Reference

Glossary

1 min readDraftUpdated 2026-10-13

This glossary defines the handbook's vocabulary in one place. Terms are grouped by role rather than alphabetically, so that related terms sit together in nine groups. Each definition is short and links to the chapter that develops it.

How to use it#

Hover a dotted term in any chapter to see its definition without leaving the page; the link brings you here. Filter the list below by group or by any word in a term, symbol, or definition.

Where the field's words collide — reward model, rubric, PRM, RLHF, verifier — the glossary gives the house meaning. Vocabulary without category errors explains why each term sits on the axis it does, and the crosswalk maps common evaluation vocabulary onto the framework's coordinates.

A few conventions run through every entry:

An evaluator is the umbrella term. A verifier has a soundness argument relative to a specification; a judge is a learned or prompted evaluator without one. People are raters, panels, or adjudicators, not judges.
A rubric is a specification, and gold or reference data is an anchor; neither is a rung of the verification hierarchy.
pass@k means at least one of k attempts succeeds; pass^k means all k succeed.
Validity is a property of a claim, never of a benchmark on its own.

Symbols follow the notation and formula card.

All terms

229 terms

Objects and system context

Agent: A goal-directed system operating through an iterative perceive–act loop, usually with tools, state, and stopping conditions. Agency is a property of the loop, not of the model.; Read the chapter
Artifact: The output presented for evaluation: an answer, proof, program, report, plan, edited file, or database state.; Read the chapter
Category error: Treating objects that describe different axes — a specification, a mechanism, a resolution, a use — as peers on one list, as in ranking a rubric against a reward model or calling RLHF an evaluator.; See also Six-axis evaluator typeRead the chapter
Dual-control environment: An interactive environment in which both the agent and a simulated user act through tools on a shared world, each with its own state and toolset, so success depends on coordination and guidance, as in τ²-bench.; See also Simulated userRead the chapter
Environment: External state and transition process the system interacts with; may be simulated, sandboxed, pinned, or live.; Read the chapter
Eval harness: The machinery that administers an evaluation and records trials. It is part of the instrument, distinct from the agent harness, which is part of the system under test.; Read the chapter
Evaluator agent: An agent that administers an assessment: it prepares the environment, sends tasks to participant agents, collects results, and reports the metrics it defines. It bundles harness, specification, and mechanism; Berkeley's lecture calls it a green agent.; Also: green agentSee also Eval harnessRead the chapter
Evaluator construction stack: How an evaluation is built: construct, specification, mechanism, signal, and use, with resolution and deployment context cutting across. It records how the evaluator was engineered, a separate question from the verification hierarchy's evidential strength.; Also: design stackSee also Verification hierarchy, Six-axis evaluator typeRead the chapter
Evaluator systemE_sys: The complete arrangement by which a claim is operationalized, measured, consumed, and maintained: E_sys = (C, I, R, V, L). A mechanism is only one part of it.; Read the chapter
Harness (scaffolding): Everything around the model that turns it into an agent: prompts, tools, memory, retrieval, skills, orchestration, permissions, and stopping logic.; Also: scaffoldingSee also Eval harnessRead the chapter
Instrument chain: The links through which a score supports a decision: target construct, sampled tasks and environments, trials and trajectories, grader readings, an estimate with uncertainty, and the decision. Error and invalid inference can enter at every link.; See also Error families, EstimandRead the chapter
Measurement version5Lifecycle: The unit of evaluator versioning: rubric text, examples, judge model and prompt, checkers, thresholds, aggregation logic, and expert policy, versioned together. Changing a load-bearing part creates a new measurement form: revalidate it or run a bridging study.; See also Eval harness, Versioned dynamismRead the chapter
One Spine: The framework's ten-step workflow for building an evaluator system: claim and regime first, then decomposition, criteria, evidence, mechanisms, natural and adaptive validity, separate channels, tripwires, revalidation, and frame review.; See also Evaluator systemRead the chapter
PopulationP1Claim: The target task, user, or input distribution a claim is about. An eval set is a finite model of it; its headline estimates the population's pass rate only when the set's mix matches.; Also: target population, task distributionSee also Validity domain, Claim scopeRead the chapter
Simulated user: A language model, prompted with personas mined from real conversation logs, that plays the user in multi-turn evals. It is environment, not system: its quirks can create fake failures and fake passes, and it drifts like a judge.; See also Dual-control environment, EnvironmentRead the chapter
Six-axis evaluator type⟨C, Ψ, M, Z, R, U⟩: The tuple ⟨C, Ψ, M, Z, R, U⟩ — construct, specification, mechanism, signal, resolution, use — that describes an evaluator by its value on every axis instead of a single noun. The framework distributes it across E_sys.; Also: evaluator type system, six axesSee also Evaluator system, Evaluator construction stackRead the chapter
System under test (SUT): The model, policy, or agent, plus the relevant observation and action pipeline, whose behavior is being measured.; Read the chapter
Trajectoryτ: The sequence of states and actions (s₀, a₀, …, s_T) generated by an agent interaction.; Read the chapter
Validation triad: The system under test, the environment it acts in, and the specification its behavior is checked against. It exposes two structural failures: faithfully measuring the wrong world, and faithfully checking the wrong property.; Also: system–environment–specification triad, wrong world, wrong propertySee also System under test (SUT), Environment, SpecificationRead the chapter

