Glossary

Overview

Each entry gives a one-line orientation and links to the section that defines the term and describes it in full; the definition is on that page.

Some concepts appear on both kinds of PCC node but take a different form on each, so the right page depends on the node type:

Requests and routing

Application protocol

The protocol a client and node use to exchange an encrypted request and response. See Application protocol.

Data Encryption Key (DEK)

The per-request symmetric key a client generates to encrypt the request payload. See Application protocol.

Fraud Detection Service (FDS)

The service that makes anonymous fraud determinations about entities requesting tokens. See Fraud data protocol.

k

The number of independent nodes that can decrypt a given request; this number bounds the strength of target diffusion. See Selecting the size of k.

Node selection

How PCC chooses which candidate nodes may serve a request. See Node selection.

Node validation

The checks performed on a node’s attestation before user data reaches it, whether the user’s device performs them or a proxy node performs them on the device’s behalf. See Node validation.

Non-targetability

The requirement that reaching a specific user’s data take a broad compromise of the fleet, not of a chosen node. See Non-Targetability.

Oblivious Gateway (OG)

The Apple-operated gateway that decrypts the Oblivious HTTP layer of a request; clients encrypt to its public key configuration. See Network transport.

Oblivious Relay (OR)

A third-party relay that hides a device’s source IP address from PCC infrastructure. See Network transport.

One-Time Token (OTT)

The single-use credential that authorizes a request without tying it to a user. See Client authentication.

PCC Gateway

The service that routes incoming requests to nodes advertising a matching capability. See Node selection.

Request Encryption Key (REK)

The key pair a node publishes, and to which users’ devices encrypt requests. See Attested Request Handling.

Target diffusion

The bound on the share of incoming requests any single node can decrypt. See Target diffusion.

Token Granting Service (TGS)

The service that issues one-time tokens to authorized clients. See Client authentication.

Transitive trust

The mechanism by which a node that received a request extends its guarantees to nodes it forwards work to. See Transitive Trust.

Nodes and node roles

Agent node

A proxy node running PCC Agent, which orchestrates multi-step and tool-using workflows. See PCC Agent.

Cloud application

A PCC application that handles incoming requests on a node. See PCC cloud applications.

CloudBoard

The node’s front end for requests, which terminates each request and isolates it in its own set of per-request processes. See CloudBoard.

Proxy application

A cloud application that runs on a proxy node and distributes a request to other nodes. See PCC cloud applications.

Releases and transparency

Attestation bundle

The structure a node produces that carries its measurements and its attested request key. See The attestation bundle for Apple silicon and Appendix: Google Cloud Attestation Bundle Contents for Google Cloud.

Release

A structured record of the measurements describing the software a node runs, and the unit recorded in the transparency log. See Recording a release.

Transparency log

The append-only log of releases that lets a researcher enumerate every release a node may run. See Release Transparency; for the log’s wire format, see Appendix: Transparency Log.

Trusted release set

The set of releases a node accepts when extending trust to another node. See Trusted release sets.

Inference

First token (FT) and extend token (ET)

The names that some PCC code, logs, and tools use for the prefill and decode phases. See Prefill-decode disaggregation.

Interaction, turn, and sub-turn

The units of a multi-turn conversation, which determine what state PCC may reuse. See Interactions, turns, and sub-turns.

KV cache

The intermediate model state from processing a prompt, which PCC may reuse across the turns of one interaction, and may transfer between the nodes that split one turn. See KV cache reuse for Apple silicon and Per-user KV cache for Google Cloud.

KV-Cache Data Encryption Key (KVDEK)

The key derived from a request’s DEK to encrypt KV cache data in transport. See Prefill-decode disaggregation.

Prefill-decode disaggregation

Splitting a request’s prompt processing and token generation across separate nodes. See Prefill-decode disaggregation.

Recycling

Returning a Google Cloud node’s inference process to a clean state between requests. See Recycling.

Static prefix cache

A cache of Apple-supplied prompt material, holding no user input. See Static prefix cache for Apple silicon and System-prompt prefix cache for Google Cloud.

The Inference Engine (TIE)

The engine that runs model inference on a PCC node. See The Inference Engine for Apple silicon and The inference pipeline for Google Cloud.

Thinking trace

An intermediate sequence of reasoning tokens, returned to the device in encrypted form. See Encrypted thinking traces.

Apple silicon nodes

Apple Component Attestation Authority (CAA)

The system that records the sealed identities of a node’s components. See Hardware life cycle.

APTicket

The Image4 boot manifest carrying signed measurements of the boot loader and each firmware image. See Secure and measured boot.

Codeless cryptex

A cryptex carrying only data and configuration, with no executable code. See Codeless cryptexes.

Cryptex

A signed, integrity-verified software distribution package that extends a node’s base software. See Cryptexes.

darwin-init

The configuration a node applies at boot; its attestation covers the configuration’s policy state and a selected set of keys. See darwin-init.

Data Center Identity Key (DCIK)

The key establishing a node’s cryptographic identity as authentic PCC hardware. See Cryptographic identity.

Ensemble

A grouping of nodes that serve one request together, with a leader that coordinates them. See Distributed inference.

Ensemble attestation

The mechanism binding every node in an ensemble into the attestation a client verifies. See Ensemble attestation.

Ephemeral Data Mode

The mode in which a node holds user data on a per-boot volume that cannot outlive the boot. See Ephemeral Data Mode.

Image4

Apple’s Secure Boot specification, which we use to manage the authorized lifetime of code. See Secure and measured boot.

Private Cloud Extensions

The cryptex layer providing PCC-specific services above the base operating system. See Private Cloud Extensions.

Public-Key Accelerator (PKA)

The Secure Enclave hardware block that generates keys entangled with the node’s UID. See Cryptographic identity.

Restricted Execution Mode (REM)

The node state that locks out code loading, debugging, and similar interfaces before user data arrives. See Restricted Execution Mode.

Secure Boot

The measured boot chain from Boot ROM through the kernel and into user space. See Secure Boot.

Secure Enclave Processor (SEP)

The separate processor holding a node’s hardware root of trust and attestation keys. See Immutable features of Apple silicon.

Secure Page Table Monitor (SPTM) and Trusted Execution Monitor (TXM)

The monitor layers that own memory management and code-execution policy, outside the kernel. See Secure Boot.

Software Sealed Registers (SSRs)

SEP registers into which a node ratchets post-boot measurements; attestations capture their contents. See Software Sealed Registers.

Trust cache

The list of code a node may execute, itself covered by attestation. See Appendix: Apple Silicon Trust Cache Format.

Trusted Signing Service (TSS)

The service through which we manage the authorized lifetime of signed code. See Secure and measured boot.

Unique ID (UID)

The per-device secret fused into the Secure Enclave during manufacturing. See Cryptographic identity.

Google Cloud nodes

aCOS

Google’s attested Container-Optimized OS, the open source image that supplies a node’s host and guest operating systems, its hypervisor, and its Key Protection Service. See The software on a node.

AFM Runtime

The runtime that executes the language model and the vision encoder. See The inference pipeline.

Bones

The library that executes the language model for the AFM Runtime, including the CUDA kernels beneath it. See The inference pipeline.

Confidential VM (CVM)

A virtual machine whose memory the CPU’s confidential-computing hardware isolates from the host. See Confidential computing.

Device Identity Manifest (DIM)

The manifest binding together the identities of the parts that make up one node. See Device identity.

Digest catalog and execution allow list

The build-time inventory of the executables in the Megacontainer, and the kernel-enforced policy loaded from it. See Execution enforcement.

Hardened and debug images

The two build variants of a node’s images, distinct measured artifacts differing in the debugging surface they expose. See Appendix: Google Cloud Debug and Hardened Images.

Hardware ledger

The append-only record of the Google Cloud hardware that is part of the PCC fleet, which a client uses to check a node. See Hardware ledger.

Intel Trust Domain Extensions (TDX)

The CPU technology that isolates a confidential VM from its host and measures the VM’s initial state. See Confidential computing.

Key Protection Service

The service that holds request-decryption keys in a confidential VM of its own, isolated from the VM that parses requests. See Protecting request keys.

Measured, unmeasured, and dynamic configuration

The three classes of node configuration, which differ in whether attestation covers them. See Measured, unmeasured, and dynamic configuration.

Measurement Register for the Trust Domain (MRTD)

The TDX register recording a confidential VM’s initial memory contents. See Guest measurements.

Megacontainer

The single measured container image that packages our runtime components onto a node. See Workload Megacontainer.

NVIDIA Confidential Computing

The GPU mode that encrypts traffic into and out of the GPU and decrypts it only within the GPU’s protected memory. See Confidential computing.

Runtime Measurement Registers (RTMRs)

The TDX registers recording firmware, boot loader, and container-launch measurements. See Guest measurements.

Titan

Google’s hardware root-of-trust chip, which measures the host boot chain and holds the host’s measurements. See Multiple hardware roots of trust.

Warm-reset attack

An attack that resets a node to reuse its attested state; the evidence binding detects the attempt. See Detecting a warm-reset attack.

Workload Services Daemon

The daemon exposing a node’s key-management and attestation interfaces to the workload. See Appendix: Google Cloud Workload Services Daemon.

Workload VM and Key Protection Service VM

The two confidential VMs on a node, one running the workload that parses requests and the other holding the keys that decrypt them. See Protecting request keys.

Applications

Private Filter Service (PFS)

The service that answers membership queries over the fingerprints PCC Agent generates from model output. See Private Filter Service.

Private Information Retrieval (PIR)

The technique that lets a device query a database without revealing which record it retrieved. See PIR fundamentals.

Recitation prevention

The mechanism that helps keep a model from reproducing memorized training data verbatim. See Recitation prevention.

Safety override

A mechanism for rapidly adjusting model behavior in response to a safety issue. See Safety overrides.

Tool gateway

The service outside the PCC trust boundary that relays PCC Agent’s external tool calls to downstream tool providers. See External tool calls.

Visual Lookup (VLU)

The feature that identifies visual entities by matching image embeddings against a database of known entities. See Visual Lookup.

Encodings

Abstract Syntax Notation One (ASN.1)

The ITU-T notation that defines the structure of Image4 manifests, releases, and certificates. See ITU-T X.680.

Distinguished Encoding Rules (DER)

The ASN.1 encoding that gives each value exactly one encoding, so a digest over a DER-encoded structure is stable. See ITU-T X.690.

Research and operations

pccvre

The tool that creates and manages VRE instances, and inspects attestations, releases, and the transparency log on either platform. See Get Started with the VRE and Google Cloud Research Tooling.

Security event

A limited, privacy-preserving signal a node reports to support compromise detection. See Security events for Apple silicon and Security event monitoring for Google Cloud.

Virtual Research Environment (VRE)

The environment in which researchers can run and inspect the Apple silicon PCC software stack. See Virtual Research Environment.