Status: Generated reference · Definitions reviewed 2026-10-25
Use this page when a term stops you understanding a sentence. The context matters: the same word or letter can have different meanings in different experiments.
For arithmetic, read the formula guide. For the overall idea, read the plain-English tour. Scientific sources explains the research connections; current ISC status gives results rather than definitions.
168 contextual definitions, sorted alphabetically. Examples explain the term; they are not extra measured results.
| Term and context | Meaning | Why it matters in practice | Example |
|---|---|---|---|
| Ablation (Tests, controls and claims) | Disable a component for comparison | Asks whether that component matters to the outcome | A quality loss after removal can support a bounded causal claim |
| Action (Using a browser or screen) | A click, key or other interface operation | Attempts a state change | A dispatched action can fail to affect the intended object |
| Activation (Language models and internal steering) | A value computed inside a forward pass | Can carry an intervention into generation | A change at one layer may have a different effect from the same change elsewhere |
| Anchor (ISC) | A saved comparison moment from which experimental arms begin. | Accepted anchors, collected anchors and independent observations are different counts. | A collected moment can be unusable if one required arm is unreadable. |
| Argmax (Mathematical claim limits) | The choice whose score is greatest among the candidates considered. | Scores can change without changing the winner. | Scores (10, 1) and (9, 2) choose the first action both times. |
| Arm (Experiment) | One treatment or control path in a comparison. | Arms must share the intended starting conditions without sharing later residue. | One copy receives an intervention; another is left unchanged. |
| Arousal (Affect and operating condition) | An internal activation or intensity coordinate | Can change the strength or urgency of processing | It is separate from whether a condition is positive or negative |
| Artifact (Logs, measurements and evidence trails) | A retained output used as evidence | Allows later inspection or reproduction | A filename does not prove that the file contains measured data |
| Assumption (Engineering calculations and drawings) | A condition taken as given | Defines where the result applies | Changing the load case can invalidate the result |
| Authority (Permission, actions and receipts) | The rule or actor entitled to make a decision | Defines who can permit a consequence | A generated plan cannot grant itself new authority |
| Autocorrelation (Statistics) | A signal resembles its own earlier values. | Closely related samples can carry less independent information than their count suggests. | Memory usage in adjacent seconds is often similar. |
| Backpressure (Memory, speed and resource limits) | A request to slow or yield when a consumer cannot keep up | Limits queues and competing work | An expected yield should not automatically be reported as a dead system |
| Baseline (Affect and operating condition) | A recorded reference level | Makes a change interpretable relative to ordinary state | Above usual and below usual need an actual reference |
| Baseline (Claims about the physical world) | The comparison without the proposed effect | Separates the result from ordinary variation | A warm room can heat a sensor without the device’s contribution |
| Bootstrap (Statistics) | Recalculate a statistic on resampled data to estimate variability. | The resampling unit and dependence assumptions affect the interval. | Resample the declared comparison units rather than treating every time frame as independent. |
| Calibration (Measurement) | Check whether scores or forecasts match observed outcomes under stated conditions. | A confident number needs evidence that its interpretation is justified. | Score a resource forecast when its deadline arrives. |
| Candidate (Attention and competing concerns) | A proposed item for attention | Enters a competition | A memory and an intention can have different content types |
| Capability (Permission, actions and receipts) | A scoped authorisation or declared ability, depending on context | Must be read with its scope | A token for one file is not unrestricted file access |
| Carrier (ISC and the subjecthood research) | The process proposed to carry the state being studied | Must be identified by the carrier experiment | A name or a passing synergy statistic does not select it |
| Causal claim (Experiment) | A claim about what changes because a factor is changed. | A correlation or a module name alone does not supply that comparison. | Changing internal valence changes the measured response under controls. |
| Causal closure (ISC) | The proposed boundary is adequate under the declared outside-information test and margin. | It is conditional on measurement coverage, estimator and environment. | An outside driver may improve prediction and keep the boundary open. |
| Certificate (Language models and internal steering) | A record of a measured basis under a stated protocol | Binds a channel claim to its tested conditions | A channel certificate can pass while installation is refused |
