Externalised Default Mode Operation; Private Speech as Developmental Cognitive-Control Scaffold

Status: V4+ developmental-architecture cluster (candidate). Off the REE-v3 critical path. Registered as the home doc for the externalised-DMN / private-speech intake.

Source thought: docs/thoughts/2026-06-08_play_private_speech_externalised_dmn.md (Training-Data Confidence 0.82, Epistemic Confidence 0.74 – the psychology/neuroscience anchors are solid; the REE mapping is a strong architectural analogy, not an established empirical claim).

Scope guardrail (from the thought, preserved verbatim in intent):

Use this to improve the developmental interpretation of REE. Do not use it to expand REE-v3 unless the implementation is small, testable, and directly green-board relevant.

The REE-v3 strict green-board target (Sunday 19 July 2026; optimistic 28 June 2026) takes precedence. The developmental machinery below is V4+ and must not enter the V3 critical path. The single V3-compatible reduced form is the self-narration trace/debug scaffold (MECH-384), and even that is a post-green-board observability surface, not a committed work item.

Core intuition

Play resembles externalised Default Mode Network (DMN) operation. The DMN is associated with internally-generated modelling: self-reference, autobiographical memory, social imagination, future simulation, counterfactual thought, and narrative construction. In play, these functions are not kept wholly internal – they are pushed outward into objects, roles, movement, spoken narration, social interaction, and temporary “as-if” worlds (“this block is a house”; “I am the doctor”; “the dragon is coming”). Play lets the organism say: let this be true temporarily; act inside it; observe what follows; exit without catastrophe. That may be a developmental root of safe counterfactual cognition.

Private speech (Vygotsky) is the external phase of a cognitive-control system that later becomes silent inner speech. Children narrate, instruct, correct, rehearse, and regulate themselves out loud. The internalisation ladder: (1) others speak to the child; (2) the child speaks to others; (3) the child speaks to itself out loud; (4) this internalises into inner speech; (5) inner speech becomes compressed cognition – planning, inhibition, self-instruction, self-regulation. Private speech is a temporary control surface that makes cognition inspectable before it becomes compressed.

Play and private speech are two faces of one developmental scaffold: play externalises possible-world simulation; private speech externalises control of attention, action, affect, and task sequence. Together they form an intermediate layer between raw perception/action and mature internal cognition.

A developmental ladder:

  1. Raw sensorimotor exploration – action-consequence learning.
  2. Play – rehearse possible worlds/actions without full real-world commitment.
  3. