Evaluation: 4-round policy, validations store, prev/next, copy/clear, build flag

- Bump policy from 3 → 4 rounds per (match, map): first, last, 2 PRNG-picked
middle rounds (deterministic on match_id+map). Fully reviewing all 4
without flagging implies the (match, map) is good.
- New localStorage store opencs2:eval:validations:v1: clicking Next
validates the current candidate. firstUnreviewed/nextUnreviewed skip
already-flagged or already-validated candidates so a returning reviewer
resumes where they left off (header eval entry jumps to first unreviewed).
- Eval bar adds Previous (←) for backtracking and a "N reviewed" counter.
- Flag dialog gets a "Local review data" footer with live flag/validation
counts, a Copy button (full JSON to clipboard via exportReviews), and a
Trash button to clear both stores. Note field already free-text.
- New "Don't flag" entry: "Recording ends right at death" — POVs cut on the
exact death tick (intentional, to avoid camera glitches).
- Build flag PUBLIC_DISABLE_EVAL=1 hides the entire eval surface (header
toggle, eval bar, flag dialog) — module stays in bundle but never
renders. Documented in README.

README.md

@@ -115,6 +115,10 @@ bun run check    # svelte-check
 
 The repo ships a `Dockerfile` and `serve.ts` that mirror the Hugging Face Space deployment.
 
+### Build flags
+
+- `PUBLIC_DISABLE_EVAL=1` — hide the evaluation surface (header toggle, eval bar, flag dialog) for public deploys. Set at build time only (`PUBLIC_DISABLE_EVAL=1 bun run build`); the eval module stays in the bundle but is never rendered.
+
 ## Viewer internals
 
 - **`hyparquet` + `hyparquet-compressors`** read the match and round parquet shards directly from `hf://` URLs. No server, no DuckDB, no WASM bundle larger than necessary.

src/lib/components/eval-bar.svelte

@@ -1,16 +1,21 @@
 <script lang="ts">
 	import { Button } from '$lib/components/ui/button';
-	import * as DropdownMenu from '$lib/components/ui/dropdown-menu';
 	import * as Tooltip from '$lib/components/ui/tooltip';
 	import FlagIcon from 'phosphor-svelte/lib/FlagIcon';
-	import InfoIcon from 'phosphor-svelte/lib/InfoIcon';
 	import ArrowRightIcon from 'phosphor-svelte/lib/ArrowRightIcon';
+	import ArrowLeftIcon from 'phosphor-svelte/lib/ArrowLeftIcon';
 	import XIcon from 'phosphor-svelte/lib/XIcon';
 	import { goto } from '$app/navigation';
 	import { page } from '$app/state';
-	import { addFlag, FLAG_REASONS, evalUrl, type EvalCandidate, type FlagReason } from '$lib/eval';
+	import {
+		addValidation,
+		evalUrl,
+		nextUnreviewed,
+		reviewedKeySet,
+		type EvalCandidate
+	} from '$lib/eval';
 	import { toast } from 'svelte-sonner';
-	import EvalInfoDialog from './eval-info-dialog.svelte';
+	import EvalFlagDialog from './eval-flag-dialog.svelte';
 
 	interface Props {
 		queue: EvalCandidate[];
@@ -21,26 +26,41 @@
 	}
 	let { queue, index, matchId, mapName, round }: Props = $props();
 
-	let infoOpen = $state(false);
+	let flagOpen = $state(false);
+	// Bumped after each review action so the dialog count and the Next-skip
+	// logic re-read localStorage without needing a reactive store.
+	let reviewBump = $state(0);
 
-	function flag(reason: FlagReason, label: string) {
-		addFlag({ matchId, mapName, round, reason });
-		toast.success(`Flagged: ${label}`, {
-			description: `match ${matchId} · ${mapName} · round ${round}`
-		});
-	}
+	const reviewedCount = $derived.by(() => {
+		void reviewBump;
+		const reviewed = reviewedKeySet();
+		let n = 0;
+		for (const c of queue) if (reviewed.has(`${c.matchId}|${c.mapName}|${c.round}`)) n++;
+		return n;
+	});
 
 	function next() {
-		const ni = index < 0 ? 0 : index + 1;
-		if (ni >= queue.length) {
+		// Hitting Next without flagging implicitly validates the current candidate.
+		if (round && index >= 0) {
+			addValidation({ matchId, mapName, round });
+			reviewBump++;
+		}
+		const ni = nextUnreviewed(queue, index < 0 ? -1 : index, reviewedKeySet());
+		if (ni < 0) {
 			toast.success('Evaluation complete', {
-				description: `Reached end of queue (${queue.length} candidates).`
+				description: `Reviewed ${reviewedCount + 1} / ${queue.length} candidates.`
 			});
 			return;
 		}
 		goto(evalUrl(queue[ni], ni));
 	}
 