| Channel (Claims about the physical world) | The route intended to produce an effect | Connects a request with a possible action | A declared channel can still be unavailable |
| Checkpoint (Reading records and research plans) | A recorded stage of work or source history | Makes a change recoverable and reviewable | A commit is not a claim that every goal is complete |
| Checkpoint (Staying alive and recovering) | Saved state for continuing work | Can limit lost progress after interruption | Its presence does not prove that restoration has been tested |
| Closure (ISC and the subjecthood research) | Whether the proposed boundary omits an important internal driver after declared input is accounted for | Can remain unresolved rather than pass by default | K* now uses a declared numeric margin about 0.0196 plus an interventional closure test; the earlier least-domain rule stays withdrawn |
| Confidence interval (Statistics) | An interval produced by a procedure designed for stated repeated-sampling coverage. | Its interpretation depends on the procedure and assumptions. | An upper bound above the declared margin cannot establish closure within that margin. |
| Configuration (Getting started) | Settings read by the program | Only a setting with a reader can change behavior | An invented environment variable changes nothing |
| Conjunction (Decision rule) | Every required condition must hold. | A strong result on one condition does not replace a failed required condition. | Four bridge grounds do not replace an unresolved fifth. |
| Context window (Language models and internal steering) | The token budget available to a model request | Limits how much text fits at once | A memory store can be larger than the material placed in one request |
| Contraction (Mathematical claim limits) | An update that shrinks distances by a factor strictly below one on a declared domain. | A near-one cosine between two states does not establish contraction. | F(x)=2x preserves direction but doubles distances. |
| Control (Tests, controls and claims) | A comparison that reveals another explanation | Can expose effects of the measuring procedure | An unchanged copy should not look like a deliberate intervention |
| Control delay (Visual control and moving worlds) | Time between choosing an action and its effect | Can change whether a plan succeeds | A correct action chosen too late may miss |
| Core (ISC and the subjecthood research) | The selected group of measured state | Defines the proposed process boundary | Useful internal causes outside it can make the description incomplete |
| Counterfactual (Causal reasoning) | What would happen under an alternative stated intervention or condition. | It makes a dependence claim testable. | What would the target do if this connection were removed? |
| Coverage (Tests, controls and claims) | What cases or paths were actually checked | Limits generalisation | An offline unit test does not cover a production microphone |
| Credential (Privacy and trust boundaries) | A secret or token used to authenticate access | Must be handled within its authorised scope | A token name in documentation is not the secret value |
| Delta, Δ (Reading mathematical symbols) | A change or configured intervention | Shows a difference rather than an absolute level | Going from 10 to 13 gives a change of +3 |
| Denotation (Mathematical claim limits) | The mathematical function or meaning assigned to a program under specified semantics. | A shorter name can preserve meaning while changing search cost. | Naming a long calculation does not add a new computable function. |
| Dependency (Architecture and runtime components) | A component or value another part relies on | Explains what must be available | A broken dependency can cause a fallback rather than the intended mechanism |
| Depth (Reasoning and checking answers) | Additional stages or repeated computation | Trades time and resources for a possible reasoning benefit | A deeper pass still needs a measured correct-answer gain |
| Drive (Affect and operating condition) | A represented need or motivation | Contributes to proposed work or choices | A curiosity value should change a relevant choice to support a causal claim |
| Effect receipt (Claims about the physical world) | A record of independent post-action measurement | Supports a scoped outcome claim | It is stronger than a transport acknowledgement |
| Effect size (Tests, controls and claims) | How much a measured quantity changed | Shows practical magnitude | Statistical significance alone does not state a useful improvement |
| Embedding (Memory and recall) | A numerical representation used for similarity search | Helps find related material | Similar wording and similar meaning can produce different retrieval behavior |
| Epsilon lower bound (Mathematical claim limits) | A positive floor on success probability for each eligible repeated trial in a stated argument. | Some positive chance at each time is weaker than a common positive floor. | With a common floor e, missing after n trials has bound (1-e)^n. |
| Epsilon or margin (Reading mathematical symbols) | A declared acceptable residual or error bound, depending on context | Defines the precision of a claim | An unnamed margin cannot be inferred from a convenient observed result |
