diff --git a/src/backend/services/realtime/OpenAIRealtimeRelayService.test.ts b/src/backend/services/realtime/OpenAIRealtimeRelayService.test.ts
index 3faa8523c..812b57dd1 100644
--- a/src/backend/services/realtime/OpenAIRealtimeRelayService.test.ts
+++ b/src/backend/services/realtime/OpenAIRealtimeRelayService.test.ts
@@ -3,7 +3,7 @@ import {
     type IncomingMessage,
     type Server as HttpServer,
 } from 'node:http';
-import type { AddressInfo } from 'node:net';
+import { connect, type AddressInfo, type Socket } from 'node:net';
 import { PassThrough } from 'node:stream';
 import {
     afterAll,
@@ -16,6 +16,7 @@ import {
     vi,
 } from 'vitest';
 import WebSocket, { WebSocketServer } from 'ws';
+import { makeActor } from '../../core/actor.js';
 import { checkRateLimit } from '../../core/http/middleware/rateLimit.js';
 import { withAiCostFactorHook } from '../../drivers/integrationTestUtil';
 import type { IConfig } from '../../types';
@@ -199,6 +200,10 @@ describe('OpenAIRealtimeRelayService relay', () => {
     let upstream: WebSocketServer;
     let user: TestUserCredentials;
     let relayUrl: string;
+    let paidPolicy: NonNullable<
+        ReturnType<PuterServer['services']['metering']['getRegisteredPolicy']>
+    >;
+    let planId = 'test_paid';
     const upstreamRequests: IncomingMessage[] = [];
     const upstreamEvents: {
         type: string;
@@ -216,13 +221,12 @@ describe('OpenAIRealtimeRelayService relay', () => {
             await server.services.auth.createSessionToken(row!, {})
         ).token;
         const metering = server.services.metering;
+        planId = 'test_paid';
+        paidPolicy.monthUsageAllowance = 1_000_000_000;
+        metering.invalidateActorSubscription(row!.uuid);
         const plan = metering.getActorSubscription.bind(metering);
-        vi.spyOn(metering, 'getActorSubscription').mockImplementation(
-            async (actor) => ({
-                ...(await plan(actor)),
-                id: 'test_paid',
-                monthUsageAllowance: 1_000_000_000,
-            }),
+        vi.spyOn(metering, 'getActorSubscription').mockImplementation((actor) =>
+            plan(actor),
         );
     });
 
@@ -473,6 +477,20 @@ describe('OpenAIRealtimeRelayService relay', () => {
             username: 'realtime-user',
             password: 'realtime-user-password',
         });
+        const userRow = await server.stores.user.getByUsername(user.username);
+        paidPolicy = {
+            ...(await server.services.metering.getActorSubscription(
+                makeActor({ user: userRow! }),
+            )),
+            id: 'test_paid',
+            monthUsageAllowance: 1_000_000_000,
+            free: false,
+            noAllowance: false,
+        };
+        server.services.metering.registerPolicy(paidPolicy);
+        server.services.metering.registerSubscriptionResolver(
+            async () => planId,
+        );
         relayUrl = `ws://api.puter.localhost:${port}${REALTIME_PATH}`;
     }, 60_000);
 
@@ -482,8 +500,12 @@ describe('OpenAIRealtimeRelayService relay', () => {
     }, 60_000);
 
     /** Open a client and collect every event until it closes. */
-    const open = (url: string, protocols?: string[]) => {
-        const ws = new WebSocket(url, protocols);
+    const open = (
+        url: string,
+        protocols?: string[],
+        headers?: Record<string, string>,
+    ) => {
+        const ws = new WebSocket(url, protocols, { headers });
         clients.add(ws);
         ws.once('close', () => clients.delete(ws));
         const events: Record<string, unknown>[] = [];
@@ -496,7 +518,10 @@ describe('OpenAIRealtimeRelayService relay', () => {
             vi.waitFor(
                 () => {
                     const e = events.find((x) => x.type === type);
-                    if (!e) throw new Error(`no ${type} yet`);
+                    if (!e)
+                        throw new Error(
+                            `no ${type} yet: ${JSON.stringify(events)}`,
+                        );
                     return e;
                 },
                 { timeout: 5_000 },
@@ -1073,7 +1098,9 @@ describe('OpenAIRealtimeRelayService relay', () => {
     });
 