+	function previous() {
+		if (index <= 0) return;
+		const pi = index - 1;
+		goto(evalUrl(queue[pi], pi));
+	}
+
 	function exitEval() {
 		const url = new URL(page.url);
 		url.searchParams.delete('eval');
@@ -60,31 +80,10 @@
 			>Eval</span
 		>
 		<span class="text-muted-foreground tabular-nums">{positionLabel}</span>
+		<span class="text-muted-foreground/60 tabular-nums">·</span>
+		<span class="text-muted-foreground tabular-nums">{reviewedCount} reviewed</span>
 
 		<div class="ml-auto flex items-center gap-1">
-			<DropdownMenu.Root>
-				<DropdownMenu.Trigger>
-					{#snippet child({ props })}
-						<Button {...props} variant="outline" size="sm">
-							<FlagIcon size={14} weight="fill" /> Flag
-						</Button>
-					{/snippet}
-				</DropdownMenu.Trigger>
-				<DropdownMenu.Content align="end" class="w-72">
-					<DropdownMenu.Label>Flag this candidate as…</DropdownMenu.Label>
-					<DropdownMenu.Separator />
-					{#each FLAG_REASONS as r (r.id)}
-						<DropdownMenu.Item
-							onclick={() => flag(r.id, r.label)}
-							class="flex-col items-start gap-0.5"
-						>
-							<span class="font-medium">{r.label}</span>
-							<span class="text-[10px] text-muted-foreground">{r.description}</span>
-						</DropdownMenu.Item>
-					{/each}
-				</DropdownMenu.Content>
-			</DropdownMenu.Root>
-
 			<Tooltip.Root>
 				<Tooltip.Trigger>
 					{#snippet child({ props })}
@@ -92,16 +91,21 @@
 							{...props}
 							variant="ghost"
 							size="icon-sm"
-							onclick={() => (infoOpen = true)}
-							aria-label="What to flag"
+							onclick={previous}
+							disabled={index <= 0}
+							aria-label="Previous candidate"
 						>
-							<InfoIcon size={14} weight="duotone" />
+							<ArrowLeftIcon size={14} weight="bold" />
 						</Button>
 					{/snippet}
 				</Tooltip.Trigger>
-				<Tooltip.Content side="bottom">How to evaluate</Tooltip.Content>
+				<Tooltip.Content side="bottom">Previous candidate</Tooltip.Content>
 			</Tooltip.Root>
 
+			<Button variant="outline" size="sm" onclick={() => (flagOpen = true)}>
+				<FlagIcon size={14} weight="fill" /> Flag
+			</Button>
+
 			<Button onclick={next} size="sm" disabled={total === 0}>
 				Next <ArrowRightIcon size={14} weight="bold" />
 			</Button>
@@ -126,4 +130,11 @@
 	</div>
 </div>
 
-<EvalInfoDialog bind:open={infoOpen} />
+<EvalFlagDialog
+	bind:open={flagOpen}
+	{matchId}
+	{mapName}
+	{round}
+	queueLength={total}
+	onChange={() => reviewBump++}
+/>