| Expected value (Learning and kept procedures) | Expected benefit minus relevant cost | Can decide whether a retained rule pays for itself | A rule that saves 5 units but costs 7 is not beneficial on that accounting |
| Expected value, EV (Reading mathematical symbols) | Expected benefit minus the relevant costs under a stated accounting | Helps decide whether a retained method is worth using | Benefit 5 and cost 7 give net value -2 |
| Experimental basis (Steering) | A measured direction or set of directions used for an intervention. | A basis used in an experiment is not automatically authorised for desktop serving. | A channel certificate can pass while an installation qualification fails. |
| Extensional property (Mathematical claim limits) | A property of what a program computes, rather than how its code is written. | Rice’s theorem needs this and other assumptions; it does not prohibit every program proof. | Two differently spelled programs can compute the same function. |
| Fail closed (Permission, actions and receipts) | Refuse or withhold when required evidence is absent | Prevents an unknown from silently becoming permission | A missing verifier is not a passing verifier |
| False positive (Measurement) | The procedure declares an effect when the relevant null condition holds. | Nulls and decision levels need a declared interpretation. | A noisy unchanged copy crosses the same threshold as a cut. |
| Falsifiable (Research) | The claim has a specified possible observation that would count against it. | A test that always passes cannot discriminate the hypothesis. | Declare the failure condition before measuring. |
| Feedback (Using a browser or screen) | The observed consequence | Checks whether the action worked | A changed page is evidence; a tool returning success is only one signal |
| Finding (Engineering calculations and drawings) | A calculated result carrying its derivation | Connects a drawing to evidence | A dimension should be traceable to its formula and inputs |
| Fingerprint or digest (Reading mathematical symbols) | A hash of specified data or configuration | Detects that an input or file changed | Two different campaign fingerprints must not silently become one result |
| Free energy (Model-dependent term) | A quantity defined by a particular inference model; its meaning needs the model and equation. | It is not automatically thermodynamic energy or a measure of felt discomfort. | Read the named implementation rather than inferring joules from the label. |
| Functional evidence (Evidence scope) | Evidence about computations, dependencies or behavior under specified tests. | It supports that mechanism without settling every interpretation of experience. | A state change reaches recall through a measured route. |
| GB and GiB (Memory, speed and resource limits) | Decimal and binary memory units: 1 GB is 1,000,000,000 bytes; 1 GiB is 1,073,741,824 bytes | Different tools can describe the same capacity with different numbers | 64 GiB is about 68.7 GB; it is not extra newly available memory |
| Generated document (Reading records and research plans) | A page rendered from another source | Must be updated through its renderer | Hand editing can be overwritten or detach it from its evidence |
| Goal (Goals and autonomous work) | A represented outcome to pursue | Organises further work | Completing a substep is different from completing the goal |
| Guide (Reading records and research plans) | An explanation intended to describe current use or concepts | Provides an entry point | Read its scope and current status links |
| Hash (Logs, measurements and evidence trails) | A fingerprint of specified bytes | Detects a changed file relative to a recorded digest | It does not establish that the original contents were correct |
| Held state (Current ISC method) | Declared attributes counted on the core side of the K* closure test. | They must not also be treated as outside predictors. | Readings under held. move with the domain they serve during a cut. |
| Held-out task (Reasoning and checking answers) | A task excluded from fitting or method selection | Checks transfer beyond the development examples | Success on a new task is stronger than replaying a fitted example |
| Historical record (Reading records and research plans) | A dated account of a run or decision | Preserves provenance | A later failure or success should be a new record |
| Hypothesis (Research) | A proposed explanation with stated predictions and scope. | The comparison must distinguish it from an alternative. | The internal state matters versus the state is unused commentary. |
| Identity continuity (Visual control and moving worlds) | Evidence that observations refer to the same object | Prevents tracking a distractor | Motion alone does not prove which object the agent controls |
| IIT (Theory) | Integrated information theory: a specific formal account of intrinsic causal structure and experience. | A similarly named graph score is not automatically the theory's full calculation. | A Gaussian integration estimate needs its own estimator label. |
| Impasse (Attention and competing concerns) | A decision that lacks a justified winner | Creates a reason for further work | An exact tie should not be disguised as a meaningful preference |