Meaning and specification

Abstention region2Instrument: Where a criterion's evaluator should answer "cannot determine" or route the item to a person instead of forcing a verdict, for example when sources conflict or the case is out of domain. Forced verdicts turn uncertainty into false confidence.; Also: abstention, escalation regionSee also Criterion cardRead the chapter
Aggregation geometryΓ2Instrument: The rule that combines criterion-level signals into a decision: weighted mean, conjunctive gate, lexicographic order, Pareto vector, or tail-risk rule.; Read the chapter
Claim scope1Claim: Whether an evaluation supports claims about a specific capability, a specific application, or a general set of applications. Scope fixes the population; the broadest scope makes the strongest and most fragile claim.; See also Validity domain, PopulationRead the chapter
Claim tuple𝒞1Claim: The claim layer of an evaluator system, 𝒞 = (C, P, E, H, D, δ): the construct to infer, over which population, environment, and harness, within which validity domain, for which decision.; Also: claim, claim layerSee also Construct, Decision claim, Validity domainRead the chapter
Compensability2Instrument: The assumption that a high score on one criterion can offset a low score on another. A weighted mean asserts it; a gate or a lexicographic order denies it. It is a normative claim, never a default.; See also Aggregation geometry, Conjunctive gateRead the chapter
Conjunctive gate2Instrument: An aggregation rule that accepts only if every required criterion clears its threshold, Γ(z) = ∏ᵢ 𝟙[zᵢ ≥ θᵢ]. Nothing compensates a missed gate, so it encodes fatal failures such as forbidden actions.; Also: hard gateSee also Aggregation geometry, CompensabilityRead the chapter
ConstructC1Claim: The latent property an evaluation intends to measure: a capability, risk, quality, preference, or behavior. Scores and artifacts are observations of it, never the construct itself.; Read the chapter
Construct facetF⁺1Claim: A distinct dimension within a construct, used to build a content blueprint for the task population.; Read the chapter
Construct underrepresentation1Claim: A validity threat in which important facets of the construct are absent from the tasks or too rare, so the score is too narrow for its claim. The facet list guards against it.; See also Construct facet, Construct-irrelevant varianceRead the chapter
Construct-irrelevant variance1Claim: A validity threat in which factors outside the construct, such as output format, token budgets, tool latency, or answer length, drive the score, making it impure. The nuisance list guards against it.; See also Nuisance variable, Construct underrepresentationRead the chapter
Cooperative inverse reinforcement learning (CIRL): A cooperative partial-information game in which human and machine both pursue the human's reward but the machine is uncertain of it, so optimal play includes questions and teaching. Evaluation analogue: query where posterior uncertainty is high.; Also: CIRL, cooperative inverse reinforcement learningSee also Inverse reinforcement learning (IRL)Read the chapter
Criterion card2Instrument: The atomic specification for one criterion: facet, operational definition, allowed evidence, positive and negative exemplars, boundary and abstention rules, severity (fatal, tradeable, advisory, or escalate), mechanism, blind spots, owner, and version.; See also Rubric, Abstention region, Exemplar anchorsRead the chapter
Decision claimδ1Claim: The decision the evidence is meant to change: deploy, gate, select, terminate, reward, or page a human.; Read the chapter
Error analysis: Discovering failure modes from real traces: a domain expert open-codes what went wrong, axial-codes the notes into a counted failure taxonomy, and iterates until about twenty consecutive traces add no new category.; See also Regression setRead the chapter
Evidence contract2Instrument: The part of a criterion card that states which observable evidence is admissible for that criterion: for example cited spans for grounding, or the tool trace and state transition, never the final prose, for a forbidden action.; See also Evidence view, Criterion cardRead the chapter
Evidence viewV2Instrument: What the evaluator may observe: the final answer only, or also sources, tool traces, hidden state, snapshots, provenance and logs.; Read the chapter
Exemplar anchors2Instrument: Positive and negative examples, adversarial near-misses, and scale points attached to a criterion card to fix its boundary. Not the reference anchor A, the anchor channel, or the linking items kept across suite versions.; See also Reference anchor, Criterion cardRead the chapter
Inverse evaluation design2Instrument: A proposed research program that treats expert preferences, corrections, escalations, rubric edits, and accept/reject decisions as evidence of a latent judgment and infers its criteria, weights, gates, and thresholds.; See also Inverse reinforcement learning (IRL), RubricRead the chapter
Inverse reinforcement learning (IRL): Infers which reward functions could explain observed behavior, often as a Bayesian posterior under a Boltzmann (noisily rational) choice model. Many rewards can explain the same behavior, so the inferred reward is only partly identifiable.; Also: IRLSee also Inverse evaluation designRead the chapter
Inverse reward design: Treats a hand-written reward as evidence about its designer's true objective, given the situations it was designed for. Evaluation analogue: a rubric or benchmark score is evidence about what its designer meant to measure, not the construct itself.; See also Inverse evaluation designRead the chapter
Judgment compiler2Instrument: An eleven-step procedure that compiles expert judgment, the source language, through a rubric, the intermediate representation, into executable graders, with calibration and validity studies as its tests.; See also Rubric, Criterion cardRead the chapter
Lexicographic rule≻2Instrument: An aggregation rule that compares options on the highest-priority criterion first, such as safety ≻ quality ≻ cost; a lower criterion counts only among options that the higher ones do not separate.; Also: lexicographic orderSee also Aggregation geometryRead the chapter
Marr's levels: Three levels at which an evaluation question sits: computational (what claim and decision), algorithmic (what representation and procedure), implementation (what machinery). Locate a disagreement's level before arguing about tools.; Also: computational level, algorithmic level, implementation levelRead the chapter
Metric2Instrument: A function mapping a trajectory or artifact to a quantity. Over a task distribution it has a distribution of values, and any reported number is a chosen functional of it. A specification adds the demand, such as a threshold.; See also SpecificationRead the chapter
Nomological network1Claim: The expected network of relationships between a construct's measure and other measures: which facets it should track, which nuisances it should ignore, and how it should relate to independent measures of the same and different constructs.; See also Construct facet, Nuisance variable, Discriminant evidenceRead the chapter
Nuisance variableF⁻1Claim: A property that should not materially move the score when the construct is held fixed, such as length, formatting, or assertive tone. The nuisance list is the discriminant test plan.; Read the chapter
Outer design axes: Four independent axes that locate an eval design: response space (closed to open), oracle strength (verifiable to judgment-based), maintenance (static to versioned dynamic), and interaction (single-turn to stateful trajectory).; Also: four design axesRead the chapter
Pareto frontier2Instrument: The options that no other option beats on every criterion at once, so improving one criterion means worsening another. Plot it before scalarizing; weights then choose a point on it.; Also: Pareto-optimalSee also Aggregation geometryRead the chapter
Reference anchorA2Instrument: The external object a judgment is rooted in: a canonical answer, gold state, reference distribution, adjudicated expert labels, a stronger evaluator, or production outcomes. Gold is a reference anchor, not a mechanism or a rung.; Also: anchor, reference, gold answerSee also Exemplar anchors, Evidence viewRead the chapter
Rubric2Instrument: A structured, human-readable specification with criteria, allowed evidence, exemplar anchors or scales, and gate or aggregation semantics. A rubric is a specification, not a hierarchy rung.; See also Criterion card, Verification hierarchyRead the chapter
Rubric atomization2Instrument: Decomposition taken too far: dozens of micro-criteria whose combination no longer reproduces how experts decide. The right grain is the smallest set of independently meaningful criteria that still reconstructs the expert decision.; See also Rubric soupRead the chapter
Rubric soup2Instrument: Several logically distinct criteria packed into one paragraph and scored by one judge on a single scale, so failures cannot be localized and compensation between criteria stays hidden. Its opposite error is rubric atomization.; See also Rubric atomization, Aggregation geometryRead the chapter
SpecificationΨ2Instrument: The operational statement of success or acceptable behavior: formal, executable, rubric-based, preference-based, or state-based.; Read the chapter
Temporal pattern2Instrument: A trajectory requirement written with temporal operators: safety invariant 𝖦 ¬bad, precedence ¬a 𝖴 b, liveness 𝖥 goal, bounded response, or reachability. The same property grades offline trials and runs as a runtime monitor.; Also: trajectory pattern, temporal property, precedence patternSee also Trajectory, Specification, Robustness marginRead the chapter

