Spread across 433 papers. Each one was re-derived by a second, independent AI session before it counted. The audit grades 92 as major.', - flagged: true, + key: 'method', + label: 'Handmer, “Method and limits: how to audit a field by machine”, 2026', + url: 'https://science.caseyhandmer.com/physics/method.html', }, { - value: '159', - label: 'errors repeated by later papers', - detail: 'At least one later paper carried the error forward. In total, 367 later papers did.', - flagged: true, + key: 'audit', + label: 'Handmer, “A field of physics, audited by machine”, 2026', + url: 'https://science.caseyhandmer.com/physics/', }, { - value: '47', - label: 'errors criticised by a later paper', - detail: - 'Only 47 of the 872 errors were ever challenged in print. For 21 of them, the copying continued after the criticism.', - flagged: true, + key: 'errors', + label: 'Handmer, “872 verified errors in 433 papers”, 2026 (the audit’s full list)', + url: 'https://science.caseyhandmer.com/physics/errors.html', }, -]; - -/** How the audit's errors travelled into later papers (source 1). */ -export const SPREAD = { - copyingPapers: 367, - objectingPapers: 68, - challenged: 47, - stillCopied: 21, - laterCopies: 51, -}; - -/** - * Estimated model cost of reviewing 3 million papers, US$ million, by model. - * Redrawn from Rust's chart (source 6): batch pricing, November 2026. - */ -export const MODEL_COSTS: { model: string; usdMillion: number; note?: string }[] = [ - { model: 'GPT-6 Astra', usdMillion: 6.0 }, - { model: 'Claude Fable 5.2', usdMillion: 4.8 }, - { model: 'Claude Opus 5.5', usdMillion: 2.4 }, - { model: 'GPT-6.1 Sol', usdMillion: 1.2 }, - { model: 'Claude Sonnet 5.5', usdMillion: 1.2 }, - { model: 'Gemini 4 Carbon', usdMillion: 1.2, note: 'promotional rate' }, - { model: 'GPT-6.1 Terra', usdMillion: 0.6 }, -]; +] as const satisfies readonly Source[]; -export const US_RND_USD_BILLION = 993; +export type SourceKey = (typeof SOURCES)[number]['key']; -/** A stated fact: a figure and the one line that says what it is. */ -export interface Fact { - value: string; - text: string; - /** Omitted where the figure is this page's own arithmetic. */ - source?: number; +/** The number a source carries on the page. */ +export function sourceNumber(key: SourceKey): number { + return SOURCES.findIndex((s) => s.key === key) + 1; } -/** What auditing costs. The headline figure is Rust's estimate (source 6). */ -export const COST_FACTS: Fact[] = [ - { - value: '$0.20–$2', - text: 'per paper, at those rates. Our arithmetic from the estimate above.', - }, - { - value: '$2,720', - text: 'Model cost of the whole astrophysics audit at list price, or about $3 for each error found.', - source: 1, - }, - { - value: '$100,000', - text: 'Roughly what Handmer says the tokens would cost to sample every field of science.', - source: 5, - }, -]; - -/** How large the literature is, beside the money already spent on research. */ -export const SCALE_FACTS: Fact[] = [ - { - value: '8.4 million', - text: 'open-access papers in PubMed Central alone.', - source: 8, - }, - { - value: '3.5 million', - text: 'new science and engineering articles published in 2024.', - source: 9, - }, - { - value: '< 0.001%', - text: 'of one year of US research spending: what the upper estimate comes to.', - }, -]; - /** - * What finding the errors could change. Each is a possibility, stated with - * the fact it rests on. + * The problem, three ways, each readable at a glance: a short statement, a + * bar where there is a share to show, and the figure behind it in small + * print. */ -export const OUTCOMES: { title: string; text: string; sources: number[] }[] = [ - { - title: 'Errors are fixed before they spread.', - text: 'Today more than five later papers repeat an error for each one that objects. An error flagged at its source can be corrected before the next paper cites it.', - sources: [1], - }, +export interface Problem { + title: string; + /** The share that held up (0–1), drawn as a bar against the share that did not. */ + split?: { share: number; good: string; bad: string }; + note: string; + source: SourceKey; +} + +export const PROBLEMS: Problem[] = [ { - title: 'New work stands on results that hold.', - text: 'The audit estimates 34,000 errors (25,000 to 48,000) still undetected in one