ADUS v1.4 — Glossary#

Companion reference to ADUS Framework – Technical Report (v1.4). Entries are alphabetical. Section §-references point to the technical report.

Entries marked [gap] are terms the report uses without defining; a proposed definition is given, but it is a proposal, not a restatement. These are collected again in the final section.

If the notation is new to you: start with one ordinary example. Learning to drive begins as a step-by-step Zone-3 skill, becomes a partly effortful Zone-2 skill, and may become a fluent Zone-1 skill. Do_A describes that growing automaticity; Aw asks whether the driver can accurately notice what the automatic process is doing; Cn asks whether it can be stopped or redirected. If the lesson is forgotten, written in a notebook, reinforced by an instructor, or deliberately practised by the driver, those are respectively C-N, C-X, C-T, and C-S consolidation channels. The entries below make those distinctions precise.

Reading order#

For a first pass, the framework decomposes into five layers:

1. Nodes — A, D, U, S. What kinds of working part exist.
2. State parameters — Do_A, Aw, Cn, PS. What state a component is in.
3. Zones — Zone-1/2/3. Position on the load axis.
4. Dynamics — plasticity, consolidation channels, matrix edges. How components move and interact.
5. Boundary conditions — V, coverage, reachable set. What the environment and the architecture must supply.

A#

A — see Abilities.

A-qualitative change (§3.1) — a change to A that alters which U and S are reachable at all. The criterion for Level 2. Paradigm case: a new sensory or motor modality.

A-quantitative change (§3.1) — a change to A that alters the ceiling, rate, or capacity within an unchanged reachable set. Width scaling, parameter growth, and context extension are A-quantitative by default.

A-realisation (§3.1) — the emergence under training of a capacity that the architecture already afforded. Not new A. Distinguishes "the model developed X during training" from "the model acquired the ability to develop X."

Abilities (A) (§3) — substrate-bounded core capacities, biological or architectural. The node that sets ceilings on what U and S can be acquired. Sub-domains: cognitive, perceptual, motor, sensory, consolidative. In artificial systems: context length, attention capacity, parameter count, inductive biases, presence of a consolidation route.

Ablation — removing or zeroing a model component to test whether behaviour depends on it. Used in the Aw assay (§8.1) as an implementation-level intervention measuring a functional quantity.

Activation patching — replacing activations at a chosen site with those from a different input, to test causal contribution. Same role as ablation in §8.1.

Architecturally afforded (§3.1) — the framework's stance on A: architecture bounds the reachable set; training determines which of the afforded set is realised. Contrasts with "architecturally given," under which A would be fully fixed at design time.

Automaticity — see Do_A.

Awareness — see Aw.

Aw( ) (§3.4) — awareness parameter, or "re-conscious-isation gain," 0–1 per node. 1 = the actor can voluntarily return a Zone-1 process to conscious working memory and report its content veridically. 0 = opaque under explicit probing. Veridicality, not fluency, is the criterion.

Aw(U) — awareness of the understanding store specifically. Central to the framework's LLM diagnosis: frontier models are hypothesised to have Aw(U) ≈ 0 for provenance, so metacognition (the U-row) runs off a store the system cannot veridically inspect.

C#

C-N / C-X / C-T / C-S (§3.3) — the consolidation channels, in ascending order of substrate involvement:

The ladder is a substrate ordering with an initiation qualifier on the substrate rung, not two independent axes.

Ceiling edge (§5) — the A→U and A→S cells of the interaction matrix. Where A constrains what U and S can be acquired. The reachability criterion (§3.1) is stated in terms of these edges.

Cn( ) (§3.4) — control parameter, or "veto/redirect gain," 0–1 per node. 1 = complete on-line suppression or re-routing of an automatism; 0 = no detectable override. Assay: stop-signal with automatism pre-loaded, Cn = 1 − p_fail × (SSRT / median_go_RT). Deliberately pure suppression: amplification is modelled separately as D-driven potentiation.

Confabulation (§3.4) — the Aw = 0 failure mode with high verbal output. Confident, articulate self-report with no predictive relationship to the underlying process. The reason fluency cannot serve as an awareness measure.

Consolidation (§3.2, §3.3) — the process by which Zone-2/3 content becomes Zone-1 content; equivalently, movement along the load axis that persists beyond the episode.

Consolidation route (§3.2) — an available path for consolidation. Its existence, not merely its speed, determines reachability, which is why consolidation capacity is classified as A.

Consolidative (A sub-domain) (§3.2) — the ability sub-domain covering consolidation capacity. New in v1.4. Its absence in frontier LLMs is the framework's central AI diagnosis.

