Files
PIG-Demo/src/components/demo/BlindCompare.tsx
T
karti-ai 408ce4a525 Capture harness, fixture verification, CI, and the public README
The site does no live inference. Rollouts are captured once against spark-1 and
replayed at their recorded wall-clock — a public demo with no auth cannot hold
an API key, and a recorded run can be scrubbed, permalinked, blind-compared and
verified in ways a live one cannot. What stops it being a video is that the
browser re-derives every number from the recorded moves.

verify_fixtures.py is the Python half of that: it replays every committed
fixture through the engine and reproduces its own rewards. All 16 land at
delta 0.0. A fixture that cannot be regenerated is a claim with no receipt.

First real measurement, thinking off, 8 seeds: solved 0/8. The model repeats
guesses it has already played, invents words (trape, slith, postt, boomy),
and contradicts its own feedback — consistency 0.09 to 0.17. That is the
published failure taxonomy showing up in our own data on the first run, and it
is why `consistency` is a reward component rather than a footnote.

A capture failure is recorded as a turn with a null reply, never dropped. A
capture that silently discarded failed turns would be reporting a better model
than the one that ran.

CI gates both halves and four things that fail silently in production: the word
lists must rebuild byte-identically, the prerendered routes must carry their own
baked og tags (crawlers do not run JS, so without them every shared link
previews as the homepage), no blob: URL may reach the bundle (the site's CSP has
no worker-src, so it falls back to default-src 'self' and a blob worker is
blocked with no error), and the conformance digest must match across languages.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019mt6sHQHEnEYrJZvoMCJSB
2026-08-28 15:47:31 -07:00

193 lines
7.0 KiB
TypeScript

import { useState } from 'react';
import type { ComponentType } from 'react';
import { Eye, Trophy } from 'lucide-react';
import type { DemoStep } from '@/lib/demo-kit/types';
import { cn } from '@/lib/utils';
import { formatOrDash } from './format';
export interface BlindCompareRun<T> {
runId: string;
/** The identity, revealed only after the visitor commits. */
label: string;
model: string;
/** Required by the contract on an `intervened` run; shown at reveal. */
intervention?: string;
steps: DemoStep<T>[];
total: number | null;
}
export type BlindVote = 'A' | 'B' | 'tie';
export interface BlindCompareProps<T> {
/** Both runs must be the same seed or the comparison is meaningless. */
seed: number;
a: BlindCompareRun<T>;
b: BlindCompareRun<T>;
Surface: ComponentType<{ state: T; compact?: boolean }>;
question?: string;
onVote?: (vote: BlindVote) => void;
className?: string;
}
/**
* Two runs on the same seed, unlabelled, until you commit.
*
* The point is not the vote. The point is that the visitor forms an opinion
* from the behaviour BEFORE they learn which one had the better prompt, the
* bigger model or the training run — because once they know, they cannot
* unknow it, and every "obviously the trained one looks better" is worthless
* after the fact.
*
* There is a "reveal without voting" escape on purpose: a visitor who does not
* want to play should not be held hostage by a modal-shaped page.
*/
export function BlindCompare<T>({
seed,
a,
b,
Surface,
question = 'Which agent would you rather have running this?',
onVote,
className,
}: BlindCompareProps<T>) {
const [vote, setVote] = useState<BlindVote | null>(null);
const [revealed, setRevealed] = useState(false);
const commit = (choice: BlindVote) => {
setVote(choice);
setRevealed(true);
onVote?.(choice);
};
const winner: 'A' | 'B' | 'tie' =
a.total === null || b.total === null
? 'tie'
: a.total > b.total
? 'A'
: b.total > a.total
? 'B'
: 'tie';
const sides: { id: 'A' | 'B'; run: BlindCompareRun<T> }[] = [
{ id: 'A', run: a },
{ id: 'B', run: b },
];
return (
<section aria-label="Blind comparison" className={cn('space-y-3', className)}>
<div className="flex flex-wrap items-baseline gap-x-3 gap-y-1">
<h3 className="text-sm font-semibold">{question}</h3>
<p className="nums text-xs text-muted">Same puzzle, same seed ({seed}).</p>
</div>
<div className="grid gap-3 sm:grid-cols-2">
{sides.map(({ id, run }) => {
const last = run.steps[run.steps.length - 1];
const picked = vote === id;
return (
<div
key={id}
className={cn(
'card flex flex-col gap-3 p-3 transition-colors duration-2 ease-enter',
picked && 'border-brand',
revealed && winner === id && 'border-positive',
)}
>
<div className="flex items-center justify-between gap-2">
<h4 className="text-sm font-semibold">Agent {id}</h4>
<span className="nums text-xs text-muted">{run.steps.length} steps</span>
</div>
<div className="rounded-lg bg-surface-2 p-3">
{last ? (
<Surface state={last.state} />
) : (
<p className="text-sm text-muted">This run recorded no steps.</p>
)}
</div>
{!revealed ? (
<button
type="button"
onClick={() => commit(id)}
className="tap w-full rounded-lg bg-primary px-3 text-sm font-medium text-primary-foreground transition-colors duration-2 ease-enter hover:bg-primary/90"
>
Agent {id} is better
</button>
) : (
<dl className="space-y-1 text-sm">
<div className="flex items-baseline justify-between gap-2">
<dt className="text-muted">Identity</dt>
<dd className="text-right font-medium">{run.label}</dd>
</div>
<div className="flex items-baseline justify-between gap-2">
<dt className="text-muted">Model</dt>
<dd className="nums text-right font-mono text-xs">{run.model}</dd>
</div>
<div className="flex items-baseline justify-between gap-2">
<dt className="text-muted">Intervention</dt>
<dd className="text-right text-xs">
{run.intervention ?? (
<span className="text-muted">none plain rollout</span>
)}
</dd>
</div>
<div className="flex items-baseline justify-between gap-2 border-t border-border pt-1">
<dt className="text-muted">Reward</dt>
<dd className="nums flex items-center gap-1 text-right font-mono font-semibold">
{revealed && winner === id ? (
<Trophy className="h-3.5 w-3.5 text-positive" aria-hidden="true" />
) : null}
{formatOrDash(run.total)}
</dd>
</div>
</dl>
)}
</div>
);
})}
</div>
{!revealed ? (
<div className="flex flex-wrap gap-2">
<button
type="button"
onClick={() => commit('tie')}
className="tap rounded-lg border border-border px-3 text-sm font-medium transition-colors duration-2 ease-enter hover:bg-surface-2"
>
Too close to call
</button>
<button
type="button"
onClick={() => setRevealed(true)}
className="tap inline-flex items-center gap-1.5 rounded-lg px-3 text-sm font-medium text-muted transition-colors duration-2 ease-enter hover:text-fg"
>
<Eye className="h-4 w-4" aria-hidden="true" />
Just show me
</button>
</div>
) : (
<p role="status" className="card bg-surface-2 p-3 text-sm leading-relaxed">
{vote === null ? (
<>Revealed without a vote. </>
) : vote === winner ? (
<>
<span className="font-semibold text-positive">You picked the higher-scoring run.</span>{' '}
</>
) : (
<>
<span className="font-semibold text-warning">
You picked the lower-scoring run.
</span>{' '}
</>
)}
The environment scored these two with the same grader, on the same seed. The difference
between them is stated above and if it is a prompt change rather than a training run,
it says so, because a prompt change presented as a training result is the oldest trick
in this business.
</p>
)}
</section>
);
}