src/lib/components/eval-flag-dialog.svelte

@@ -0,0 +1,207 @@
+<script lang="ts">
+	import * as Dialog from '$lib/components/ui/dialog';
+	import { Button } from '$lib/components/ui/button';
+	import { Label } from '$lib/components/ui/label';
+	import {
+		addFlag,
+		clearFlags,
+		clearValidations,
+		exportReviews,
+		FLAG_REASONS,
+		KNOWN_MINOR_ISSUES,
+		loadFlags,
+		loadValidations,
+		type FlagReason
+	} from '$lib/eval';
+	import ArrowSquareOutIcon from 'phosphor-svelte/lib/ArrowSquareOutIcon';
+	import ClipboardIcon from 'phosphor-svelte/lib/ClipboardIcon';
+	import TrashIcon from 'phosphor-svelte/lib/TrashIcon';
+	import { toast } from 'svelte-sonner';
+
+	interface Props {
+		open?: boolean;
+		matchId: number;
+		mapName: string;
+		round: number;
+		queueLength: number;
+		onChange?: () => void;
+	}
+	let { open = $bindable(false), matchId, mapName, round, queueLength, onChange }: Props = $props();
+
+	let reason = $state<FlagReason | null>(null);
+	let note = $state('');
+	// Bumped to refresh the counts when the user copies/clears or saves a flag.
+	let storeBump = $state(0);
+
+	$effect(() => {
+		if (!open) {
+			reason = null;
+			note = '';
+		} else {
+			storeBump++;
+		}
+	});
+
+	const flagsCount = $derived.by(() => {
+		void storeBump;
+		return loadFlags().length;
+	});
+	const validationsCount = $derived.by(() => {
+		void storeBump;
+		return loadValidations().length;
+	});
+
+	function save() {
+		if (!reason) return;
+		const trimmed = note.trim();
+		addFlag({ matchId, mapName, round, reason, note: trimmed || undefined });
+		const label = FLAG_REASONS.find((r) => r.id === reason)?.label ?? 'Flagged';
+		toast.success(`Flagged: ${label}`, {
+			description: `match ${matchId} · ${mapName} · round ${round}`
+		});
+		onChange?.();
+		open = false;
+	}
+
+	async function copyAll() {
+		const data = exportReviews(queueLength);
+		try {
+			await navigator.clipboard.writeText(JSON.stringify(data, null, 2));
+			toast.success('Review data copied to clipboard', {
+				description: `${data.flags.length} flags · ${data.validations.length} validations`
+			});
+		} catch {
+			toast.error('Could not copy to clipboard');
+		}
+	}
+
+	function clearAll() {
+		clearFlags();
+		clearValidations();
+		storeBump++;
+		onChange?.();
+		toast.success('Cleared local review data');
+	}
+</script>
+
+<Dialog.Root bind:open>
+	<Dialog.Content
+		class="flex max-h-[min(80vh,40rem)] w-full max-w-xl flex-col gap-3 overflow-hidden p-0 sm:max-w-xl"
+	>
+		<Dialog.Header class="border-b px-4 pt-4 pb-3">
+			<Dialog.Title>Flag this candidate</Dialog.Title>
+			<Dialog.Description>
+				match {matchId} · {mapName} · round {round}. Pick a reason and (optionally) leave a note.
+			</Dialog.Description>
+		</Dialog.Header>
+
+		<div class="flex min-h-0 flex-col gap-3 overflow-y-auto px-4">
+			<div class="grid gap-1.5">
+				{#each FLAG_REASONS as r (r.id)}
+					<button
+						type="button"
+						data-active={reason === r.id || undefined}
+						class="flex flex-col items-start gap-0.5 rounded-md border border-border px-2.5 py-1.5 text-left text-xs transition hover:bg-muted/50 data-active:border-amber-500 data-active:bg-amber-500/10"
+						onclick={() => (reason = r.id)}
+					>
+						<span class="flex items-center gap-2">
+							<span class="font-medium text-foreground">{r.label}</span>
+							<span
+								data-severity={r.severity}
+								class="rounded-sm border px-1 py-0 text-[9px] font-semibold tracking-wider text-muted-foreground/80 uppercase data-[severity=major]:border-rose-500/30 data-[severity=major]:text-rose-600 dark:data-[severity=major]:text-rose-400"
+								>{r.severity}</span
+							>
+						</span>
+						<span class="text-[11px] leading-snug text-muted-foreground">{r.description}</span>
+						{#if r.examples?.length}
+							<span class="mt-1 flex flex-wrap items-center gap-x-2 gap-y-0.5 text-[10px]">
+								<span class="tracking-wider text-muted-foreground/70 uppercase">Examples:</span>
+								{#each r.examples as href, i (href)}
+									<a
+										{href}
+										target="_blank"
+										rel="noreferrer noopener"
+										onclick={(e) => e.stopPropagation()}
+										class="inline-flex items-center gap-0.5 text-muted-foreground underline-offset-2 hover:text-foreground hover:underline"
+									>
+										#{i + 1}
+										<ArrowSquareOutIcon size={10} weight="bold" />
+									</a>
+								{/each}
+							</span>
+						{/if}
+					</button>
+				{/each}
+			</div>
+
+			<div class="grid gap-1.5">
+				<Label for="flag-note" class="text-xs">Notes (optional)</Label>
+				<textarea
+					id="flag-note"
+					bind:value={note}
+					placeholder="e.g. player 4 spawn position is at mid instead of T spawn from t=0"
+					rows="3"
+					class="w-full resize-y rounded-md border border-input bg-input/20 px-2 py-1.5 text-xs outline-none focus-visible:border-ring focus-visible:ring-2 focus-visible:ring-ring/30 dark:bg-input/30"
+				></textarea>
+			</div>
+
+			{#each KNOWN_MINOR_ISSUES as k (k.label)}
+				<div class="rounded-md border border-dashed border-muted-foreground/20 p-2.5 text-[11px]">
+					<div
+						class="mb-0.5 text-[10px] font-semibold tracking-wider text-muted-foreground/80 uppercase"
+					>
+						Don't flag — known minor
+					</div>
+					<div class="font-medium text-foreground">{k.label}</div>
+					<div class="mt-0.5 leading-snug text-muted-foreground">{k.description}</div>
+					{#if k.examples?.length}
+						<div class="mt-1 flex flex-wrap gap-2 text-[10px]">
+							{#each k.examples as href, i (href)}
+								<a
+									{href}
+									target="_blank"
+									rel="noreferrer noopener"
+									class="inline-flex items-center gap-0.5 text-muted-foreground underline-offset-2 hover:text-foreground hover:underline"
+								>
+									#{i + 1}
+									<ArrowSquareOutIcon size={10} weight="bold" />
+								</a>
+							{/each}
+						</div>
+					{/if}
+				</div>
+			{/each}
+
+			<div class="rounded-md border bg-muted/30 px-2.5 py-2 text-[11px]">
+				<div class="flex items-center justify-between gap-2">
+					<div>
+						<div class="font-medium text-foreground">Local review data</div>
+						<div class="mt-0.5 text-[10px] text-muted-foreground">
+							<span class="tabular-nums">{flagsCount}</span> flags ·
+							<span class="tabular-nums">{validationsCount}</span> validations · stored in
+							<code class="font-mono">localStorage</code> (persists across sessions)
+						</div>
+					</div>
+					<div class="flex shrink-0 items-center gap-1">
+						<Button variant="outline" size="sm" onclick={copyAll}>
+							<ClipboardIcon size={12} weight="duotone" /> Copy
+						</Button>
+						<Button
+							variant="ghost"
+							size="icon-sm"
+							onclick={clearAll}
+							aria-label="Clear local review data"
+						>
+							<TrashIcon size={12} weight="duotone" />
+						</Button>
+					</div>
+				</div>
+			</div>
+		</div>
+
+		<Dialog.Footer class="border-t px-4 pt-3 pb-4">
+			<Button variant="ghost" size="sm" onclick={() => (open = false)}>Cancel</Button>
+			<Button size="sm" onclick={save} disabled={!reason}>Save flag</Button>
+		</Dialog.Footer>
+	</Dialog.Content>
+</Dialog.Root>

