Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.
Python agentic-aiai-securityalert-triageblue-teamcybersecuritydetection-engineeringincident-responselanggraphllm-agentsmcpmitre-attackmodel-context-protocolpurple-teamsecurity-operationsself-hostedsiemsoarsocthreat-detectionthreat-intelligence
'use client';
/**
* "Benchmarked, not vibes" — `benchmark` section from §6.9 of the
* brief.
*
* Renders a three-up band of `NumberTicker`-driven metrics over a
* brand-tinted gradient. Each tile shows:
* - A big mono metric (animated count-up from 0 when the band enters
* the viewport; suppressed under `prefers-reduced-motion`).
* - A caption labelling the metric source (substrate / wet-eval) so
* the page never claims a "real" number for a substrate gate. This
* matches the v1.4 eval-harness honesty contract from AGENTS.md.
*
* Two CTAs sit below the tiles: read the methodology, open the public
* scoreboard. Both deep-link to /benchmark routes that already exist.
*/
import Link from 'next/link';
import { motion, useReducedMotion } from 'framer-motion';
import { ArrowRight, ExternalLink, GitCompare } from 'lucide-react';
import { NumberTicker } from '@/components/magicui/NumberTicker';
import { docs } from '@/lib/docs';
interface Metric {
/**
* If `value` is set, the tile animates a 0→value count-up. Otherwise
* `label` is the visible metric (e.g. "Sub-minute p50").
*/
value?: number;
suffix?: string;
label: string;
caption: string;
}
const METRICS: ReadonlyArray<Metric> = [
{
value: 97,
suffix: '%',
label: '97.0%',
caption: 'MITRE-tactic accuracy · substrate · per-case',
},
{
// `mtc_p50_seconds` is null on every row of the published scoreboard: the
// weekly wet-eval has never run against a funded provider key. This tile
// used to read "Sub-minute p50 · wet-eval target", which put an aspiration
// in a row of measurements and let a reader take it for one.
value: undefined,
label: 'Not measured',
caption: 'Live-LLM investigation latency · no funded provider key',
},
{
value: 35,
suffix: ' ms',
label: '35 ms',
caption: 'Full substrate suite runtime on a laptop',
},
];
export function BenchmarkBand() {
const prefersReducedMotion = useReducedMotion();
const initial = prefersReducedMotion ? false : { opacity: 0, y: 16 };
return (
<section
id="benchmark"
aria-labelledby="benchmark-heading"
className="relative py-20 sm:py-24 lg:py-28"
>
<div className="mx-auto max-w-7xl px-4 sm:px-6 lg:px-8">
<div className="mx-auto max-w-3xl text-center">
<p className="text-xs font-semibold uppercase tracking-[0.18em] text-velvet-emerald-mint">
Reproducible by anyone
</p>
<h2
id="benchmark-heading"
className="font-velvet-display font-normal mt-3 text-3xl tracking-tight text-velvet-content-primary sm:text-4xl lg:text-[40px] lg:leading-[1.15] lg:tracking-[-0.015em]"
>
Benchmarked, not vibes.
</h2>
<p className="mt-4 text-base leading-relaxed text-velvet-content-secondary sm:text-lg">
Five pytest suites gate every PR over 200 synthetic incidents drawn
from 55 templates plus a 361-event telemetry corpus across 14 log
sources, and per-template macros catch the regression a per-case
mean hides. Every figure carries the label that says what it
measures: a substrate gate grades the harness’s own
consistency, not the agent, and a figure nobody has measured says
so rather than showing a zero.
</p>
</div>
<div className="relative mt-12 overflow-hidden rounded-3xl border border-velvet-border bg-velvet-surface-raised/60 shadow-[0_30px_80px_-32px_rgba(15,23,42,0.7)] lg:mt-16">
<div
aria-hidden="true"
className="absolute inset-0 bg-velvet-pillars-grad opacity-[0.08]"
/>
<ul className="relative grid grid-cols-1 divide-y divide-velvet-border md:grid-cols-3 md:divide-x md:divide-y-0">
{METRICS.map((metric, idx) => (
<motion.li
key={metric.label}
initial={initial}
whileInView={{ opacity: 1, y: 0 }}
viewport={{ once: true, margin: '-15%' }}
transition={{
duration: 0.55,
ease: [0.16, 1, 0.3, 1],
delay: idx * 0.08,
}}
className="flex flex-col items-center gap-2 px-6 py-10 text-center"
>
<p className="font-mono text-4xl font-semibold leading-none tracking-tight text-velvet-content-primary tabular-nums sm:text-5xl lg:text-[56px]">
{metric.value !== undefined ? (
<>
<NumberTicker
value={metric.value}
decimalPlaces={metric.value < 100 && metric.value !== 35 ? 1 : 0}
/>
{metric.suffix ?? ''}
</>
) : (
metric.label
)}
</p>
<p className="max-w-[20ch] text-xs leading-relaxed text-velvet-content-tertiary">
{metric.caption}
</p>
</motion.li>
))}
</ul>
</div>
<div className="mx-auto mt-10 flex max-w-3xl flex-col items-center justify-center gap-3 text-center sm:flex-row sm:gap-5">
<Link
href="/benchmark"
className="group inline-flex h-11 items-center gap-2 rounded-md bg-velvet-emerald-cta px-6 text-sm font-semibold text-velvet-content-primary shadow-[0_1px_0_rgba(255,255,255,0.18)_inset] transition-shadow duration-200 ease-landing-out-quart motion-safe:hover:shadow-glow-emerald-sm focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-velvet-emerald-mint focus-visible:ring-offset-2 focus-visible:ring-offset-velvet-surface-base"
>
<GitCompare className="h-4 w-4" aria-hidden="true" />
Read the methodology
<ArrowRight
className="h-4 w-4 transition-transform duration-200 group-hover:translate-x-0.5 motion-reduce:transition-none motion-reduce:group-hover:translate-x-0"
aria-hidden="true"
/>
</Link>
<Link
href={docs('benchmark')}
className="inline-flex items-center gap-1 text-sm font-medium text-velvet-content-tertiary transition-colors duration-200 hover:text-velvet-content-primary focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-velvet-emerald-mint focus-visible:ring-offset-2 focus-visible:ring-offset-velvet-surface-base"
>
Open the public scoreboard
<ExternalLink className="h-3.5 w-3.5" aria-hidden="true" />
</Link>
</div>
</div>
</section>
);
}