     it('refuses accounts without a subscription', async () => {
-        // The test user's own plan is a free one.
+        planId = 'user_free';
+        const row = await server.stores.user.getByUsername(user.username);
+        server.services.metering.invalidateActorSubscription(row!.uuid);
         vi.mocked(server.services.metering.getActorSubscription).mockRestore();
         const c = open(`${relayUrl}?auth_token=${user.token}`);
         expect(await c.waitFor('error')).toMatchObject({
@@ -1099,13 +1126,15 @@ describe('OpenAIRealtimeRelayService relay', () => {
             const metering = server.services.metering;
             const resolvers = (metering as unknown as Resolvers)
                 .subscriptionResolvers;
-            resolvers.splice(resolvers.indexOf(failing), 1);
+            const index = resolvers.indexOf(failing);
+            if (index >= 0) resolvers.splice(index, 1);
             // The fallback stays cached for its window otherwise.
             const u = await server.stores.user.getByUsername(user.username);
             metering.invalidateActorSubscription(u!.uuid);
         });
 
         it('asks the client to retry instead of reporting no subscription', async () => {
+            planId = 'user_free';
             const u = await server.stores.user.getByUsername(user.username);
             vi.mocked(
                 server.services.metering.getActorSubscription,
@@ -1235,6 +1264,252 @@ describe('OpenAIRealtimeRelayService relay', () => {
         ws.close();
     });
 