field’s most cited papers, and it has recomputed only 4% of them. Each one found is a result nobody builds on by mistake.', - sources: [1, 2], + title: 'Fewer than half of results replicate.', + split: { share: 0.46, good: 'Replicated', bad: 'Did not' }, + note: '46% in a major replication project in cancer biology.', + source: 'rpcb', }, { - title: 'Open questions are answered sooner.', - text: 'The audit listed 155 open questions in its field. A model judged 8 answerable now from existing data and code, and 119 that AI would speed up. Nobody has tried yet.', - sources: [1], + title: 'Peer reviewers miss most errors.', + split: { share: 2.58 / 9, good: 'Caught', bad: 'Missed' }, + note: 'Reviewers caught 2.6 of 9 planted errors, on average.', + source: 'reviewers', }, { - title: 'Every new paper is checked as it appears.', - text: 'About 3.5 million articles are published a year. Rust proposes running the same review on every new preprint, so that checking keeps pace with publishing.', - sources: [9, 6], + title: '$28 billion a year funds work that cannot be reproduced.', + note: 'In the US alone, by one estimate with a wide range.', + source: 'freedman', }, ]; -/** What the pool pays for. DRAFT scope. */ -export interface Track { - id: string; - name: string; - lead: string; - body: string; - asks: string[]; +/** The audit the pool funds, step by step. DRAFT scope. */ +export interface AuditStep { + title: string; + text: string; } -export const TRACKS: Track[] = [ - { - id: 'field', - name: 'Audit a field', - lead: 'Take one field and read all of it.', - body: 'Map a field’s literature, screen what is open, and recompute its most relied-upon papers. Astrophysics is one field, mostly theory, around one group. Nobody knows yet what the same audit finds in cell biology, economics or clinical trials.', - asks: [ - 'A defined corpus and how it is drawn', - 'Every error re-derived independently before it counts', - 'The full list published, with each derivation', - ], - }, - { - id: 'verify', - name: 'Check the checker', - lead: 'Put people on the flags.', - body: 'No field-scale audit has had its findings adjudicated by experts. Fund the reading that turns a model’s claim into a result: specialists confirm or reject a sample of flags, and the precision is published, whatever it is.', - asks: [ - 'Named domain experts and a blind protocol', - 'A sample large enough to give a real error bar', - 'The rejected flags published beside the upheld ones', - ], - }, - { - id: 'correct', - name: 'Carry the correction', - lead: 'Follow each error to where it went.', - body: 'An error matters most in the papers that built on it. Trace confirmed errors through their citations, tell the authors first, and take corrections to journals and preprint servers so that the record changes, not only a list.', - asks: [ - 'Authors contacted before anything is public', - 'A right of reply, published with the finding', - 'Corrections filed and their outcomes tracked', - ], - }, - { - id: 'method', - name: 'Build it in the open', - lead: 'Make the next audit cheaper and better.', - body: 'Pipelines, prompts and benchmarks anyone can rerun: tests on real, author-confirmed errors; false-positive rates on clean papers; and standing audits that read new preprints as they are posted.', - asks: [ - 'Open code, prompts and data under an open licence', - 'Measured recall and false-positive rates', - 'A method that works on a field its authors did not build it for', - ], - }, +export const AUDIT_STEPS: AuditStep[] = [ + { title: 'AI reads the papers', text: 'Every open paper in one field.' }, + { title: 'It flags possible errors', text: 'Statistics, arithmetic, figures, data.' }, + { title: 'Experts check each flag', text: 'Specialists confirm or reject it.' }, + { title: 'Errors get corrected', text: 'Authors are told. Results are public.' }, ]; -/** Conditions on funded work. DRAFT. */ -export const RULES: { title: string; text: string }[] = [ - { - title: 'Two derivations, not one', - text: 'An error counts only when a second, independent pass reaches it without seeing the first.', - }, - { - title: 'People before publication', - text: 'A named paper is called wrong only after a qualified person has checked the claim and its authors have had the chance to answer.', - }, - { - title: 