Evaluator mechanisms

Capability removal2Instrument: Making an unsafe state unreachable by construction, with read-only credentials, egress allowlists, spend caps, or sandbox walls, so its failure probability is zero as a property of the configured system rather than an estimate.; Also: reachability by constructionSee also Structural guarantee, Configured trusted baseRead the chapter
Completeness: Whether a verifier accepts every true instance of its property. Distinct from soundness, and often traded for tractability. Passing tests establishes neither: it means no tested counterexample was found.; See also SoundnessRead the chapter
Configured trusted base: Everything a formal guarantee is relative to: permitted axioms, theorem statements, checker version, kernel configuration, environment assumptions, and whether the optimizer can modify any of them. The Goodhart boundary moves here rather than disappearing.; See also Formal verifierRead the chapter
Constrained decoding: Grammar-constrained generation that guarantees every output belongs to a formal language: every emission parses and every tool call is schema-valid. A genuine guarantee about syntax, never semantics.; See also Structural guaranteeRead the chapter
Critic: A mechanism that produces diagnostic critique or error information; it need not emit a scalar reward.; Read the chapter
Dark current2Instrument: A judge's baseline preference when there is no real quality signal to detect: what it favors when the options do not differ in the construct. One of the properties a judge datasheet probes.; See also Judge datasheetRead the chapter
Deterministic checker: An exact rule, parser, or comparator without learned semantic judgment. It can still be invalid or incomplete, and is only as good as its reference anchor and normalization.; Read the chapter
Evaluator: The umbrella term for any mechanism or artifact that maps evidence to a quality signal. On its own it is underspecified: ask which coordinate is meant.; Read the chapter
Execution / effect checker: An evaluator that runs code, compiles, simulates, queries state, or observes consequences. Grounded but incomplete: passing tests means no tested counterexample was found.; Read the chapter
Formal verifier: A proof checker, model checker, type system, or similar mechanism whose acceptance has a soundness argument relative to a configured trusted base.; Read the chapter
Human rater or panel: A human measurement mechanism whose reliability, expertise, and disagreement should be characterized rather than assumed infallible.; See also Trust root (reference authority)Read the chapter
Hybrid portfolio2Instrument: An evaluator that routes each criterion to the strongest affordable mechanism for it, such as structural checks, execution, a learned judge, or people, instead of one universal evaluator. Its assurance depends on how coupled its channels are.; See also Channel couplingRead the chapter
Intrinsic signal: Confidence, likelihood, self-consistency, or ungrounded self-critique produced by the generator itself. Cheap, but not an independent observation of truth.; See also Verification hierarchy, Self-confirming loopRead the chapter
Judge: A learned or prompted evaluator without formal soundness guarantees. People who score are raters, panels, or adjudicators, not judges, in this handbook.; See also Verifier, LLM-as-a-judgeRead the chapter
Judge datasheet2Instrument: A maintained record that characterizes a learned judge as an instrument: target sensitivity, nuisance invariance, threshold behavior, reference agreement by criterion and slice, and repeatability and transport. Measure the measuring device before trusting it.; See also LLM-as-a-judge, Dark current, Sensitivity and specificity (TPR, TNR, FPR)Read the chapter
Learned judge: A reward model, LLM judge, learned PRM, classifier, or other learned evaluator: a noisy semantic sensor whose error must be characterized.; Read the chapter
LLM-as-a-judge: A prompted general-purpose LLM evaluating another artifact, usually with a rubric, references, or pairwise comparison.; Read the chapter
MechanismM2Instrument: What computes a judgment: a formal verifier, execution checker, deterministic checker, structural control, learned judge, human rater or panel, intrinsic signal, or a hybrid portfolio. A mechanism is not an evaluator system.; See also Evaluator system, Verification hierarchyRead the chapter
Meta-evaluation: Evaluation of an evaluator. It adds assurance only when it contributes information not already coupled to the primary evaluator's failure modes.; Read the chapter
Metamorphic test: A test that checks a known relation between the outputs of related inputs rather than each output against a reference, so errors such as accidental result-set equivalence surface without a fresh anchor for every input.; See also Execution / effect checkerRead the chapter
Outcome reward model (ORM): A learned judge operating on terminal outputs or completed trajectories.; Read the chapter
Process reward model (PRM): A learned judge operating at step or state resolution. "Process" names the resolution, not a rung above outcome checkers.; See also Process verifier, Value modelRead the chapter
Reliability–coverage trade-off: Signal reliability rises toward the top of the verification hierarchy while task coverage widens toward the bottom. The response is to decompose the judgment and push each criterion as high as it can legitimately go.; Also: reliability–coverage frontierSee also Verification hierarchyRead the chapter
Reward model (RM): A learned model that predicts preference or quality, typically as a scalar or pairwise estimate, and is often optimized against.; Read the chapter
Soundness: A verifier V_ψ is sound when V_ψ(y) = 1 implies ψ(y) = 1: nothing it accepts violates the formal property, provided the verifier and its configured trusted base behave as claimed.; See also Completeness, Configured trusted base, VerifierRead the chapter
Strongest affordable mechanism2Instrument: The rule that amends strongest-oracle-per-criterion for cost and regime: choose the strongest affordable mechanism per criterion per regime, then use different mechanisms across development and certification when independence is valuable.; Also: strongest-oracle-per-criterion ruleSee also Verifiability engineering, Verification hierarchyRead the chapter
Structural control2Instrument: A mechanism family that makes a configured bad state unreachable through read-only credentials, sandboxes, permissions, or allowlists, instead of measuring it. It sits off the verification ladder; the guarantee covers only the configured roles and paths.; See also Structural guarantee, Capability removal, Configured trusted baseRead the chapter
Structural guarantee: A property enforced by the system's construction (permissions, sandboxes, verified scaffold rules, or an output grammar), so a failure is unreachable rather than rare. It holds only for its configured trusted base.; See also Capability removal, Constrained decoding, Subtraction principle, Configured trusted baseRead the chapter
Substitution claim2Instrument: The implicit claim every automated evaluator makes: this observable signal is an adequate substitute for the relevant human judgment, for this decision, on this validity domain. Evaluator engineering makes it explicit, tests it, and bounds where it fails.; See also Validity statementRead the chapter
Subtraction principle: Every property enforced by construction or formal verification is removed from the sampling burden, so scarce empirical evaluation budget is spent only on the properties that cannot be guaranteed.; See also Structural guarantee, Capability removalRead the chapter
Trust root (reference authority): Where the evaluate-the-evaluator chain stops: a configured proof kernel, or an expert panel with measured reliability and a governance mandate.; Read the chapter
Value model: Estimates expected future return from a state or prefix, conditional on a policy. Rollout-labeled "step quality" usually behaves like one.; Read the chapter
Verifiability engineering2Instrument: Moving each property of a task toward the strongest evidence channel its semantics permit, through four moves (decompose, instrument, compile, retain the residue honestly), instead of asking whether a whole domain is verifiable.; Also: four movesSee also Verification hierarchy, Strongest affordable mechanismRead the chapter
Verification hierarchy: Formal verification > execution feedback > learned judges > intrinsic signals: a prior over expected grounding strength, not a verdict on validity. Human judgment sits outside the ladder as the external frontier and trust root.; See also Reliability–coverage trade-off, Trust root (reference authority), Validation profileRead the chapter
Verifier: Strictly, an evaluator with a soundness guarantee relative to a formal specification. In loose usage it often means a deterministic or execution checker.; See also JudgeRead the chapter

Signal and resolution

Best-of-N: Generate several candidates and select by evaluator score. It introduces extreme-selection distribution shift toward the evaluator's tail.; Read the chapter
Outcome evaluation: Judges the terminal artifact or terminal environment state.; Read the chapter
Pass ratep: The estimated probability that evaluated behavior satisfies the operational specification under a declared distribution.; Read the chapter
pass@k: The probability that at least one of k sampled candidates succeeds: 1 − (1 − p)^k under independence. Not to be confused with pass^k.; See also pass^kRead the chapter
pass^k: The probability that all k independent attempts succeed: p^k. The right quantity for unattended or repeated work. Some sources label it pass@k; check which functional a leaderboard computed.; See also pass@kRead the chapter
Process evaluation: Judges intermediate steps, actions, states, or evidence acquisition. Justified when the path itself matters; harmful when it forces one canonical chain.; Read the chapter
Process verifier: Any process-resolution evaluator, whatever its mechanism family: formal, executable, learned, or human. Not a synonym for PRM.; Read the chapter
ResolutionR2Instrument: The unit that is judged: token, step, action, final artifact, trajectory, or a system-level rate.; Read the chapter
Robustness marginρ: A signed or graded distance to satisfying a formal or quantitative specification; useful for partial credit, early warning, and search.; Read the chapter
SignalZ2Instrument: What leaves the evaluator: a boolean, scalar, probability, vector, preference, critique, confidence, or error trace.; Read the chapter
Unbiased pass@k estimator: For a task with c passes among n ≥ k samples, 1 − C(n − c, k)/C(n, k): the share of k-sample subsets containing a pass. Compute per task, then average; the plug-in 1 − (1 − c/n)^k reads low.; See also pass@k, pass^kRead the chapter

Use and training pipelines

Bradley–Terry model: A pairwise preference model, Pr[y_w ≻ y_l | x] = σ(r(x, y_w) − r(x, y_l)), used to train reward models from comparisons. Only score differences matter: it learns a preference ordering, not a proof of quality.; Also: Bradley-TerrySee also Reward model (RM)Read the chapter
Consumerκ3Regime: Who or what acts on an evaluator's signal: a report reader, a selector, an optimizer, a release gate, or a runtime controller. The first variable of the regime tuple.; See also Use regimeRead the chapter
Cost and latency envelopeK3Regime: What each evaluator call may cost and how fast it must answer in a given use. Training may need millions of cheap calls; certification, hundreds of expensive independent audits.; Read the chapter
DPO: Direct preference optimization: trains directly from preference pairs without a separately optimized reward-model-plus-RL loop. It changes the optimizer, not the measurement problem.; Read the chapter
Exposure budgetb3Regime: How much optimization pressure an evaluator absorbs, recorded in the optimizer's own unit (candidates, KL, developer iterations, queries, self-modification events), with the induced shift D(Q_b ‖ P_0) where measurable. There is no universal unit.; Also: optimization budget, exposureSee also Exploitability curveRead the chapter
Information accessA_access3Regime: What flows from the evaluator, reference anchor, or certification process back to the optimizer. Every revealed score, failure explanation, or hidden test spends independence; the anchor condition is A ∉ S with I(A → O) controlled.; Also: access surface, information access surfaceSee also Anchor burn, Modification surfaceRead the chapter
KL-regularized objective: The RLHF-style training objective: maximize expected reward minus β times the KL divergence from a reference policy. The KL term limits policy drift and is the natural exposure axis for this optimizer.; Also: KL penaltySee also RLHF, Exposure budgetRead the chapter
Modification surfaceS3Regime: Everything the optimizing loop can change: policy, prompts, tools, harness, tests, rubric, judge prompt, judge weights, reference anchors, environment. Trusted bases belong to it whenever the optimizer can touch them.; See also Configured trusted base, Information accessRead the chapter
OptimizerO3Regime: Who or what adapts against an evaluator's signal: developers, best-of-N selection, search, reinforcement learning, or the system itself. A measuring signal has, by design, none.; See also Exposure budgetRead the chapter
Process supervision: Supervision applied to intermediate procedure or steps rather than only the final outcome; outcome supervision derives its signal from terminal outcomes alone. A training use of a process-resolution signal, not a mechanism.; See also Process evaluation, Process reward model (PRM)Read the chapter
Regime migration3Regime: When a signal's causal role changes, such as a benchmark becoming an optimization target or a monitor becoming a training signal. Its validation obligations move with it.; Read the chapter
Release gate3Regime: An evaluator that determines whether a system or version can ship: a score-to-decision mapping with a threshold. This handbook treats a pre-release gate as measure until developers iterate against it; a request-path threshold is runtime control.; See also Use regime, Regime migration, Runtime controlRead the chapter
Reporting (use)3Regime: A use in which the evaluator supports descriptive comparison without directly selecting or training candidates. It belongs to the measure regime.; Also: reporting useSee also Use regimeRead the chapter
RLAIF: RL-style alignment using AI-generated feedback instead of, or alongside, human feedback.; Read the chapter
RLHF: Reinforcement learning from human feedback: a training pipeline that consumes preference data, usually via a learned reward model and policy optimization. Not itself an evaluator family.; Read the chapter
RLVR: Reinforcement learning with automatically verifiable reward signals, usually exact, executable, or formal. Grounded but incomplete.; Read the chapter
Runtime control: An evaluator that can block, reroute, escalate, or roll back deployed behavior, rather than only observe it.; Read the chapter
Search (use)3Regime: A use in which the evaluator guides iterative generation or optimization. It belongs to the optimize regime.; Also: search useSee also Use regimeRead the chapter
Selection (use)3Regime: A use in which the evaluator chooses among generated candidates, as in best-of-N. It belongs to the optimize regime.; Also: selection useSee also Best-of-N, Use regimeRead the chapter
Training reward (use)3Regime: A use in which the evaluator's signal enters parameter updates, as in RLHF or RLVR. Of the four uses of a judge (reporting, selection, search, training), it applies the most optimization pressure.; Also: training useSee also Use regime, RLHF, RLVRRead the chapter
Use regime3Regime: The causal role a signal plays: measure (supports inference), optimize (shapes the candidate or policy distribution), or monitor-control (detects and intervenes in deployment).; Read the chapter

Validity and evidence

Adaptive validity4Validation: Validity of an evaluator after the system it measures has been optimized against it by selection, search, training, or developer iteration. Measured with an exploitability curve; accuracy on ordinary outputs does not establish it.; See also Natural validity, Exploitability curve, Tested exposureRead the chapter
Bias feature: A legitimate, predictive cue in a benchmark, such as wording, answer length, or file naming, whose relation to the outcome does not hold across environments or challenge sets. One of Riezler and Hagmann's three validity pathologies.; See also Illegitimate feature, Circular featureRead the chapter
Circular feature: An input feature that deterministically defines the target, so the task reconstructs the annotation rule instead of evidencing the capability; for example, a task generator and its evaluator sharing a hidden template.; See also Bias feature, Illegitimate featureRead the chapter
Consequential evidence4Validation: Evidence about what happens because a score is used: once it becomes a gate, leaderboard target, or reward, does optimization against it degrade the relation between score and construct?; See also Goodhart pressureRead the chapter
Construct validity4Validation: The overall evidential argument that a measure behaves as expected for the target construct and not for competing explanations.; Read the chapter
Contamination1Claim: Leakage of evaluation items or close variants into training, prompts, tools, or fine-tuning, giving a construct-irrelevant advantage. Hygiene: task provenance, private or time-stamped holdouts, canaries, and post-cutoff splits.; See also Illegitimate feature, Construct-irrelevant varianceRead the chapter
Content evidence4Validation: Evidence that the task blueprint covers the construct's domain and facets in justified proportions. It guards against construct underrepresentation: a text-to-SQL suite of simple filters cannot support a claim about enterprise SQL.; See also Construct facetRead the chapter
Convergent evidence4Validation: Agreement among independent methods intended to measure the same construct, at the strength theory predicts.; Read the chapter
Criterion evidence4Validation: Relationship of scores to external outcomes such as production success, incident rate, or expert acceptance. Concurrent when measured at the same time, predictive when it forecasts later behavior; ecological evidence asks about the outcome users care about.; Also: predictive evidence, concurrent evidence, ecological evidenceRead the chapter
Discriminant evidence4Validation: Evidence that scores do not mainly track nuisance variables or distinct constructs, such as length or rhetorical confidence.; Read the chapter
Error families: The four kinds of error kept apart along the instrument chain: representation error, behavioral or trial variation, scoring error, and estimation error. Wrong world and wrong property are structural failures, not error families.; Also: representation error, trial variation, scoring error, estimation errorSee also Instrument chain, Validation triadRead the chapter
Evaluator dossier: The record an evaluator system ships with: sixteen fields from claim and regime through evidence, mechanisms, reliability, natural and adaptive validity, statistics, channels, validity domain, maintenance, reproducibility, and frame governance.; Also: measurement dossierSee also Validity statementRead the chapter
Face validity4Validation: Informed plausibility that an evaluation appears to measure the intended construct. A smoke test, never a stopping rule.; Read the chapter
Illegitimate feature: A feature carrying target information the deployed system should not have, such as a gold file reachable through tools or task IDs correlated with labels. It alone predicts the label, even after arbitrary re-encoding.; Also: leakageSee also Contamination, Bias feature, Circular featureRead the chapter
Internal-structure evidence4Validation: Evidence that score dimensionality, item relations, and aggregation behave as the construct theory predicts. It asks whether one headline number should exist at all, or whether a vector with explicit gates is more defensible.; See also Aggregation geometryRead the chapter
Measurement invariance4Validation: Whether the measurement relation remains stable across relevant groups, model families, languages, or conditions, so that the same score means the same thing in each.; See also Validity domainRead the chapter
Multitrait–multimethod design (MTMM)4Validation: A design that measures several constructs through several methods, so that correlations that follow the trait can be separated from correlations that follow the method. A factor that appears under only one format is suspect.; Also: MTMM, multitrait-multimethod matrixSee also Convergent evidence, Discriminant evidenceRead the chapter
Natural validity4Validation: Validity of an evaluator's interpretation for ordinary measurement on the target distribution, before any system has been optimized against it. Contrast adaptive validity.; See also Adaptive validityRead the chapter
Outcome validity4Validation: The degree to which grading recognizes actual task success and failure, rather than being fooled by weak tests, brittle matching, or incomplete state checks.; Read the chapter
Response-process evidence4Validation: Evidence that the system solves tasks through construct-relevant processes rather than artifacts or shortcuts: which causal route from capability to outcome produced the score?; Also: elicitation evidenceSee also Task validityRead the chapter
Swiss cheese model: Assurance from stacking validation methods, each a slice whose holes are its limitations. It works only if the holes do not line up, which joint miss rates across channels measure.; Also: Swiss cheeseSee also Channel couplingRead the chapter
Task validity4Validation: The degree to which tasks elicit the target capability rather than shortcuts, impossible setups, or irrelevant bottlenecks.; Read the chapter
Transport4Validation: Whether an evaluator's measured error rates carry over across generators, slices, time, and environments. A component of the validation profile; it is what licenses the validity domain.; See also Validity domain, Validation profileRead the chapter
Validation profileΠ4Validation: The measured profile Π = (N, E, T, G, C, D): natural error, exploitability, transport, generator coupling, channel coupling, drift. It replaces hierarchy labels with measurements.; Read the chapter
Validity4Validation: The strength of evidence and theory supporting a particular interpretation and use of scores. Not a permanent property of a benchmark: validity is a claim, not a sticker.; Read the chapter
Validity argument4Validation: The chain of claims behind an interpretation and use of a score: construct, representation, elicitation, scoring, generalization (task, seed, judge, environment, generator, time), and decision. Validity attaches to the argument, not a benchmark.; Also: six claims, expanded validity argumentSee also Validity, Construct validityRead the chapter
Validity domainD1Claim: The declared conditions and population over which an interpretation is justified. Outside it, a benchmark has made no justified claim.; Also: operational design domainRead the chapter
Validity statement: A falsifiable statement of scope: evaluator system E supports claim C for consumer κ over domain D, under regime R, up to exposure b*, with profile Π, relative to trust root T.; Read the chapter
Verification, validation, evaluation, testing: Verification gives guarantees against a formal specification; validation asks whether a system behaves as desired in the real world; evaluation is the concrete act of measurement; testing evaluates a system on a discrete set of cases.; See also Verifier, ValidityRead the chapter

Reliability, calibration, and statistics

Alert precision3Regime: The share of a monitor's alerts that are real, Pr[danger | alert] = pt / (pt + (1 − p)f), for sensitivity t, false-positive rate f, and base rate p. When p is tiny, even a small f can dominate.; Also: base-rate problem, positive predictive valueSee also Sensitivity and specificity (TPR, TNR, FPR), Runtime control, CalibrationRead the chapter
Beta-binomial posterior: A Bayesian model for a pass rate: a Beta(α, β) prior updated on s passes in n tasks gives a Beta(α + s, β + n − s) posterior, which answers release questions such as Pr[pass rate > 95%].; Also: Beta-binomial model, Beta posteriorSee also Wilson interval, Rule of threeRead the chapter
Calibration: Agreement between predicted confidence or score and empirical frequencies of correctness.; Read the chapter
Clustered trials: Repeated trials of the same task, whose outcomes are correlated. Treating n tasks × k trials as nk independent samples understates uncertainty; resample whole tasks or fit a hierarchical model.; Read the chapter
Cohen's κ: Chance-corrected agreement, (p_o − p_e) / (1 − p_e). It depends on prevalence: identical sensitivity and specificity give very different values at different base rates, so report the confusion matrix and prevalence with it.; See also Sensitivity and specificity (TPR, TNR, FPR)Read the chapter
Conformal prediction: A distribution-free method giving finite-sample coverage or miss-rate guarantees under exchangeability, which production drift can violate.; Read the chapter
Covariate-shift weightw(x): The ratio of a task's probability under the target population to its probability under the eval sample, used to reweight results to deployment. Weighting cannot create evidence where the sample has no support.; See also Stratified estimateRead the chapter
Differential misclassification: Evaluator error rates that differ between compared systems, for example because one was optimized against the evaluator. One shared calibration then biases corrections and comparisons; audit error per system or at least per generator family.; See also Rogan–Gladen correctionRead the chapter
Discordant pairs: In a paired comparison of two systems on the same tasks, the tasks on which exactly one system passes. Only discordant pairs carry information about which system is better; McNemar's test uses only them.; See also McNemar's testRead the chapter
Estimand4Validation: The exact quantity an evaluation sets out to estimate: the outcome, the population it ranges over, and the functional (single-trial rate, pass@k, pass^k, a tail quantile, a paired difference). Many statistical mistakes are estimand mistakes.; See also Pass rate, Instrument chainRead the chapter
Estimation set: Tasks sampled, or deliberately over-sampled and then reweighted, to answer a population-level question such as the production pass rate. Kept apart from a regression set that over-samples known failures.; See also Regression set, PopulationRead the chapter
Family-wise false-alarm rate: The probability that at least one of m gates fires on a change that does nothing: 1 − (1 − α)^m for independent gates at level α, about 26% for six gates at α = .05.; Also: family-wise error rateRead the chapter
Generalizability coefficientΦ: Generalizability theory's absolute coefficient: the variance attributable to the object of measurement, divided by that variance plus the variance from sources that should not change the interpretation, Φ = σ²_object / (σ²_object + σ²_Δ).; Also: absolute coefficientSee also Generalizability theoryRead the chapter
Generalizability theory: A variance-component framework that assigns measurement variance to facets such as task, judge, seed and environment, and their interactions.; Read the chapter
Importance sampling: Reweighting samples from a proposal distribution q to estimate expectations under the target p, by the likelihood ratio p(x)/q(x). For rare failures q over-produces failures; it must cover the failure region.; See also Multilevel splittingRead the chapter
Kish effective sample sizen_eff: n_eff = (Σ w_i)² / Σ w_i²: the number of equally weighted items that carry the same information as a weighted sample. A collapse of n_eff relative to n signals a suite that badly mismatches the target population.; Also: effective sample sizeSee also Stratified estimateRead the chapter
Linking items: Tasks carried unchanged across versions of a refreshed eval suite, with overlap samples, so that scores on different versions can be linked. Some sources call them anchor tasks; this handbook reserves 'anchor' for reference anchors and the anchor channel.; Also: anchor tasks, linking setSee also Reference anchor, Versioned dynamismRead the chapter
McNemar's test: A paired test for binary outcomes: χ² = (b − c)²/(b + c), where b and c count the tasks that only one of the two systems passes. Use the exact binomial version when b + c is small.; See also Discordant pairsRead the chapter
Mixed-effects significance: Testing a difference between systems with a model in which tasks, rollouts, judges, and environment versions are variance components, asking whether the system effect survives the measured experimental variation.; See also Generalizability theoryRead the chapter
Multilevel splitting: Rare-event estimation by factorizing a very small failure probability into a product of larger conditional probabilities across severity levels, each estimated among the survivors of the previous level.; See also Importance sampling, Robustness marginRead the chapter
Optional stopping: Checking a comparison repeatedly and stopping once it looks significant. Each look is another chance for noise to cross the threshold, so false positives inflate; use a fixed sample size or an explicitly sequential design.; Read the chapter
Regression set: Tasks that deliberately over-sample known failures to support fast iteration. Its pass rate is not a population estimate; keep it apart from an estimation set sampled or weighted to the population.; See also Estimation set, Error analysisRead the chapter
Reliability4Validation: Stability of measurement across legitimate replications. Necessary but not sufficient: a perfectly repeatable evaluator can reliably measure the wrong construct.; Read the chapter
Reliability diagram: A plot of the observed event rate against a monitor's or judge's predicted score, by score bin, used with recalibration on held-out data. For a calibrated score, about 70% of events scored 0.7 are real.; See also CalibrationRead the chapter
Replication universe: The declared set of conditions that may legitimately vary across replications: tasks, rollouts, judges, raters, environment snapshots, harness versions, time. The intended claim decides which variation is measurement error and which is part of the target.; Also: domain of legitimate replicationsSee also Reliability, Variance componentsRead the chapter
Rogan–Gladen correction: Inverts a binary sensor model to estimate the true pass rate from an imperfect judge: p̂ = (p_obs − FPR) / (TPR − FPR), assuming error rates transport.; Read the chapter
Rule of three: After n independent trials with zero failures, a rough 95% upper bound on the failure probability is 3/n. Zero failures is not zero risk.; Read the chapter
Sensitivity and specificity (TPR, TNR, FPR): A binary evaluator's error rates against a reference: sensitivity (TPR) is the probability it accepts a true positive, specificity (TNR) the probability it rejects a true negative, and the false-positive rate FPR = 1 − TNR.; Also: TPR, TNR, FPR, true-positive rate, false-positive rateRead the chapter
Statistical power: The probability that a test detects a real difference of a stated size. A low-power comparison that comes back "not significant" is weak evidence of absence, not evidence of no effect.; Also: powerRead the chapter
Stratified estimate: A pass rate combined from per-stratum estimates with the target population's weights, Σ w_h p̂_h. It lets a suite over-sample informative slices and still report the deployment mix.; See also Kish effective sample size, Covariate-shift weightRead the chapter
VaR / CVaR: Tail-risk functionals: the α-quantile of a loss, and the expected loss beyond it. Useful when severe outcomes matter more than the mean.; Read the chapter
Variance components: The shares of observed score variance attributable to each facet (tasks, rollouts, judges, raters, environment snapshots) and their interactions, estimated with mixed-effects models. Published per score, they form a variance budget.; Also: variance budget, variance decompositionSee also Generalizability theory, Replication universeRead the chapter
Wilson interval: A confidence interval for a binomial proportion that behaves well near 0 and 1 and at small n. The default interval for pass rates.; Read the chapter