+    it('applies the IP gate before opening an upstream connection', async () => {
+        const before = upstreamRequests.length;
+        const deny = (_key: string, data: { allow: boolean }) => {
+            data.allow = false;
+        };
+        server.clients.event.on('ip.validate', deny);
+        try {
+            const c = open(`${relayUrl}?auth_token=${user.token}`);
+            expect(await c.waitFor('error')).toMatchObject({
+                error: { code: 'permission_denied' },
+            });
+            await c.closed;
+            expect(upstreamRequests).toHaveLength(before);
+        } finally {
+            server.clients.event.off('ip.validate', deny);
+        }
+    });
+
+    it('does not allocate credential limiter buckets for many unique invalid signatures', async () => {
+        const before = upstreamRequests.length;
+        const auth = vi.spyOn(server.services.auth, 'authenticate');
+        const redis = vi.spyOn(server.clients.redis, 'multi');
+        const relay = new OpenAIRealtimeRelayService(
+            { domain: 'puter.localhost', trust_proxy: 'loopback' } as IConfig,
+            server.clients,
+            server.stores,
+            server.services,
+        );
+        const http = createServer();
+        relay.attachHttpServer(http);
+        await new Promise<void>((resolve) =>
+            http.listen(0, '127.0.0.1', resolve),
+        );
+        try {
+            for (let i = 0; i < 64; i++) {
+                const c = open(
+                    `ws://127.0.0.1:${(http.address() as AddressInfo).port}${REALTIME_PATH}?auth_token=invalid-${i}`,
+                    undefined,
+                    {
+                        host: 'api.puter.localhost',
+                        'x-forwarded-for': `203.0.113.${i + 1}`,
+                    },
+                );
+                expect(await c.waitFor('error')).toMatchObject({
+                    error: { code: 'invalid_api_key' },
+                });
+                await c.closed;
+            }
+            expect(auth).not.toHaveBeenCalled();
+            expect(redis).not.toHaveBeenCalled();
+            expect(upstreamRequests).toHaveLength(before);
+        } finally {
+            await relay.onServerPrepareShutdown();
+            await new Promise<void>((resolve) => http.close(() => resolve()));
+        }
+    });
+
+    it('uses non-touching authentication on connection and recheck', async () => {
+        const auth = vi.spyOn(server.services.auth, 'authenticate');
+        const c = open(`${relayUrl}?auth_token=${user.token}`);
+        await c.waitFor('session.created');
+        kv.del(realtimeAuthKey(user.token));
+        await server.services.openaiRealtime.recheckSessions();
+        expect(auth.mock.calls.length).toBeGreaterThanOrEqual(2);
+        expect(auth.mock.calls.every(([, ctx]) => ctx?.touch === false)).toBe(
+            true,
+        );
+        c.ws.close();
+    });
+
+    it.each([
+        [false, '203.0.113.9, 10.0.0.1', '127.0.0.1'],
+        [['loopback', '10.0.0.0/8'], '203.0.113.9, 10.0.0.1', '203.0.113.9'],
+        [['loopback'], '203.0.113.9, 198.51.100.5', '198.51.100.5'],
+    ])(
+        'resolves forwarded chains using trust policy %j',
+        async (trustProxy, forwarded, expected) => {
+            const relay = new OpenAIRealtimeRelayService(
+                {
+                    domain: 'puter.localhost',
+                    trust_proxy: trustProxy,
+                } as IConfig,
+                server.clients,
+                server.stores,
+                server.services,
+            );
+            const http = createServer();
+            relay.attachHttpServer(http);
+            await new Promise<void>((resolve) =>
+                http.listen(0, '127.0.0.1', resolve),
+            );
+            let seen = '';
+            const deny = (
+                _key: string,
+                data: { allow: boolean; ip: string },
+            ) => {
+                seen = data.ip;
+                data.allow = false;
+            };
+            server.clients.event.on('ip.validate', deny);
+            try {
+                const ws = new WebSocket(
+                    `ws://127.0.0.1:${(http.address() as AddressInfo).port}${REALTIME_PATH}`,
+                    {
+                        headers: {
+                            host: 'api.puter.localhost',
+                            'x-forwarded-for': forwarded as string,
+                        },
+                    },
+                );
+                ws.on('error', () => {});
+                await new Promise((resolve) => ws.once('close', resolve));
+                expect(seen).toBe(expected);
+            } finally {
+                server.clients.event.off('ip.validate', deny);
+                await relay.onServerPrepareShutdown();
+                await new Promise<void>((resolve) =>
+                    http.close(() => resolve()),
+                );
+            }
+        },
+    );
+
+    it.each([undefined, '', 'not-a-jwt', 'x'.repeat(8193)])(
+        'rejects missing or malformed credentials before upstream admission',
+        async (token) => {
+            const before = upstreamRequests.length;
+            const c = open(
+                token === undefined
+                    ? relayUrl
+                    : `${relayUrl}?auth_token=${token}`,
+            );
+            expect(await c.waitFor('error')).toMatchObject({
+                error: { code: 'invalid_api_key' },
+            });
+            await c.closed;
+            expect(upstreamRequests).toHaveLength(before);
+        },
+    );
+
+    it('reports a distributed limiter outage as retryable without authenticating or connecting upstream', async () => {
+        const before = upstreamRequests.length;
+        const auth = vi.spyOn(server.services.auth, 'authenticate');
+        vi.spyOn(server.clients.redis, 'multi').mockImplementationOnce(() => {
+            throw new Error('limiter unavailable');
+        });
+        const c = open(`${relayUrl}?auth_token=${user.token}`);
+        expect(await c.waitFor('error')).toMatchObject({
+            error: { code: 'internal_error' },
+        });
+        expect((await c.closed).code).toBe(1011);
+        expect(auth).not.toHaveBeenCalled();
+        expect(upstreamRequests).toHaveLength(before);
+    });
+
+    it('keeps recovery attempt capacity after the shared handover budget is exhausted', async () => {
+        const holder = await createTestUser(server, {
+            username: 'handover-attempts',
+            password: 'handover-attempts-password',
+        });
+        for (let i = 0; i < 20; i++)
+            await checkRateLimit(
+                `${realtimeConnectKey(holder.token)}:handover`,
+                20,
+                60_000,
+            );
+        const probe = open(
+            `${relayUrl}?auth_token=${holder.token}&purpose=handover`,
+        );
+        expect(await probe.waitFor('error')).toMatchObject({
+            error: { code: 'rate_limit_exceeded' },
+        });
+        await probe.closed;
+        const recovery = open(`${relayUrl}?auth_token=${holder.token}`);
+        await recovery.waitFor('session.created');
+        expect(
+            await checkRateLimit(realtimeConnectKey(holder.token), 2, 60_000),
+        ).toBe(true);
+        recovery.ws.close();
+    });
+    it('bounds unverified socket allocation while rejected clients withhold close acknowledgments', async () => {
+        const relay = new OpenAIRealtimeRelayService(
+            { domain: 'puter.localhost', trust_proxy: 'loopback' } as IConfig,
+            server.clients,
+            server.stores,
+            server.services,
+        );
+        const http = createServer();
+        relay.attachHttpServer(http);
+        await new Promise<void>((resolve) =>
+            http.listen(0, '127.0.0.1', resolve),
+        );
+        const raw = new Set<Socket>();
+        const accepted: WebSocket[] = [];
+        const upgrade = WebSocketServer.prototype.handleUpgrade;
+        vi.spyOn(WebSocketServer.prototype, 'handleUpgrade').mockImplementation(
+            function (this: WebSocketServer, req, socket, head, done) {
+                upgrade.call(this, req, socket, head, (ws, request) => {
+                    accepted.push(ws);
+                    done(ws, request);
+                });
+            },
+        );
+        const request = (index: number) =>
+            new Promise<string>((resolve, reject) => {
+                const socket = connect(
+                    (http.address() as AddressInfo).port,
+                    '127.0.0.1',
+                );
+                raw.add(socket);
+                socket.on('error', () => {});
+                socket.once('error', reject);
+                socket.once('data', (data) => resolve(data.toString()));
+                socket.once('connect', () =>
+                    socket.write(
+                        [
+                            `GET ${REALTIME_PATH}?auth_token=invalid-${index} HTTP/1.1`,
+                            'Host: api.puter.localhost',
+                            'Connection: Upgrade',
+                            'Upgrade: websocket',
+                            'Sec-WebSocket-Version: 13',
+                            'Sec-WebSocket-Key: dGhlIHNhbXBsZSBub25jZQ==',
+                            `X-Forwarded-For: 203.0.113.${index + 1}`,
+                            '',
+                            '',
+                        ].join('\r\n'),
+                    ),
+                );
+            });
+        try {
+            for (let i = 0; i < 64; i++)
+                expect(await request(i)).toMatch(/^HTTP\/1\.1 101/);
+            await vi.waitFor(() =>
+                expect(
+                    accepted.every((ws) => ws.readyState === WebSocket.CLOSING),
+                ).toBe(true),
+            );
+            expect(await request(64)).toMatch(/^HTTP\/1\.1 503/);
+            expect(accepted).toHaveLength(64);
+        } finally {
+            for (const socket of raw) socket.destroy();
+            await relay.onServerPrepareShutdown();
+            await new Promise<void>((resolve) => http.close(() => resolve()));
+        }
+    });
+
     it('closes a session whose credential stops authenticating', async () => {
         const relay = server.services.openaiRealtime;
         const auth = vi.spyOn(server.services.auth, 'authenticate');
diff --git a/src/backend/services/realtime/OpenAIRealtimeRelayService.ts b/src/backend/services/realtime/OpenAIRealtimeRelayService.ts
index 2e3f3b92f..24f828b71 100644
--- a/src/backend/services/realtime/OpenAIRealtimeRelayService.ts
+++ b/src/backend/services/realtime/OpenAIRealtimeRelayService.ts
@@ -20,6 +20,8 @@
 import { createHash } from 'node:crypto';
 import type { IncomingMessage, Server as HttpServer } from 'node:http';
 import type { Duplex } from 'node:stream';
+import express from 'express';
+import type { AnyTokenPayload } from '../auth/types.js';
 import WebSocket, { WebSocketServer, type RawData } from 'ws';
 import type { Actor } from '../../core/actor.js';
 import {
@@ -148,8 +150,13 @@ const HEARTBEAT_MS = 30_000;
 const UPSTREAM_HANDSHAKE_MS = 10_000;
 const SESSION_SETUP_MS = 15_000;
 const DEFAULT_MAX_PER_USER = 20;
-// Session attempts per credential, checked before authenticating.
+// Session attempts per signature-verified credential.
 const CONNECTS_PER_MINUTE = 60;
+const HANDOVER_CONNECTS_PER_MINUTE = 20;
+const PREAUTH_PER_IP = 120;
+const PREAUTH_PER_NODE = 600;
+const MAX_AUTHENTICATING = 64;
+const MAX_TOKEN_BYTES = 8192;
 // Each response holds this much audio, at the session model's rates, from the
 // account's budget until its usage is recorded.
 const RESPONSE_HOLD_INPUT_TOKENS = 8_000;
@@ -533,6 +540,8 @@ export class OpenAIRealtimeRelayService extends PuterService {
     // One re-authentication per token at a time, shared by its sessions.
     #authChecks = new Map<string, Promise<Actor | null>>();
     #draining = false;
+    #authenticating = 0;
+    #pendingUpgrades = 0;
     #sweeping = false;
     attachHttpServer(server: HttpServer): void {
         this.#wss = new WebSocketServer({
@@ -546,6 +555,9 @@ export class OpenAIRealtimeRelayService extends PuterService {
             handleProtocols: (protocols) =>
                 protocols.has('realtime') ? 'realtime' : false,
         });
+        // Use Express itself so upgrades follow the ordinary HTTP trust policy.
+        const ipApp = express();
+        ipApp.set('trust proxy', this.config.trust_proxy ?? false);
         const domain = (this.config.domain ?? '').toLowerCase().trim();
         const apiHost = `api.${domain}`;
 
@@ -553,23 +565,101 @@ export class OpenAIRealtimeRelayService extends PuterService {
             if (!req.url?.startsWith(REALTIME_PATH)) return;
             const url = new URL(req.url, 'http://placeholder');
             if (url.pathname !== REALTIME_PATH) return;
+            socket.on('error', () => socket.destroy());
+            const refuseUpgrade = (
+                status:
+                    | '404 Not Found'
+                    | '429 Too Many Requests'
+                    | '503 Service Unavailable',
+            ) => {
+                socket.end(
+                    `HTTP/1.1 ${status}\r\nConnection: close\r\n\r\n`,
+                    () => socket.destroy(),
+                );
+            };
             const host = (req.headers.host ?? '').split(':')[0].toLowerCase();
             if (!domain || host !== apiHost) {
-                // Node drops its socket error listener once a request upgrades.
-                socket.on('error', () => socket.destroy());
-                socket.end('HTTP/1.1 404 Not Found\r\n\r\n');
+                refuseUpgrade('404 Not Found');
                 return;
             }
-            this.#wss!.handleUpgrade(req, socket, head, (client) => {
-                this.#runSession(client, req, url).catch((err: unknown) => {
-                    console.warn('[realtime] session setup failed', err);
-                    if (client.readyState !== WebSocket.OPEN) return;
-                    client.send(
-                        errorEvent('internal_error', 'Realtime setup failed'),
-                    );
-                    client.close(1011, 'internal error');
-                });
+            if (
+                this.#authenticating >= MAX_AUTHENTICATING ||
+                this.#pendingUpgrades >= MAX_AUTHENTICATING
+            ) {
+                refuseUpgrade('503 Service Unavailable');
+                return;
+            }
+            const ipRequest = Object.assign(Object.create(ipApp.request), {
+                headers: req.headers,
+                connection: req.socket,
+                socket: req.socket,
             });
+            const clientIp = ipRequest.ip as string;
+            this.#pendingUpgrades++;
+            let pendingUpgrade = true;
+            const releaseUpgrade = () => {
+                if (!pendingUpgrade) return;
+                pendingUpgrade = false;
+                this.#pendingUpgrades--;
+                socket.removeListener('close', releaseUpgrade);
+            };
+            socket.once('close', releaseUpgrade);
+            void (async () => {
+                try {
+                    if (
+                        !(await checkRateLimit(
+                            'realtime:preauth:node',
+                            PREAUTH_PER_NODE,
+                            60_000,
+                            'memory',
+                        )) ||
+                        !(await checkRateLimit(
+                            `realtime:preauth:ip:${clientIp}`,
+                            PREAUTH_PER_IP,
+                            60_000,
+                            'memory',
+                        ))
+                    ) {
+                        refuseUpgrade('429 Too Many Requests');
+                        return;
+                    }
+                    if (socket.destroyed) return;
+                    if (this.#authenticating >= MAX_AUTHENTICATING) {
+                        refuseUpgrade('503 Service Unavailable');
+                        return;
+                    }
+                    this.#wss!.handleUpgrade(req, socket, head, (client) => {
+                        this.#authenticating++;
+                        this.#runSession(
+                            client,
+                            req,
+                            url,
+                            clientIp,
+                            releaseUpgrade,
+                        )
+                            .catch((err: unknown) => {
+                                console.warn(
+                                    '[realtime] session setup failed',
+                                    err,
+                                );
+                                if (client.readyState !== WebSocket.OPEN)
+                                    return;
+                                client.send(
+                                    errorEvent(
+                                        'internal_error',
+                                        'Realtime setup failed',
+                                    ),
+                                );
+                                client.close(1011, 'internal error');
+                            })
+                            .finally(() => {
+                                this.#authenticating--;
+                            });
+                    });
+                } catch {
+                    refuseUpgrade('503 Service Unavailable');
+                }
+            })();
         });
 
         this.#heartbeatTimer = setInterval(() => {
@@ -654,6 +744,8 @@ export class OpenAIRealtimeRelayService extends PuterService {
         client: WebSocket,
         req: IncomingMessage,
         url: URL,
+        clientIp: string,
+        onAdmitted: () => void,
     ): Promise<void> {
         // Buffer what the client sends while auth and the upstream handshake
         // run; OpenAI clients send `session.update` immediately on open.
@@ -774,7 +866,11 @@ export class OpenAIRealtimeRelayService extends PuterService {
                 if (kv.get(realtimeAuthKey(token)) !== undefined) return;
                 rechecking = true;
                 try {
-                    const next = await this.#authenticateCached(token, req);
+                    const next = await this.#authenticateCached(
+                        token,
+                        req,
+                        clientIp,
+                    );
                     if (closed) return;
                     if (!next || next.user?.uuid !== actor.user?.uuid) {
                         reject(
@@ -1174,12 +1270,44 @@ export class OpenAIRealtimeRelayService extends PuterService {
             return;
         }
 
+        const ipGate = { allow: true, ip: clientIp };
+        await this.clients.event.emitAndWait('ip.validate', ipGate, {});
+        if (closed) return;
+        if (!ipGate.allow) {
+            reject('permission_denied', 'Connection refused');
+            return;
+        }
+        try {
+            if (!token || Buffer.byteLength(token) > MAX_TOKEN_BYTES)
+                throw new Error('invalid token');
+            this.services.token.verify<AnyTokenPayload>('auth', token);
+        } catch {
+            reject('invalid_api_key', 'Authentication failed');
+            return;
+        }
         if (
-            token &&
+            url.searchParams.get('purpose') === 'handover' &&
             !(await checkRateLimit(
-                realtimeConnectKey(token),
+                `${realtimeConnectKey(token!)}:handover`,
+                HANDOVER_CONNECTS_PER_MINUTE,
+                60_000,
+                undefined,
+                { failOpen: false },
+            ))
+        ) {
+            reject(
+                'rate_limit_exceeded',
+                'Too many realtime handover attempts',
+            );
+            return;
+        }
+        if (
+            !(await checkRateLimit(
+                realtimeConnectKey(token!),
                 CONNECTS_PER_MINUTE,
                 60_000,
+                undefined,
+                { failOpen: false },
             ))
         ) {
             reject(
@@ -1190,7 +1318,9 @@ export class OpenAIRealtimeRelayService extends PuterService {
         }
         if (closed) return;
         try {
-            actor = token ? await this.#authenticateCached(token, req) : null;
+            actor = token
+                ? await this.#authenticateCached(token, req, clientIp)
+                : null;
         } catch (err) {
             // A lookup that failed says nothing about the credential.
             console.warn('[realtime] credential lookup failed', err);
@@ -1390,6 +1520,7 @@ export class OpenAIRealtimeRelayService extends PuterService {
             // The holds are released once this charge is recorded.
             meter(unreportedResponseRows(billingModel, inFlight), ...holds);
         });
+        onAdmitted();
     }
 