'Publish the misses', - text: 'Every audit reports its false positives and what it could not read, alongside what it found.', - }, - { - title: 'Errors, not accusations', - text: 'The subject is mistakes in honest work. The astrophysics screen looked for paper-mill language and chatbot residue and found none.', - }, - { - title: 'Open by default', - text: 'Findings, data, prompts and code are published on ResearchHub under an open licence.', - }, -]; +/** What the audit is for: four things it makes possible. */ +export interface Benefit { + title: string; + text: string; +} -/** What is known about how well models find real errors. */ -export const LIMITS: { value: string; text: string; source: number }[] = [ - { - value: '21%', - text: 'The best model’s recall on 91 real, author-confirmed errors in 2025. Its precision was 6%.', - source: 10, - }, - { - value: '83%', - text: 'Of 316 mistakes a checker flagged in published AI papers, the share human experts confirmed.', - source: 11, - }, - { - value: '12–71%', - text: 'Detection of errors injected into astrophysics preprints across twelve AI reviewers, with false alarms on under 9% of clean copies.', - source: 12, - }, - { - value: '33 of 40', - text: 'Errors from the astrophysics audit upheld when a third session re-checked them blind. No person has checked any yet.', - source: 2, - }, +export const BENEFITS: Benefit[] = [ + { title: 'Science that holds up', text: 'Researchers build on results that have been checked.' }, + { title: 'Funding that goes further', text: 'Money goes to work that stands, not to dead ends.' }, + { title: 'Treatments, sooner', text: 'Fewer years lost chasing results that were never right.' }, + { title: 'Smarter AI', text: 'Models learn from a literature that has been corrected.' }, ]; export interface Faq { q: string; a: string; + sources?: SourceKey[]; } export const FAQ: Faq[] = [ { q: 'What makes a strong proposal?', - a: 'It names its corpus and how it is drawn, says how an error will be confirmed before it counts, and commits in advance to reporting its false-positive rate. It includes people qualified to judge the field, and it publishes everything, including the flags that did not hold up.', + a: 'It names its corpus and how it is drawn, says how an error will be confirmed before it counts, and commits in advance to reporting its false-positive rate. It includes people qualified to judge the field.', }, { - q: 'Which fields are in scope?', - a: 'Any field where a claim can be checked against the paper itself: derivations, statistics, units, tables, code and data. Quantitative fields are the obvious start, but a well-argued proposal in any discipline is welcome.', + q: 'Does an audit show whether a result replicates?', + a: 'No. An audit does not rerun the experiment. It checks what the paper already shows: whether the statistics, arithmetic, figures and data agree with each other and with the conclusions. Those errors are one reason results fail to hold up, and they are the part a machine can check at scale.', }, { - q: 'Does an audit need people, or only models?', - a: 'Both. Models make reading a whole field affordable; they also flag things that are not errors. Proposals that include expert adjudication, or that are only expert adjudication of an existing audit, are in scope.', + q: 'How accurate is AI at finding errors?', + a: 'It varies a great deal. On one 2025 test of real, author-confirmed errors the best model found 21%. In another, human experts confirmed 83% of the mistakes a checker flagged. In the astrophysics audit a blind re-check by a third AI session agreed with 33 of 40 errors, and no person has checked them yet. That is why this call pays for expert checking as well as model time.', + sources: ['spot', 'bianchi', 'method'], }, { q: 'Will papers and authors be named?', - a: 'Only for errors a qualified person has confirmed, and after the authors have been told and given the chance to reply. The call funds work on honest mistakes, not allegations of misconduct.', + a: 'Only for errors a qualified person has confirmed, and after the authors have been told and given the chance to reply.', }, { - q: 'What about papers behind paywalls?', - a: 'Audits work from text they are licensed to read: open-access articles and preprints. The