src/lib/components/eval-info-dialog.svelte

@@ -1,43 +0,0 @@
-<script lang="ts">
-	import * as Dialog from '$lib/components/ui/dialog';
-	import { FLAG_REASONS } from '$lib/eval';
-
-	interface Props {
-		open?: boolean;
-	}
-	let { open = $bindable(false) }: Props = $props();
-</script>
-
-<Dialog.Root bind:open>
-	<Dialog.Content class="max-w-xl">
-		<Dialog.Header>
-			<Dialog.Title>How to evaluate</Dialog.Title>
-			<Dialog.Description>
-				The queue samples 3 rounds per match — round 1, the last round, and one deterministic round
-				in the middle. Watch each candidate, flag any issue you spot, then click <span
-					class="font-medium text-foreground">Next</span
-				> to advance.
-			</Dialog.Description>
-		</Dialog.Header>
-
-		<div class="mt-3 space-y-3 text-sm">
-			<div>
-				<div class="font-semibold text-foreground">Flag if you see:</div>
-				<ul class="mt-1.5 space-y-1.5 text-xs/relaxed text-muted-foreground">
-					{#each FLAG_REASONS as r (r.id)}
-						<li class="flex flex-col gap-0.5">
-							<span class="font-medium text-foreground">{r.label}</span>
-							<span>{r.description}</span>
-						</li>
-					{/each}
-				</ul>
-			</div>
-
-			<p class="border-t pt-3 text-xs text-muted-foreground">
-				Flags are stored in your browser's <code class="font-mono">localStorage</code> under
-				<code class="font-mono">opencs2:eval:flags:v1</code>. Export them by copying that key from
-				devtools.
-			</p>
-		</div>
-	</Dialog.Content>
-</Dialog.Root>