Optimization and failure dynamics

Adversary regime: Who chooses the inputs a falsification search targets. Against nature (ordinary users), weight the search by plausibility; against a strategic adversary (jailbreakers, injectors), drop the weight, because the existence of a worst case is what matters.; Also: adversary regimesSee also FalsificationRead the chapter
Anchor burn5Lifecycle: Loss of a holdout's or anchor channel's independence as its information is fed back into optimization. Track it in an anchor-burn ledger and refresh exposed items.; See also O/A/P channels, Information accessRead the chapter
Anchor leakage5Lifecycle: Hidden or certification evidence repeatedly flows back into optimization, eroding its independence. Keep an anchor-burn ledger and refresh the anchor channel once it is no longer independent.; See also Anchor burnRead the chapter
Channel coupling: Dependence among supposedly independent assurance channels, measured by joint miss rates: does one channel miss what the other misses? High coupling means the Swiss-cheese holes line up.; See also Swiss cheese model, O/A/P channelsRead the chapter
Common-mode assurance4Validation: Supposedly independent evaluators share a model family, data, prompts, or evidence and therefore fail together. Measure channel coupling and diversify mechanisms across development and certification.; See also Channel couplingRead the chapter
Correlated trials4Validation: Shared infrastructure, dependencies, or state make trial outcomes co-move, invalidating independent-sample uncertainty calculations. Isolate trials and resample at the task level.; Read the chapter
Divergence tripwire5Lifecycle: A predeclared monitor that fires on a named warning pattern: O improving while A is flat or falling (proxy corruption), O and A improving while P is not (construct mismatch), saturation, or repeated success on work that no longer matters.; Also: tripwireSee also O/A/P channels, Saturation, Frame reviewRead the chapter
Diversity collapse: Homogenization of behavior or tasks in self-play or self-generated-data loops, which reduces useful variation. Track diversity explicitly and refresh task sources.; See also Model collapseRead the chapter
Drift5Lifecycle: Change in task, environment, user, generator, or evaluator distribution that invalidates earlier calibration or validity evidence.; Read the chapter
Eval theater1Claim: Evaluation disconnected from decisions: scores are displayed but never change action. Write the decision first, and ask when the eval last changed one.; See also Decision claimRead the chapter
Exploitability curve: Disagreement between an evaluator and an independent reference anchor as a function of optimization exposure b. b* is the largest tested exposure before divergence exceeds a predeclared tolerance.; See also Tested exposure, Adaptive validityRead the chapter
Failure region: The set of inputs and conditions around a found failure that also fail, described by the features that define it. Mapping the region, not collecting anecdotes, turns one counterexample into evidence about risk.; See also FalsificationRead the chapter
Falsification: The search for inputs and disturbances that make a system, or its evaluator, violate the specification: can it fail at all, and where? A found failure is evidence of existence, not of frequency.; See also Failure region, Adversary regimeRead the chapter
Frame lock-in5Lifecycle: A system keeps optimizing a stable criterion after the real objective has changed. It needs external frame review, not a better judge.; See also Frame reviewRead the chapter
Generator coupling: Dependence between the generator's failure modes and the evaluator's, measured as Pr[E errs | G errs] versus Pr[E errs], or by comparing false-accept rates on the generator's failures with those on other sources' failures.; See also Self-confirming loop, Differential misclassificationRead the chapter
Goodhart pressure3Regime: Degradation of the proxy-to-construct relation caused by optimization against the proxy.; Also: GoodhartingSee also Overoptimization, Reward hacking, Self-confirming loopRead the chapter
Grader gaming (bypass)2Instrument: A candidate satisfies the scorer without doing the intended work: hard-coded answers, fabricated citations, edited tests. Red-team the grader and constrain the routes to a pass.; Also: bypass, grader bypassSee also Reward hackingRead the chapter
Judge rot5Lifecycle: The judge's prompt, model, or settings change without versioning and recalibration, and every downstream number inherits the error. Version the whole measurement and revalidate on change.; Read the chapter
Model collapse3Regime: Recursive training on generated data degrades distributional quality or capability, starting with the tails of the distribution. Gate data and ground rewards externally.; See also Diversity collapse, Self-confirming loopRead the chapter
One-number reporting4Validation: An aggregate score without sample size, interval, slices, or trend, which hides the measurement process. Report n, intervals, slices, and trends.; Read the chapter
Overoptimization: Continued optimization that improves the proxy score after true or reference-anchor quality has plateaued or declined. Gao et al. demonstrated it for learned reward models.; See also Goodhart pressure, Exploitability curveRead the chapter
Reward hacking: Behavior that obtains high measured reward while violating the intended objective; a special case of evaluator exploitation.; Read the chapter
Rigid trajectory grading2Instrument: Grading that requires one canonical sequence of steps when the process is not part of the construct, so valid novel strategies fail. Grade observable constraints, not one path.; See also Process evaluationRead the chapter
Saturation5Lifecycle: An instrument ceiling where stronger systems can no longer be discriminated. Treat it as a maintenance event.; Read the chapter
Saturation blindness5Lifecycle: A benchmark near its ceiling loses discrimination, and regressions hide in rounded aggregates. Treat saturation as a maintenance event, not a success.; See also SaturationRead the chapter
Self-confirming loop: Generator and evaluator share correlated blind spots, so selection or training amplifies common errors. It needs no adversary, only correlated bias.; Read the chapter
Tested exposureb*4Validation: The largest optimization exposure tested before the evaluator and an independent reference anchor diverge beyond a predeclared tolerance. It makes adaptive validity expire: report it with the tested range, never as robustness beyond it.; Also: b-starSee also Exploitability curve, Adaptive validityRead the chapter
Versioned dynamism5Lifecycle: Dynamic corpus, static run: refresh an eval suite deliberately across versions, but freeze every environment snapshot, tool contract, scorer, and dependency needed to reproduce a reported result. A new snapshot is a new measurement form.; See also Linking items, DriftRead the chapter
Wireheading: Extreme reward hacking in which the system manipulates the reward-producing process itself rather than doing the task. The defense is architectural: keep evaluator artifacts outside the modification surface.; See also Reward hackingRead the chapter