astrophysics audit could not read about a quarter of the papers on its map, and said so. Proposals should report their own coverage the same way.', + q: 'Why biomedicine, and what about other fields?', + a: 'It has an open literature large enough to check at scale, the best-documented record of results that fail to hold up, and errors that can end up reaching patients. Other fields may get calls of their own.', }, { q: 'Who can apply?', @@ -387,11 +167,12 @@ export const FAQ: Faq[] = [ a: 'Proposals are published on ResearchHub and peer reviewed in the open. The ResearchHub Foundation picks which to fund, guided by those reviews.', }, { - q: 'Where does money added to the pool go?', - a: 'Only to proposals funded under this call. The pool is the call’s whole budget, built by everyone who adds to it.', + q: 'Can we give model credits instead of cash?', + a: 'Yes. Model time is a real cost of every audit. If you run an AI lab or a compute provider and would rather give credits, book a call with us.', }, { - q: 'Is this call run with Casey Handmer or Ruslan Rust?', - a: 'No. Their public posts prompted it and are cited as sources. Neither is affiliated with this call or has endorsed it.', + q: 'What inspired this call?', + a: 'Casey Handmer’s AI audit of theoretical astrophysics, published in November 2026. It read 50,269 papers in full and reported 872 errors in 433 of them, for about $2,720 in model cost. No person has checked its findings yet, and he is not involved in this call.', + sources: ['audit', 'errors'], }, ]; diff --git a/app/audit/data/rfp.config.json b/app/audit/data/rfp.config.json index 5d2c7d358..2b70898fe 100644 --- a/app/audit/data/rfp.config.json +++ b/app/audit/data/rfp.config.json @@ -1,3 +1,3 @@ { - "productionPostId": null + "productionPostId": 33884 } diff --git a/app/audit/data/rfp.ts b/app/audit/data/rfp.ts index aac74b789..44e27292b 100644 --- a/app/audit/data/rfp.ts +++ b/app/audit/data/rfp.ts @@ -8,8 +8,8 @@ import config from './rfp.config.json'; * (`app/consciousness/data/rfp.ts`), with its own post and terms. * * Post ids differ between environments, so the id is picked per environment: - * - production deploys use `productionPostId` in `rfp.config.json` (null - * until the RFP is posted); + * - production deploys use `productionPostId` in `rfp.config.json`, which + * `next.config.js` also reads to send /rfp// here; * - anywhere else, `AUDIT_RFP_POST_ID` names one, and `?rfp=` * previews any. * With none, the page shows an empty pool and its buttons only explain. @@ -21,16 +21,16 @@ export function rfpPostId(preview?: string | string[]): number | null { } const fromEnv = Number(process.env.AUDIT_RFP_POST_ID); if (Number.isInteger(fromEnv) && fromEnv > 0) return fromEnv; - const fromConfig: number | null = config.productionPostId; - return isProductionDeploy ? fromConfig : null; + return isProductionDeploy ? config.productionPostId : null; } /** * The campaign's own terms, which the grant does not hold. * * DRAFT: every value here is a placeholder until the call's terms are set. - * `taxDeductibleLine` stays null unless gifts to this pool are confirmed to - * be deductible; the page then shows no tax wording at all. + * `taxDeductibleLine` is the short line in the hero card, worded as on + * /consciousness (Kobe, 2026-11-20). `taxDeductibleNote`, the longer note + * that names the fund, stays null until that is confirmed for this pool. */ export const CAMPAIGN: { goalUsd: number; @@ -42,7 +42,7 @@ export const CAMPAIGN: { goalUsd: 250_000, goalBy: '2027-01-31', awardRangeLabel: '$5k – $100k', - taxDeductibleLine: null, + taxDeductibleLine: 'Contributions are tax deductible', taxDeductibleNote: null, }; @@ -62,7 +62,7 @@ export type RfpStatus = 'OPEN' | 'PENDING' | 'CLOSED'; export interface RfpView { /** Null when no RFP is configured or it failed to load. */ grantId: number | null; - /** The RFP's own page: its full proposal list and details. */ + /** The RFP's own page, kept reachable with `?view=rfp`. */ rfpHref: string | null; raisedUsd: number; goalUsd: number; @@ -91,7 +91,7 @@ export function rfpView(work: Work | null): RfpView { if (grant && grant.status !== 'PENDING') status = isGrantActive(grant) ? 