src/lib/components/header.svelte

@@ -11,7 +11,8 @@
 	import HfLogo from '$lib/components/hf-logo.svelte';
 	import { site } from '$lib/site';
 	import type { Match } from '$lib/types';
-	import { buildEvalQueue, evalUrl } from '$lib/eval';
+	import { buildEvalQueue, evalUrl, firstUnreviewed, EVAL_ENABLED } from '$lib/eval';
+	import { toast } from 'svelte-sonner';
 
 	const inEval = $derived(page.url.searchParams.get('eval') === '1');
 	const matches = $derived((page.data?.matches ?? []) as Match[]);
@@ -26,7 +27,15 @@
 		}
 		const queue = buildEvalQueue(matches);
 		if (!queue.length) return;
-		goto(evalUrl(queue[0], 0));
+		// Resume on the first un-reviewed candidate so prior sessions aren't replayed.
+		const start = firstUnreviewed(queue);
+		if (start < 0) {
+			toast.success('All eval candidates already reviewed', {
+				description: `${queue.length} total. Clear local data in the Flag dialog to redo.`
+			});
+			return;
+		}
+		goto(evalUrl(queue[start], start));
 	}
 </script>
 
@@ -42,27 +51,29 @@
 		</a>
 
 		<div class="ml-auto flex items-center gap-1">
-			<Tooltip.Root>
-				<Tooltip.Trigger>
-					{#snippet child({ props })}
-						<Button
-							{...props}
-							variant="ghost"
-							size="icon-sm"
-							onclick={toggleEval}
-							disabled={!matches.length}
-							data-active={inEval || undefined}
-							class="data-active:bg-amber-500/15 data-active:text-amber-700 dark:data-active:text-amber-300"
-							aria-label={inEval ? 'Exit evaluation' : 'Start evaluation'}
-						>
-							<ListChecksIcon size={16} weight="duotone" />
-						</Button>
-					{/snippet}
-				</Tooltip.Trigger>
-				<Tooltip.Content side="bottom">
-					{inEval ? 'Exit evaluation' : 'Start evaluation'}
-				</Tooltip.Content>
-			</Tooltip.Root>
+			{#if EVAL_ENABLED}
+				<Tooltip.Root>
+					<Tooltip.Trigger>
+						{#snippet child({ props })}
+							<Button
+								{...props}
+								variant="ghost"
+								size="icon-sm"
+								onclick={toggleEval}
+								disabled={!matches.length}
+								data-active={inEval || undefined}
+								class="data-active:bg-amber-500/15 data-active:text-amber-700 dark:data-active:text-amber-300"
+								aria-label={inEval ? 'Exit evaluation' : 'Start evaluation'}
+							>
+								<ListChecksIcon size={16} weight="duotone" />
+							</Button>
+						{/snippet}
+					</Tooltip.Trigger>
+					<Tooltip.Content side="bottom">
+						{inEval ? 'Exit evaluation' : 'Start evaluation'}
+					</Tooltip.Content>
+				</Tooltip.Root>
+			{/if}
 			<Tooltip.Root>
 				<Tooltip.Trigger>
 					{#snippet child({ props })}

src/lib/eval.ts

@@ -1,6 +1,16 @@
 import { browser } from '$app/environment';
+import { env } from '$env/dynamic/public';
 import type { Match } from '$lib/types';
 
+/**
+ * Build-time switch. Set `PUBLIC_DISABLE_EVAL=1` (or `=true`) before running
+ * `bun run build` to ship a viewer with the entire evaluation surface
+ * (header button, eval bar, flag dialog) hidden. The eval module stays in
+ * the bundle but is never rendered.
+ */
+export const EVAL_ENABLED =
+	env.PUBLIC_DISABLE_EVAL !== '1' && env.PUBLIC_DISABLE_EVAL?.toLowerCase() !== 'true';
+
 export type EvalCandidate = {
 	matchId: number;
 	mapName: string;
@@ -8,45 +18,119 @@ export type EvalCandidate = {
 };
 
 export type FlagReason =
+	| 'victory_screen'
+	| 'wrong_initial_position'
+	| 'no_animation'
+	| 'missing_video'
 	| 'missing_audio'
 	| 'av_misaligned'
 	| 'pov_desync'
 	| 'uninteresting'
-	| 'missing_video'
 	| 'other';
 