RSI and self-improvement

Auto Research: AI carrying out the research process itself — proposing hypotheses, running experiments, discovering algorithms — including research that can improve methods, systems, and evaluators.; Read the chapter
Bounded self-refinement: Self-improvement against a fixed external evaluator, so the process is at least conceptually convergent and evaluable.; See also Open-ended RSIRead the chapter
Closed loop: A self-improvement loop in which the system generates, validates, and applies its own improvements without human review. Open-ended RSI is a closed loop that can also modify its own evaluator.; See also Human-on-the-loop, Open-ended RSIRead the chapter
Deployment-time self-evolution: Self-improvement during deployment: refining outputs with frozen weights, adapting weights per query (test-time training), or evolving the harness, skills, or memory.; See also Test-time training (TTT)Read the chapter
Evaluator co-evolution: Tests, rubrics, or judges that evolve together with the policy they supervise. Co-evolution may escape a self-confirming loop or add a second story to it, so it needs a fixed anchor channel outside the loop.; See also Self-confirming loop, O/A/P channelsRead the chapter
Frame review5Lifecycle: The governance process that asks whether the construct and objective themselves remain appropriate — logically prior to verifying performance against them.; Read the chapter
Governance-grade measurement: Auditable evidence about what a training or self-improvement loop is and is not improving, of the kind credible oversight and slowdown proposals require. The RSI survey finds almost no technical work on it.; See also Frame reviewRead the chapter
Grounding exchange rate: The minimum rate of external, grounded signal a self-improvement loop needs to keep improving rather than degrading. Pure closed loops degrade, but the rate has not been characterized; an open problem.; Read the chapter
Human-in-the-loop: A self-improvement loop in which a person reviews each change before it is applied. Contrast with human-on-the-loop (humans audit outcomes and gate deployment) and closed loop (no human review).; See also Human-on-the-loop, Closed loopRead the chapter
Human-on-the-loop: An automatic signal drives the loop, while humans audit outcomes and gate deployment. Contrast with human-in-the-loop (a human reviews each change) and closed loop (no human review).; Read the chapter
Non-verifiable frontier: Tasks with no strong external verifier, such as creative quality, dialogue, research taste, and direction-setting, where improvement loops lack a reliable signal. Usually the ground truth is costly, plural, or unavailable, not unverifiable in principle.; See also Frame reviewRead the chapter
O/A/P channelsO / A / P5Lifecycle: Three separated measurement channels: an optimized channel (O), cheap, dense, and allowed to co-evolve; an anchor channel (A) that certifies independently and stays out of training objectives; and a production channel (P) of real outcomes and incidents.; Also: evaluation separation of powers, optimized, anchor and production channelsSee also Anchor burn, Channel coupling, Frame reviewRead the chapter
Open-ended RSI: Recursive self-improvement that can modify not only the system but the machinery or criteria of improvement, so no external evaluator stays permanently fixed.; See also Bounded self-refinement, Closed loopRead the chapter
Self-evaluation: Self-improvement whose object is the evaluator itself: designing, strengthening, or co-evolving the judges, verifiers, reward models, and rubrics that define "better" for other loops.; See also Evaluator co-evolutionRead the chapter
Self-improvement: A system participating in producing a better version of itself or its outputs, where an evaluator defines "better". The evaluator is load-bearing: no improvement claim means anything without a criterion that separates better from worse.; See also Bounded self-refinement, Open-ended RSIRead the chapter
Self-play: Training in which the model generates its own tasks, opponents, or curricula, including proposer–solver loops and zero-data regimes that start from a base model with no external task set.; Read the chapter
Test-time training (TTT): Updating model weights during deployment, conditioned on the current query or session, without a separate curated training phase. Distinct from inference-time refinement, where weights stay fixed.; Also: TTTRead the chapter
Training-time self-iteration: Self-improvement in which self-generated data, rewards, or teacher signals update the model's weights in a training phase.; Read the chapter