     async #record(
@@ -1415,19 +1546,25 @@ export class OpenAIRealtimeRelayService extends PuterService {
     #authenticateCached(
         token: string,
         req: IncomingMessage,
+        clientIp: string,
     ): Promise<Actor | null> {
         const key = realtimeAuthKey(token);
         let check = this.#authChecks.get(key);
         if (!check) {
             check = (async () => {
                 try {
-                    const actor = await this.#authenticate(token, req);
-                    if (actor) {
+                    const actor = await this.#authenticate(
+                        token,
+                        req,
+                        clientIp,
+                    );
+                    if (actor && this.#authChecks.get(key) === check) {
                         kv.set(key, actor.user!.uuid, { EX: AUTH_RECHECK_SEC });
                     }
                     return actor;
                 } finally {
-                    this.#authChecks.delete(key);
+                    if (this.#authChecks.get(key) === check)
+                        this.#authChecks.delete(key);
                 }
             })();
             this.#authChecks.set(key, check);
@@ -1438,10 +1575,12 @@ export class OpenAIRealtimeRelayService extends PuterService {
     async #authenticate(
         token: string,
         req: IncomingMessage,
+        clientIp: string,
     ): Promise<Actor | null> {
         const ua = req.headers['user-agent'];
         const result = await this.services.auth.authenticate(token, {
-            ip: req.socket.remoteAddress,
+            ip: clientIp,
+            touch: false,
             userAgent: typeof ua === 'string' ? ua : undefined,
         });
         if (result.blocked) return null;
diff --git a/src/backend/services/realtime/README.md b/src/backend/services/realtime/README.md
index 9d949f2a7..d92eb77fd 100644
--- a/src/backend/services/realtime/README.md
+++ b/src/backend/services/realtime/README.md
@@ -68,7 +68,17 @@ Each response holds a fixed amount of the account's budget
 until its usage is recorded, so concurrent sessions can't overdraw it; a session
 opens only if one hold fits. Input transcription is priced at the costlier of
 the transcription models in force at commit and at completion. Session attempts
-are limited per credential before authentication runs.
+are limited per credential after signature verification. Local, bounded IP and
+node admission limits protect signature verification and authentication lookups.
+Upgrades use Express's configured proxy trust policy and the `ip.validate` gate.
+Authentication uses `touch: false` on both initial connection and rechecks, so
+opening or maintaining a realtime socket does not update session activity metadata.
+
+Text, transcripts and raw audio are forwarded without new content logging.
+The relay does not invoke `ai.prompt.validate`. A realtime adapter needs the host
+listeners' chat payload and denial contracts before that hook can safely apply
+to instructions and conversation text. Raw audio and generated transcripts need
+separate policy decisions.
 
 Credential checks run in batches of at most 16 with no overlapping sweeps.
 When the drain starts, open sessions get a `puter.reconnect` event and new
diff --git a/src/docs/src/AI/realtime.md b/src/docs/src/AI/realtime.md
index 27bd884b9..e7a621e29 100644
--- a/src/docs/src/AI/realtime.md
+++ b/src/docs/src/AI/realtime.md
@@ -30,6 +30,8 @@ Browser WebSocket clients pass `realtime` and `openai-insecure-api-key.<PUTER_TO
 
 The socket can open before authentication and provider setup finish. Listen for errors and closure immediately. Wait for `session.created`, send your session settings, then wait for `session.updated` before submitting a turn.
 