-export const FLAG_REASONS: { id: FlagReason; label: string; description: string }[] = [
+export type FlagSeverity = 'major' | 'minor';
+
+export type FlagReasonInfo = {
+	id: FlagReason;
+	label: string;
+	description: string;
+	severity: FlagSeverity;
+	examples?: string[];
+};
+
+// Severity is informational — both major and minor are still flaggable. The
+// `examples` URLs are stable links to a representative case so a reviewer can
+// confirm what the failure mode looks like.
+export const FLAG_REASONS: FlagReasonInfo[] = [
+	{
+		id: 'victory_screen',
+		label: 'Victory screen instead of POV',
+		severity: 'major',
+		description:
+			"Round-end / scoreboard screen renders in place of the player's first-person view, usually for the whole round on the same player slot. Renders are essentially unusable.",
+		examples: [
+			'https://blanchon-opencs2-dataset-viewer.hf.space/match/2393397/de_overpass?round=1&player=7&view=grid'
+		]
+	},
+	{
+		id: 'wrong_initial_position',
+		label: 'Wrong initial position',
+		severity: 'major',
+		description:
+			'Player is not at their spawn point at the very first tick of the round (most visible at t=0).',
+		examples: [
+			'https://blanchon-opencs2-dataset-viewer.hf.space/match/2392873/de_dust2?round=1&player=2&view=grid',
+			'https://blanchon-opencs2-dataset-viewer.hf.space/match/2392873/de_mirage?round=1&player=0&view=grid'
+		]
+	},
+	{
+		id: 'no_animation',
+		label: 'No animation',
+		severity: 'major',
+		description:
+			'Player or world animations stop playing — character moves through space but limbs / weapons / world stay frozen.',
+		examples: [
+			'https://blanchon-opencs2-dataset-viewer.hf.space/match/2392131/de_mirage?round=1&player=2'
+		]
+	},
+	{
+		id: 'missing_video',
+		label: 'Missing video',
+		severity: 'major',
+		description:
+			'A POV stream stays blank after the round has buffered (i.e. not just a slow-network hiccup).'
+	},
 	{
 		id: 'missing_audio',
 		label: 'Missing audio',
+		severity: 'major',
 		description: 'No audio at all when there should be (gunfire, footsteps, callouts).'
 	},
 	{
 		id: 'av_misaligned',
 		label: 'Audio out of sync',
-		description: 'Audio is offset from on-screen action — gunshots before the muzzle flash, etc.'
+		severity: 'major',
+		description:
+			'Audio is offset from on-screen action — gunshots before the muzzle flash, footsteps lagging the movement, etc.'
 	},
 	{
 		id: 'pov_desync',
 		label: 'POVs out of sync',
+		severity: 'major',
 		description:
 			'In grid mode, two players who should see the same moment are time-offset from each other.'
 	},
 	{
 		id: 'uninteresting',
 		label: 'Uninteresting gameplay',
+		severity: 'minor',
 		description: 'Pure AFK, intentional griefing, or otherwise unusable footage.'
 	},
 	{
-		id: 'missing_video',
-		label: 'Missing video',
+		id: 'other',
+		label: 'Other',
+		severity: 'minor',
+		description: 'Something else worth recording — use the notes field to describe.'
+	}
+];
+
+// Cosmetic / known issues that look like problems but aren't — listed in the
+// flag dialog so reviewers stop reporting them.
+export const KNOWN_MINOR_ISSUES: { label: string; description: string; examples?: string[] }[] = [
+	{
+		label: '"Terrorist/CT win" tail at round start',
 		description:
-			'A POV stream stays blank after the round has buffered (i.e. not just a slow-network hiccup).'
+			'A short sting from the previous round can leak into the start of the next round. Cosmetic, not a render bug — please skip.',
+		examples: [
+			'https://blanchon-opencs2-dataset-viewer.hf.space/match/2392131/de_mirage?round=16&player=0&view=grid',
+			'https://blanchon-opencs2-dataset-viewer.hf.space/match/2393398/de_dust2?round=16&player=4&view=grid',
+			'https://blanchon-opencs2-dataset-viewer.hf.space/match/2393178/de_dust2?round=27&player=0&view=grid'
+		]
 	},
 	{
-		id: 'other',
-		label: 'Other',
-		description: 'Something else worth recording.'
+		label: 'Recording ends right at death',
+		description:
+			"Each POV is cut on the exact frame the player dies. On headshots that can feel like the recording ended too soon — it's intentional, to avoid the camera snapping to the killer or glitching out around the death tick."
 	}
 ];
 