| Initiative (Goals and autonomous work) | A proposed action or task | Can enter selection or governance | A proposal can be refused or deferred |
| Input (Engineering calculations and drawings) | A supplied quantity with a defined meaning and unit | Feeds the calculation | A length in metres cannot silently become millimetres |
| Intervention (Experiment) | A deliberate, declared change to a measured system. | Its writer, size, duration and consumers determine what was tested. | Change one source domain while keeping the other declared conditions fixed. |
| Interventional closure (Current ISC method) | The specified outside-state swaps leave the measured core and held state unchanged at the tested anchors and horizon. | It checks actual effects beyond the observational prediction comparison. | The current rule checks bit-for-bit equality within two turns when untouched twins agree. |
| Invariant (Architecture and runtime components) | A condition required to stay true | Can identify the stage that broke an assumption | A checker that fails to run is not evidence the condition held |
| J-star (ISC bridge) | The combined description of carrier, content relations and causal lineage. | It is a structured identification, not a scalar score to maximise. | No carrier means the other recorded grounds cannot manufacture J-star. |
| K* (Current ISC method) | The proposed core K plus declared held-state dependencies its transition reads. | It changes which readings count as inside the candidate boundary. | An identified history buffer is included alongside the original ten-domain vector. |
| KV cache (Model runtime) | Stored attention keys and values reused during generation. | Cache ownership, context size and invalidation affect memory and correctness. | A longer conversation can use more cache memory even with fixed model weights. |
| Lambda (Reading mathematical symbols) | A context-dependent coefficient or residual-control symbol | Must be read from the local definition | The same letter can mean a decay setting in one model and an outside-control quantity in another |
| Latency (Memory, speed and resource limits) | Elapsed time for a defined step | Measures delay experienced by a user or pipeline | State the start and end points, not just one undifferentiated speed number |
| Latent state (Reasoning and checking answers) | Internal numerical representation rather than generated words | Can carry computation without printing each step | A latent operation needs evidence that its output affects the answer |
| Likelihood ratio (Probability) | The evidence probability under one hypothesis divided by that under another. | A project score is not a calibrated likelihood ratio merely because it has that label. | Identical stipulated observation distributions give ratio one. |
| Liveness (Staying alive and recovering) | Whether a component is still making the required kind of progress | Helps distinguish stalled from merely slow | A heartbeat alone may not show that a request is advancing |
| Log (Logs, measurements and evidence trails) | A recorded event or message | Retains context for diagnosis | A message saying success is weaker than a verified outcome |
| Logit (Language model) | A model score before conversion into a token probability. | A logit is not already a probability. | Temperature rescales relative logits before sampling. |
| Matrix, often W (Reading mathematical symbols) | A table of numbers used in a transformation | Can map one numerical representation to another | In the ISC domain labels, W instead means world model |
| Mean (Reading mathematical symbols) | The arithmetic average | Summarises a collection while hiding variation between its entries | Changes +1 and +5 have mean +3 |
| Metric (Logs, measurements and evidence trails) | A declared numerical measurement | Supports comparison over time | Its unit, scope and missing-value behavior matter |
| MIP (Integration search) | Minimum-information partition under a specified objective and search. | A candidate cut and a certified minimum are different claims. | Exhaustive search can certify the small-grain minimum. |
| Model lane (Getting started) | A route to a particular model or model service | The route can change resource needs and what answers a request | A smaller configured lane may answer while a larger one is unavailable |
| Momentum (Affect and operating condition) | How a state carries forward while new input arrives | Controls persistence and settling | A value overwritten immediately cannot carry a sustained intervention |
| Non-inferiority (Mathematical claim limits) | A comparison asking whether a treatment is no worse than a baseline by more than a predeclared acceptable margin. | Overlapping confidence intervals do not establish it. | A declared loss margin needs an appropriate comparison, not a visual overlap. |
| Norm (Mathematics) | A declared way to measure the size of a vector. | Different coordinate scales can change an unnormalised distance. | The Euclidean norm of (3, 4) is 5. |
| Null (Experiment) | The specified comparison model or condition representing the alternative being tested. | There is no universal null that rules out every explanation. | A shuffled comparison destroys a selected relationship while preserving others. |