+To open a replacement while another connection remains healthy, set the optional `purpose=handover` query parameter. These attempts share a smaller per-credential budget, checked before the general connection-attempt budget. Initial connections and dropped-connection recovery use the URL without this parameter. The limits are listed in [rate limits and quotas](/rate-limits-and-quotas/).
+
 ## Text example
 
 Install the `ws` package and set `PUTER_TOKEN` to your Puter token. Run this example as a Node.js ES module:
@@ -109,7 +111,7 @@ A new connection starts an empty conversation. To keep context, add earlier item
 
 ## Errors, billing, and limits
 
-API errors arrive as `{ "type": "error", "error": { "code": "...", "message": "..." } }`. Handle socket errors and close frames too. Invalid, missing, or revoked credentials use `invalid_api_key`. If Puter can't check a credential or the account's plan, the error is `internal_error` and the session closes with 1011, so try again after a short, randomized delay. Other relay codes include `model_not_found`, `insufficient_funds`, `subscription_required`, `rate_limit_exceeded`, `invalid_event`, `connection_timeout`, `provider_unavailable`, `upstream_error`, and `server_restarting`.
+API errors arrive as `{ "type": "error", "error": { "code": "...", "message": "..." } }`. Handle socket errors and close frames too. Invalid, missing, or revoked credentials use `invalid_api_key`. If Puter can't check a credential or the account's plan, the error is `internal_error` and the session closes with 1011, so try again after a short, randomized delay. Other relay codes include `permission_denied`, `model_not_found`, `insufficient_funds`, `subscription_required`, `rate_limit_exceeded`, `invalid_event`, `connection_timeout`, `provider_unavailable`, `upstream_error`, and `server_restarting`.
 
 Policy failures generally close with 1008, oversized client messages with 1009, upstream or setup failures with 1011, server restarts with 1012, and temporary server overload with 1013. Reconnect after 1011, 1012, or 1013. Authentication and model failures can arrive after the WebSocket handshake succeeds.
 
diff --git a/src/docs/src/rate-limits-and-quotas.md b/src/docs/src/rate-limits-and-quotas.md
index ea5b785d9..243a66621 100644
--- a/src/docs/src/rate-limits-and-quotas.md
+++ b/src/docs/src/rate-limits-and-quotas.md
@@ -111,7 +111,12 @@ Realtime sessions ([`puter.ai.convo()`](/AI/convo/), `/puterai/openai/v1/realtim
 | Upstream handshake timeout | 10 seconds | Close 1011. |
 | Heartbeat interval | 30 seconds | A missing pong terminates the connection on the next heartbeat. |
 | Credential cache and recheck | Cache 60 seconds, sweep every 15 seconds, at most 16 checks in parallel | Invalid credentials close with 1008. A check that fails at connection gets `internal_error` and closes with 1011. A failed recheck keeps the session and tries again on the next sweep. Overlapping sweeps are skipped. |
-| Session attempts per credential | 60 per minute | Error `rate_limit_exceeded`, close 1008. |
+| Session attempts per signature-verified credential | 60 per minute | Error `rate_limit_exceeded`, close 1008. |
+| Pre-auth attempts | 120 per client IP and 600 per node per minute | Local memory admission runs before the WebSocket handshake and signature verification. Exceeding either rate returns HTTP 429. Client IP follows the configured proxy trust policy. The shared memory limiter caps its state at 10,000 keys with a 1-hour retention floor and a 60-second sweep. |
+| Unauthenticated upgrade sockets, including closing sockets | 64 per node | Further upgrades receive HTTP 503. The slot is retained until successful admission or transport closure. |
+| Authentication setups in progress | 64 per node | Further upgrades receive HTTP 503. |
+| Presented credential | 8192 UTF-8 bytes | Longer or malformed credentials receive `invalid_api_key`. Unverified credentials allocate no distributed credential limiter state. Limiter lookup failures close with `internal_error` and 1011. |
+| Replacement attempts marked `purpose=handover` | 20 per credential per minute | This shared limit is checked before the general 60-attempt limit. A refused probe consumes no general attempt. Initial and dropped-session recovery attempts retain their general budget. |
 | Credit held per response | 8,000 audio input and 4,096 audio output tokens at the session model's rates, held until the response's usage is recorded | A session opens only if the account can cover one hold. A response the account can't cover gets `insufficient_funds`, close 1008, and is billed for what it generated. A response whose usage never arrives because the provider connection closed first is billed the full hold. |
 | Restart notice (`puter.reconnect`) | At most once a minute per session | Repeated announcements within the minute aren't sent. |
 | In-flight response drain at shutdown | 5 seconds | Close 1012. |