@@ -55,34 +139,77 @@ export type Flag = {
 	mapName: string;
 	round: number;
 	reason: FlagReason;
+	note?: string;
+	ts: number;
+};
+
+export type Validation = {
+	matchId: number;
+	mapName: string;
+	round: number;
 	ts: number;
 };
 
 const FLAGS_KEY = 'opencs2:eval:flags:v1';
+const VALIDATIONS_KEY = 'opencs2:eval:validations:v1';
 
-export function loadFlags(): Flag[] {
+function loadJson<T>(key: string): T[] {
 	if (!browser) return [];
 	try {
-		return JSON.parse(localStorage.getItem(FLAGS_KEY) ?? '[]') as Flag[];
+		return JSON.parse(localStorage.getItem(key) ?? '[]') as T[];
 	} catch {
 		return [];
 	}
 }
-
-export function saveFlags(flags: Flag[]) {
+function saveJson<T>(key: string, v: T[]) {
 	if (!browser) return;
-	localStorage.setItem(FLAGS_KEY, JSON.stringify(flags));
+	localStorage.setItem(key, JSON.stringify(v));
 }
 
+export const loadFlags = () => loadJson<Flag>(FLAGS_KEY);
+export const saveFlags = (v: Flag[]) => saveJson(FLAGS_KEY, v);
+export const loadValidations = () => loadJson<Validation>(VALIDATIONS_KEY);
+export const saveValidations = (v: Validation[]) => saveJson(VALIDATIONS_KEY, v);
+
 export function addFlag(f: Omit<Flag, 'ts'>) {
 	const flags = loadFlags();
 	flags.push({ ...f, ts: Date.now() });
 	saveFlags(flags);
 }
 
-// Mulberry32 — small deterministic PRNG so the picked middle round is stable
-// for a given (match_id, map_name) and the eval set doesn't shift between
-// sessions.
+export function addValidation(v: Omit<Validation, 'ts'>) {
+	const validations = loadValidations();
+	const key = `${v.matchId}|${v.mapName}|${v.round}`;
+	if (validations.some((x) => `${x.matchId}|${x.mapName}|${x.round}` === key)) return;
+	validations.push({ ...v, ts: Date.now() });
+	saveValidations(validations);
+}
+
+export function clearFlags() {
+	if (browser) localStorage.removeItem(FLAGS_KEY);
+}
+export function clearValidations() {
+	if (browser) localStorage.removeItem(VALIDATIONS_KEY);
+}
+
+const candidateKey = (c: { matchId: number; mapName: string; round: number }) =>
+	`${c.matchId}|${c.mapName}|${c.round}`;
+
+/**
+ * Set of (matchId, mapName, round) keys that have been "reviewed" — either
+ * flagged or validated by clicking Next. Used to skip already-seen
+ * candidates when computing the next position in eval mode.
+ */
+export function reviewedKeySet(): Set<string> {
+	const set = new Set<string>();
+	for (const f of loadFlags()) set.add(candidateKey(f));
+	for (const v of loadValidations()) set.add(candidateKey(v));
+	return set;
+}
+
+// Mulberry32 — small deterministic PRNG so the picked middle rounds are
+// stable for a given (match_id, map_name) and the eval set doesn't shift
+// between sessions.
 function prng(seed: number): number {
 	let t = (seed + 0x6d2b79f5) | 0;
 	t = Math.imul(t ^ (t >>> 15), t | 1);
@@ -91,9 +218,13 @@ function prng(seed: number): number {
 }
 