| Observable (Claims about the physical world) | The quantity to be measured | Defines what success would change | Temperature requires a declared sensor and unit |
| Observation (Using a browser or screen) | The interface information actually read | Constrains what the agent can infer | An old screenshot may no longer describe the page |
| Observational closure (Current ISC method) | The outside predictor adds no more than the permitted contribution under the specified estimator and controls. | A bounded predictive result needs a margin, uncertainty and measurement validity. | The upper bound must clear the named margin; a failed estimator fold cannot support closure. |
| Outcome (Goals and autonomous work) | What was observed after acting | Feeds completion and learning decisions | A successful tool call need not mean the real task succeeded |
| p value (Reading mathematical symbols) | The extremeness of a statistic under the declared null comparison | Belongs to one test and its assumptions | It is not the probability of consciousness |
| P-value (Statistics) | How extreme the statistic is under the declared null procedure and assumptions. | It is not the probability of consciousness or the probability that the hypothesis is false. | Zero exceedances among finite random draws still does not produce p = 0. |
| Parameter count, such as 27B (Language models and internal steering) | The approximate number of stored model parameters; B means billion | Describes model size, not an intelligence or consciousness score | A 27B model can have different memory needs under different numeric formats |
| Parity (ISC and the subjecthood research) | The project’s five-ground bridge conjunction under named assumptions | Organises supporting evidence | It is not an independently calibrated probability of experience |
| Periphery (ISC) | Recorded internal information outside the selected core. | Unmeasured outside state limits how broad a closure claim can be. | A service record can be relevant even though it is outside the ten domain vectors. |
| Permission (Privacy and trust boundaries) | A rule authorising a defined action | Limits consequences | Access to read does not imply authority to publish |
| Phase (Architecture and runtime components) | One stage of processing | Locates when a state is read or written | Order can determine whether a change reaches the current turn |
| Phenomenal experience (Consciousness discussion) | Subjective experience: whether there is something it is like for the system. | A software label or operational pass is not an assumption-free verdict on it. | A report can be causally grounded while its experiential interpretation needs an argument. |
| Phi (Reading mathematical symbols) | A name used for several integration-related measurements | Its formula, state description and controls determine its meaning | Different phi implementations are not interchangeable consciousness meters |
| Playback drain (Conversation and voice) | Time needed to finish audio already queued | Separates generation time from what the listener hears | The model may be finished while speech is still playing |
| Position (Visual control and moving worlds) | Where the identified object is measured | Supports the next control decision | An uncertain boundary should remain an uncertainty |
| Preregistration (Reading records and research plans) | Rules recorded before the result is read | Prevents changing a bar to fit an outcome | A revised method starts a new protocol |
| Priority (Goals and autonomous work) | Relative importance under a stated rule | Determines which work is considered next | A high priority does not bypass a required permission |
| Procedure (Learning and kept procedures) | Executable steps retained for a class of problems | Can replace repeated derivation | A solver may be kept only after matching required known answers |
| Prompt injection (Privacy and trust boundaries) | Content trying to redirect the assistant’s instructions | Must not acquire the authority of the person’s request | A downloaded document can contain adversarial instructions |
| Provenance (Memory and recall) | Where a record came from and how it changed | Lets a reader assess and correct it | An observation and an inference should retain different sources |
| q value (Reading mathematical symbols) | A multiple-testing-adjusted statistical quantity under its declared procedure | Helps control a stated error rate across many comparisons | Checking many candidate edges needs a declared family-level rule |
| Qualia (Philosophy and project names) | Qualitative aspects of experience; in code, often a label for a descriptor mechanism. | The implemented descriptor and the experiential claim must be distinguished. | A QualiaDescriptor can be inspected without establishing felt experience. |
| Quantization (Language models and internal steering) | Storing or computing numbers with reduced precision | Can reduce model memory and change quality or speed | A lower bit count does not guarantee the best practical result |
| Rank correlation (Statistics) | Agreement in the ordering of paired measurements. | It can show a monotonic relationship without showing causation. | Ranks (1, 2, 3) versus (1, 3, 2) give rho = 0.5. |