Constrained reasoning budget (§6.2) — proposed load assay for artificial systems. Vary permitted reasoning-token allocation; measure degradation. Zone-1 content survives severe constraint, Zone-3 procedures collapse. Degradation slope is the analogue of dual-task cost.

Coverage (§8.3) — environmental parameter: breadth and sequencing of available experience. Distinct from V. Curriculum is a coverage intervention, not a verification one.

D#

D — see Disposition.

D-driven potentiation (§5.1) — amplification of an automatism modelled as a positive-gain D→ matrix edge rather than as negative control. Preserves the dissociation between "can suppress when engaged" (high Cn) and "strong drive when not engaged" (strong potentiation).

Disposition (D) (§3) — affective, motivational, and habit circuits. Sub-domains: affect-valence, drive-intensity, habit-context. Assayed as three loops (see Loop A/B/C).

Do_A( ) (§3) — automaticity parameter, 0–1 per node. Higher = more automatic, lower dual-task cost, less stepwise execution. Determines zone placement but not node identity (§3, node-identity principle). Notation normalised in v1.4; v1.3 used DoCA / DoDA / DoPA / DoSA.

Dual-task cost (§3.3) — performance decrement when a component is exercised under a competing demand. The primary operationalisation of the load axis in humans.

E#

EEfRT (§6) — Effort Expenditure for Rewards Task. Effort-discounting assay used for Loop B of D.

Elastic / plastic — article vocabulary, not framework vocabulary. Elastic adaptation reverts when the episode ends (C-N); plastic adaptation persists (C-T, C-S). Note the collision with Elastic Weight Consolidation, where "elastic" denotes a restoring force that aids retention — the opposite valence.

Episodic memory (as an A candidate) (§3.1) — listed as a candidate A-qualitative capacity. Note that in LLMs the context window arguably supplies the store functionally, and what is missing is the encoding step — making this partly a consolidation question rather than an ability question.

Executive function (as an A candidate) (§3.1) — listed as a candidate A-qualitative capacity. The framework's negative verdict for current LLMs runs through Aw rather than A: monitoring that is not veridical is not executive control.

F#

Functional-primary (§2) — the level-of-analysis commitment in v1.4. Node identity, zone placement, and all §9 claims are stated functionally; implementation evidence is admissible as evidence and mechanism, not as definition.

I#

Implementation evidence (§2) — architectural facts, ablation, patching, mechanistic interpretability. Admissible under the priority rule.

Insight strength (IS) (§4) — [gap] named as a quantity carried on the S→U edge, alongside PS on U→S, but not defined. Proposed: the strength with which procedural competence yields explicit conceptual content — the reverse of proceduralisation, i.e. articulating what one has learned to do.

Interaction matrix (§5) — the 4×4 directed source→target influence structure over nodes. Interventions are held to target edges rather than nodes (§7).

Internal simulation (§8.2) — roll-forward of represented state that is not routed through the output channel. An information-flow criterion, chosen so that internality remains functional rather than implementational. Assay: constrain reasoning tokens and test whether simulation-dependent performance survives.

L#

Level 1 (§3.1) — acquisition of new U and S within a fixed reachable set.

Level 2 (§3.1) — A-qualitative change: expansion of the reachable set.

Load axis (§3.3) — the dimension along which zones are defined in v1.4: sensitivity to competing demand and dependence on explicit stepwise execution. Separated in v1.4 from plasticity, which was previously bundled into the zone definition.

Loop A / Loop B / Loop C (§6) — [gap] the three D feedback loops, identified by assay (affective-priming RT; EEfRT effort-discounting; slip-of-action habit test) rather than by definition. Proposed mapping: Loop A ≈ affect-valence, Loop B ≈ drive-intensity, Loop C ≈ habit-context, matching the sub-domain list in §3.

M#

Metacognition (§5) — defined structurally as the U-row of the interaction matrix: U→A, U→D, U→U, U→S. Knowledge acting on the other nodes. Its reliability is bounded by Aw(U), or substituted for externally by V.

Mismatch diagnosis (§7) — a diagnostic pattern across nodes rather than within one: high A + low D → under-achievement; high U + low S → knowing-without-doing.

N#

Node (§3) — one of the four components: A, D, U, S.

Node-identity principle (§3) — a node's identity is set by functional role (capacity ceiling, motivational circuit, content store, procedure), never by automaticity or trainability. Frozen, non-updatable knowledge is Zone-1 U, not A. Exists to prevent A↔U and A↔S conflation.

P#

Plasticity (§3.3) — the transition rate between zones. In v1.4 it is not part of the zone definition and it is channel-relative: a property of (component × consolidation channel). Decomposed as rate = f(A_consolidation, channel availability, Aw).

