Glossary
Acceptance set. The set of outputs a model is permitted to produce for a given input under the declared values document. Defined in Alignment.
Action chunk. A fixed-length sequence of motor commands emitted by the embodiment surface in a single forward pass. Typically steps.
Attestation. A signed statement, optionally rooted in a hardware-trusted execution environment, that a specific computation produced a specific output.
Audit fraction. The probability that a given query is dispatched to a second validator for divergence checking. Denoted .
Backbone. The cognitive transformer stack that ingests perceptual and instruction tokens and produces cognition tokens. Shared across embodiments.
Challenge. A claim, submitted on chain within the challenge window, that a validator's attested output is fraudulent. Successful challenges produce slashing.
Checkpoint. A model's weights at a specific training step, content-addressed and licensed for public release when protocol-funded.
Cognition tokens. A small fixed-count set of learnable queries that absorb the task-relevant information from the latent. The output of the backbone and the input to the action transformer.
Coordination layer. The off-chain, signed-message layer mediating between the on-chain settlement layer and the validator-local execution layer.
DPO (Direct Preference Optimization). A preference-based fine-tuning technique that updates a policy directly against preference data without an intermediate reward model.
Embodiment tag. A learned embedding identifying a registered embodiment family, injected into the cognition tokens to condition the action transformer.
Episode. A recorded trajectory of observations, actions, and outcomes from a physical or simulated body. The fundamental unit of embodied data.
Eval suite. A frozen set of test cases run against checkpoints to determine acceptance for promotion. Versioned and content-addressed.
Flow matching. A continuous-time generative modeling technique that trains a model to predict the velocity field of a probability path from noise to data.
Forgetting eval. A regression suite of historical tasks run before and after a continual-learning update to detect catastrophic forgetting.
Gateway. A public service operator that accepts user queries, forwards them to the routing service, and returns validator-produced outputs. Permissionless.
Hourglass. The codex's term for the architectural pattern in which many surfaces converge on a shared narrow latent and diverge again into surface-specific outputs.
Job ledger. The on-chain registry of posted jobs (training, inference batches, evaluations), bids, selections, and attestations.
Latent. The shared representation space through which every surface communicates. A sequence of typed tokens with discrete codes and continuous residuals.
Manifest. A content-addressed description of a dataset: sources, licenses, hashes, sampling weights. Referenced by every checkpoint.
MSAT (Multi-Stream Action Transformer). The action-side architecture in which vision-language, state-action, and physics token streams interleave through gated cross-stream attention.
P11. The design principle that the system has no permanent operators; foundation roles have scheduled ends.
PDA (Program Derived Address). An on-chain account whose address is deterministically derived from a program and seeds, used for per-job and per-validator state isolation.
Real-time chunking (RTC). The protocol that stitches consecutive action chunks together to produce continuous motion without boundary discontinuities.
Reserve floor. The minimum treasury balance, expressed in months of trailing compute spend, below which no disbursement is permitted.
Routing service. The off-chain service that maps incoming queries to validators based on the current registry, the surface's checkpoint, and the validator's tier and quality score.
Settlement layer. The on-chain layer holding validator state, checkpoint state, treasury, and job ledger. Slow, small, authoritative.
Slashing. The forfeiture of a validator's stake bond following a successful challenge or attestation forgery.
Steward. A token holder exercising governance. The term emphasizes responsibility over ownership.
Surface. A cognitive or embodiment component that reads and writes the shared latent. Language, vision, audio, reasoning, memory, world models, embodiment.
TEE. Trusted Execution Environment. Hardware-isolated execution producing attestations of code and data identity. AMD SEV-SNP, Intel TDX, NVIDIA H100 Confidential Computing.
Tenure multiplier. A governance weighting that scales with continuous-holding duration, bounded by a constitutional ceiling.
Timelock. A mandatory delay between governance approval and execution. Lengths scale with impact.
SPL token. The standard fungible-token program on Solana. OGI deploys as a plain SPL mint with revoked mint authority; the Token-2022 extension set is not used.
Raydium CPMM. Concentrated liquidity market maker on Solana, used to host the protocol-owned liquidity position whose swap fees fund the treasury.
Concentrated liquidity. A liquidity provision style in which the position is active only over a chosen price range, allowing higher fee accrual per unit of capital than uniform-range market making.
Validator. An entity running part of the model on its own hardware, signing outputs, and staking tokens against correctness. The network's compute substrate.
Verification (cryptoeconomic). Correctness enforced by the threat of slashing combined with the ability of any party to challenge, rather than by cryptographic proofs of execution.
World model. The surface predicting future latent states from current latent states and proposed actions. Latent-space, not pixel-space.