| Readiness (Staying alive and recovering) | Whether it can currently accept its intended work | Protects callers during startup or degradation | A living process can remain unready |
| Ready (Getting started) | The declared startup requirements have been met | Allows work to enter that route | A process can be running before its model is ready |
| Receipt (Permission, actions and receipts) | A structured record of a decision or effect | Makes actions inspectable | Transport success and independently measured outcome may differ |
| Refit (Estimation) | Estimate model parameters again using the declared data split. | Changing a refit rule after seeing a result can change what the experiment means. | Keep a baseline whitening transform fixed across compared arms. |
| Replication (Research) | Repeat the defined measurement on fresh data or an independently executed setup. | Reprinting the same artifact is not a new replication. | A fresh seed tests a method chosen on discovery data. |
| Resolution (Measurement) | The scale of effect the method has demonstrated it can detect under the stated conditions. | Not seeing an effect below that scale can leave the question unresolved. | A detector resolving 0.05 cannot support a 0.03 closure claim from non-detection. |
| Retention (Memory and recall) | Rules for keeping or deleting records | Controls continuity, storage and privacy | Deleting a source should have defined consequences for derived records |
| Retrieval score (Memory and recall) | The ranking produced by a particular search method | Orders candidates for recall | It is not a probability that the record is true |
| rho (Reading mathematical symbols) | A correlation coefficient; ISC uses ranked comparisons in named analyses | Describes how two sets of values or changes vary together | High agreement does not by itself establish one caused the other |
| RSS (Memory, speed and resource limits) | Memory resident for a process | Shows part of its current memory footprint | Adding process RSS can double-count shared memory |
| Salience (Attention and competing concerns) | A contribution to how much an item stands out | Can change its chance of being selected | An urgent observation can outrank a routine item |
| Sampling temperature (Language models and internal steering) | A setting affecting the distribution used to choose tokens | Can alter variability in generated text | It is different from the computer’s physical temperature |
| Service (Architecture and runtime components) | A component made available through a defined interface | Gives other components a declared dependency | A registered service still needs a caller |
| Sham (Experiment) | Repeat the intervention procedure without the intended change. | It checks whether the procedure itself creates an apparent effect. | Four untouched forks provide the matched synergy sham. |
| Spread or standard deviation, often sigma (Reading mathematical symbols) | How much values vary around a reference | Makes an intervention interpretable relative to ordinary variation | A change of 1 is large if ordinary variation is 0.1 and small if it is 10 |
| SSE (Prediction metric) | Sum of squared prediction errors over the declared cases and outputs. | Pooling errors is different from averaging arbitrary fold percentages. | Core error 100 versus wider-predictor error 80 gives a 0.20 leak. |
| Steering alpha (Language models and internal steering) | The coefficient controlling that addition | Sets an implementation-specific intervention strength | Alpha 0.5 is not 50 percent emotion |
| Steering vector (Language models and internal steering) | A direction added to selected activations | Defines which internal change is applied | Its identity and placement must match the certificate |
| Subject (Research question) | Here, a system considered as a possible bearer of experience. | The project tests operational organisation and states the assumptions of a subjecthood argument. | The name subject-core does not itself answer the question. |
| Surrogate (Measurement) | Constructed comparison data that preserve selected properties and disrupt others. | What was preserved or disrupted determines the alternative explanation tested. | Shift a recorded series to disrupt a declared cross-channel relationship. |
| Synergy (ISC and the subjecthood research) | A joint effect beyond adding the separate effects | Checks dependence of one source’s effect on another | The detailed guide works through 19 - 12 - 13 + 10 |
| Taint (Staying alive and recovering) | A retained record that an assumption or invariant broke | Keeps later green reports from concealing earlier damage | A repaired symptom does not automatically clear the historical condition |
| Target (Using a browser or screen) | The intended control or object | Binds the chosen action to a referent | Two similar buttons need a way to distinguish them |
| Telemetry (Runtime) | Measurements emitted by the system about its state or operation. | A named measurement needs a reader, units and known limitations. | A coherence score can be internal telemetry without measuring conversation quality. |
| Tensor (Computation) | An array of numbers, with a declared shape and interpretation. | Its size alone does not tell you what information it carries. | A hidden-state vector is a one-dimensional tensor. |