Predictive validity (§8.1) — the criterion in the Aw assay: does the introspective report predict behaviour under intervention. Replaces confidence and coherence as the measure.

Priority rule (§2) — where behavioural and implementation evidence conflict, the behavioural finding fixes the attribution and the implementation finding becomes the thing requiring explanation.

Process-dissociation hit-rate (§3.4) — human Aw assay. Probe-cued report of the last automatic stimulus or rule; hit-rate = Aw.

Proceduralisation strength (PS) (§3, §4, §6) — the quantity on the U→S edge measuring how far declarative content has become procedure. Assayed by retrieval latency plus dual-task cost during recall. Distinct from Do_A(U), though v1.3 presented them in the same cell.

R#

Reachable set (§3.1) — the set of U and S that a given architecture permits to be acquired at all. The central construct behind Level 1/Level 2: Level 1 moves within it, Level 2 expands it.

Reachability criterion (§3.1) — the test for A-qualitative change: does the change alter which U and S are reachable, rather than the ceiling, rate, or capacity within an unchanged set.

Re-conscious-isation (§3.4) — the process Aw measures: returning an automatised process to conscious working memory such that its content can be veridically reported.

Retest gain, channel-indexed (§6) — retest gain is now reported together with the consolidation channel under which it was measured, since under C-N every rate is zero by construction.

S#

Skills (S) (§3) — learned procedures, motor or cognitive. Sub-domains: motor, cognitive, perceptual-motor, social.

Slip-of-action test (§6) — habit assay used for Loop C of D: does a previously reinforced response persist after the contingency is devalued.

SSRT (§3.4, §6.1) — stop-signal reaction time. Component of the Cn assay.

T#

Tick (§4) — [gap] used in the Cn formulation ("for the current tick"), implying a discrete-time dynamical model that is not specified. Proposed: one update step of the interaction matrix; needs pinning down before the matrix can be simulated.

U#

Understanding (U) (§3) — the declarative and episodic content store. Sub-domains: factual, conceptual, experiential, contextual. Note that zone placement for a store is about retrieval automaticity rather than execution load; see the gaps section.

U-row — see Metacognition.

V#

V (verification strength) (§8.3) — environmental parameter: the quality of the feedback signal available in the task environment. Ordinal: formal verifier > execution feedback > learned judge > intrinsic signal. Not an ADUS node; an adjacent condition.

V × Aw substitution (§8.3, claim 7) — the claim that V and Aw are substitutes on the U-row edges: metacognitive correction can be driven externally by verification or internally by veridical self-inspection. Predicts reliable improvement at high V even where Aw ≈ 0, and stalling at low V.

Veridicality (§3.4) — the correspondence between a self-report and the process it reports on. The criterion for Aw, chosen specifically to exclude fluent confabulation.

Z#

Zone-1 (core / trait) (§3.3) — dual-task cost < 10%; no explicit stepwise execution; resistant to single-session intervention. Provisional Do_A ≥ 0.9.

Zone-2 (trainable) (§3.3) — measurable dual-task cost; partial stepwise dependence. Provisional Do_A 0.3–0.9.

Zone-3 (conscious / state) (§3.3) — performance collapses under load; requires explicit step-by-step execution. Provisional Do_A < 0.3.

Zone placement (§3) — the current state of a component within a node. Distinct from node identity, which is fixed by functional role.

Gaps: terms used but not defined in v1.4#

Collected for a v1.5 pass. Proposed definitions above are proposals only.

1. Insight strength (IS) — named in §4, never defined.
2. Loop A / B / C — identified by assay, not by construct. The mapping to D's sub-domains is inferable but unstated.
3. Tick — implies a discrete-time dynamical model that the report does not specify. Blocking for any simulation of the matrix.
4. Zone semantics for a content store. The load axis is defined by dual-task cost and stepwise execution, which are properties of processes. U is a store. In practice §6 measures U through retrieval (latency + dual-task cost during recall), so U's zone is retrieval automaticity — but this is never said, and as written the zone criteria do not straightforwardly apply to U at all. Worth an explicit sentence in §3.3.
5. PS vs Do_A(U) — v1.3 listed them in the same table cell; v1.4 asserts they are distinct without stating the relationship. Are they independent, or is PS a directional component of Do_A(U)?
6. "Developmental windows" (§4) — invoked as a mechanism locking A into Zone-1; not defined, and its artificial-system analogue is unstated.
7. Sub-domain boundaries within U — "experiential" and "contextual" are not distinguished from each other or from "factual."
8. Channel availability — appears in the plasticity decomposition (§3.2) as a term in f( ), but is not itself given a scale. Binary, or graded by bandwidth and latency?