'OPEN' : 'CLOSED'; return { grantId: grant?.id != null ? Number(grant.id) : null, - rfpHref: work && grant ? `/rfp/${work.id}/${work.slug}` : null, + rfpHref: work && grant ? `/rfp/${work.id}/${work.slug}?view=rfp` : null, raisedUsd: pool?.amountRaised.usd ?? 0, goalUsd: CAMPAIGN.goalUsd, goalBy: CAMPAIGN.goalBy, diff --git a/app/audit/layout.tsx b/app/audit/layout.tsx index d53472ea3..6c39a0205 100644 --- a/app/audit/layout.tsx +++ b/app/audit/layout.tsx @@ -23,13 +23,13 @@ const mono = IBM_Plex_Mono({ export const metadata: Metadata = { ...buildOpenGraphMetadata({ - title: 'Fund the audit of everything science has published.', + title: 'Fund the audit of biomedical research.', description: - 'A machine audit of one field of astrophysics reported 872 errors in 433 papers; 47 had ever been challenged in print. We are raising a pool to audit more fields in the open, with experts checking what the machines flag.', + 'AI can now audit scientific papers at a fraction of the cost and time it once took. We are raising funds to audit biomedical research, so that the results science builds on hold up.', url: '/audit', }), title: { - default: 'Auditing the scientific literature: request for proposals', + default: 'Auditing biomedical research: request for proposals', template: `%s | ${SITE_CONFIG.name}`, }, // Not indexed until the call's terms are final (see CAMPAIGN in data/rfp.ts). diff --git a/app/audit/page.tsx b/app/audit/page.tsx index 5880de7a6..ff4234391 100644 --- a/app/audit/page.tsx +++ b/app/audit/page.tsx @@ -2,16 +2,12 @@ import { ActionDock } from './components/ActionDock'; import { Hero } from './components/Hero'; import { AboutResearchHub, - Arithmetic, AuditFooter, - Call, - Evidence, + Benefits, Faq, - Finding, - HowItWorks, - Rules, + Problem, + Proposal, Sources, - Standing, TalkToUs, } from './components/Sections'; import { rfpPostId, rfpView } from './data/rfp'; @@ -39,12 +35,12 @@ async function loadRfp(preview: string | string[] | undefined) { /** * The landing page for the literature-audit request for proposals: fund AI - * audits of the published literature, field by field, with experts checking - * what the machines flag. It shares its hero (pool, backers and both - * buttons, from the live RFP) with /consciousness and nothing else: the - * style is a marked-up proof, and the sections tell the audit's own story. - * All text is server-rendered; the hero's 3D figure starts once its script - * loads. + * audits of the open biomedical literature, with experts confirming what + * the models flag. It shares its hero (pool, backers and both buttons, from + * the live RFP) with /consciousness and nothing else: the style is a + * marked-up proof, and the page is three short sections (the problem, the + * proposal, its benefits). All text is server-rendered; + * the hero's 3D figure starts once its script loads. */ export default async function AuditPage({ searchParams, @@ -58,13 +54,9 @@ export default async function AuditPage({ <>
    - - - - - - - + + + diff --git a/next.config.js b/next.config.js index b03fc57e6..927652212 100644 --- a/next.config.js +++ b/next.config.js @@ -16,6 +16,20 @@ const CONSCIOUSNESS_RFP_POST_ID = process.env.CONSCIOUSNESS_RFP_POST_ID || (process.env.VERCEL_ENV === 'production' ? String(consciousnessRfp.productionPostId) : null); +// eslint-disable-next-line @typescript-eslint/no-require-imports -- next.config.js is CommonJS +const auditRfp = require('./app/audit/data/rfp.config.json'); + +/** + * The RFP post that the /audit landing page stands in for, picked the same + * way: `rfp.config.json` on production deploys, `AUDIT_RFP_POST_ID` anywhere + * else. Must match `rfpPostId` in `app/audit/data/rfp.ts`. + */ +const AUDIT_RFP_POST_ID = + process.env.AUDIT_RFP_POST_ID || + (process.env.VERCEL_ENV === 'production' && auditRfp.productionPostId != null + ? String(auditRfp.productionPostId) + : null); + /** @type {import('next').NextConfig} */ const nextConfig = { reactStrictMode: false, @@ -32,6 +46,17 @@ const nextConfig = { }, ] : []), + // The audit RFP's own page sends visitors to its landing page, the same way. + ...(AUDIT_RFP_POST_ID + ? [ + { + source: `/rfp/${AUDIT_RFP_POST_ID}/:path*`, + missing: [{ type: 'query', key: 'view' }], + destination: '/audit', + permanent: false, + }, + ] + : []), // The page's first address, from its preview (PR #1298). { source: '/orch-or/figures',