| Theory (Research) | An organised account that connects assumptions, mechanisms and predictions. | An implementation inspired by a theory must still be tested on its own terms. | Workspace theory motivates a coordination mechanism; it does not validate every implementation. |
| Throughput (Memory, speed and resource limits) | Work completed per unit time | Describes sustained capacity | High tokens per second can coexist with a long startup delay |
| Time to first audio (Conversation and voice) | Elapsed time until audible output starts | Describes how immediate speech feels | Fast text generation can still be followed by slow speech synthesis |
| Timeout (Staying alive and recovering) | A declared waiting limit | Bounds a stalled operation | An overly short limit can discard correct slow work |
| Token (Conversation and voice) | A unit processed or generated by the language model | Affects context capacity and decoding cost | A word can use more than one token |
| Tolerance (Engineering calculations and drawings) | The declared allowed deviation | Makes acceptance testable | It must be defined rather than inferred from a neat drawing |
| Trace (Logs, measurements and evidence trails) | Related events along one operation | Connects stages and delays | A request identifier prevents mixing two users’ turns |
| Training set (Learning and kept procedures) | Examples used to fit a method | Shapes what the method learns | It cannot also supply the whole independent transfer claim |
| Trust boundary (Mathematical claim limits) | The part of a system whose assumptions must remain valid for a security or verification claim. | Changing an admission rule needs protection of the property it was meant to preserve. | A protected checker may reject a proposed guard that would admit every action. |
| Trust boundary (Privacy and trust boundaries) | A point where data or authority changes scope | Defines which checks are required | A local browser page can still be untrusted input |
| Turn (Conversation and voice) | One request and its associated processing | Binds work to the person who asked | An interrupted turn must not publish into the next turn |
| Universal computation (Mathematical claim limits) | An ideal program system able to represent every partial computable function given sufficient resources. | Representability does not supply the correct program or a practical runtime budget. | A universal language can express a search that the finite machine cannot finish. |
| Unmeasured (Tests, controls and claims) | The required observation is absent | Prevents missing evidence from becoming a pass | A probe that cannot run contributes no positive score |
| Valence (Affect and operating condition) | An internal positive/negative coordinate | Can influence retrieval, attention or reports | A reported rating must follow an intervention to count as grounded |
| Validation (Learning and kept procedures) | Checks used to decide whether a change is acceptable | Can reject a harmful update | A passing local check may still have a narrow scope |
| Vector, x or h (Reading mathematical symbols) | A list of numbers used to represent a state | Different coordinates can describe different properties | A two-value example could keep condition and urgency; real model coordinates need their own interpretation |
| Velocity (Visual control and moving worlds) | Change in position over time | Helps predict where an object will be | One frame cannot establish a measured velocity |
| Verifier (Reasoning and checking answers) | A mechanism checking a stated property of a result | Can prevent unsupported answers from being accepted | Its own false acceptances and false rejections also need measurement |
| Weights (Language models and internal steering) | Stored model parameters | Usually define the trained model | Changing an activation is not automatically retraining the weights |
| Whitening (Measurement) | Transform baseline coordinates so declared variation is put on a reference scale. | It helps compare changes but cannot remove shared-history contamination. | A scalar two baseline standard deviations above its mean becomes 2. |
| Wired memory (Memory, speed and resource limits) | Memory the operating system cannot page out in the ordinary way | Reduces room available for other work | Unified CPU/GPU allocation is not described by Python heap alone |
| Working memory (Memory and recall) | Information currently kept available | Shapes the next computation | A retrieved note can be available for this turn without being permanent context |
| Workspace (Attention and competing concerns) | The selected information and competition state | Makes information available to other components | A selected item can influence several readers |
| z score (Reading mathematical symbols) | Distance from a reference mean in units of its standard deviation | Describes above-usual or below-usual state under a defined reference | Value 12, mean 10 and spread 2 give z = +1 |
Definitions are rendered from the reviewed catalogue. Project labels such as qualia, will and felt state do not acquire scientific validity from their names. Read the formula, its consumers, controls and evidence scope together.