 /**
- * Evaluation policy: per (match, map), sample 3 rounds — first, last, and
- * one deterministic random round in the middle. Sorted by match_id, then
+ * Evaluation policy: per (match, map), sample 4 rounds — first, last, and
+ * two deterministic random rounds in the middle. Sorted by match_id then
  * map_index to match the rest of the app's canonical order.
+ *
+ * Validating ALL 4 candidates of a (match, map) without flagging implies the
+ * whole match-map is good; flagging any one of the 4 marks the match-map as
+ * having issues.
  */
 export function buildEvalQueue(matches: Match[]): EvalCandidate[] {
 	const sorted = matches
@@ -106,10 +237,18 @@ export function buildEvalQueue(matches: Match[]): EvalCandidate[] {
 		if (!total || total < 1) continue;
 
 		const picks = new Set<number>([1, total]);
-		if (total > 2) {
-			const seed = m.match_id * 31 + m.map_name.length * 17 + (m.map_index ?? 0);
+		// Two PRNG-picked middle rounds when there's room for them.
+		if (total >= 4) {
+			const seedBase = m.match_id * 31 + m.map_name.length * 17 + (m.map_index ?? 0);
 			const span = total - 2; // pick from [2, total - 1]
-			picks.add(2 + Math.floor(prng(seed) * span));
+			let attempts = 0;
+			while (picks.size < 4 && attempts < 16) {
+				const r = 2 + Math.floor(prng(seedBase + attempts) * span);
+				picks.add(r);
+				attempts++;
+			}
+		} else if (total === 3) {
+			picks.add(2);
 		}
 
 		for (const round of [...picks].sort((a, b) => a - b)) {
@@ -130,6 +269,33 @@ export function indexOfCandidate(
 	);
 }
 
+/**
+ * Find the next un-reviewed candidate (strictly after `fromIndex`). Returns
+ * the index in `queue`, or -1 if all remaining candidates have been
+ * reviewed.
+ */
+export function nextUnreviewed(
+	queue: EvalCandidate[],
+	fromIndex: number,
+	reviewed: Set<string> = reviewedKeySet()
+): number {
+	for (let i = fromIndex + 1; i < queue.length; i++) {
+		if (!reviewed.has(candidateKey(queue[i]))) return i;
+	}
+	return -1;
+}
+
+/** First un-reviewed candidate index, or -1 if everything is done. */
+export function firstUnreviewed(
+	queue: EvalCandidate[],
+	reviewed: Set<string> = reviewedKeySet()
+): number {
+	for (let i = 0; i < queue.length; i++) {
+		if (!reviewed.has(candidateKey(queue[i]))) return i;
+	}
+	return -1;
+}
+
 export function evalUrl(c: EvalCandidate, i: number): string {
 	const params = new URLSearchParams({
 		round: String(c.round),
@@ -140,3 +306,20 @@ export function evalUrl(c: EvalCandidate, i: number): string {
 	});
 	return `/match/${encodeURIComponent(c.matchId)}/${encodeURIComponent(c.mapName)}?${params}`;
 }
+
+export type ReviewExport = {
+	exportedAt: string;
+	totalCandidates: number;
+	flags: Flag[];
+	validations: Validation[];
+};
+
+/** Snapshot of the current localStorage state for sharing/exporting. */
+export function exportReviews(queueLength: number): ReviewExport {
+	return {
+		exportedAt: new Date().toISOString(),
+		totalCandidates: queueLength,
+		flags: loadFlags(),
+		validations: loadValidations()
+	};
+}

src/routes/match/[matchId]/[mapName]/+page.svelte

@@ -4,7 +4,7 @@
 	import { goto } from '$app/navigation';
 	import Header from '$lib/components/header.svelte';
 	import EvalBar from '$lib/components/eval-bar.svelte';
-	import { buildEvalQueue, indexOfCandidate } from '$lib/eval';
+	import { buildEvalQueue, indexOfCandidate, EVAL_ENABLED } from '$lib/eval';
 	import RoundList from '$lib/components/round-list.svelte';
 	import PlayerGrid from '$lib/components/player-grid.svelte';
 	import VideoStage from '$lib/components/video-stage.svelte';
@@ -79,7 +79,7 @@
 
 	// Evaluation mode driven by ?eval=1&i=N. Queue is deterministic from the
 	// match list so reloading or sharing a URL still lands on the same item.
-	const inEvalMode = $derived(page.url.searchParams.get('eval') === '1');
+	const inEvalMode = $derived(EVAL_ENABLED && page.url.searchParams.get('eval') === '1');
 	const evalQueue = $derived(inEvalMode ? buildEvalQueue(data.matches) : []);
 	const evalIndex = $derived.by(() => {
 		if (!inEvalMode) return -1;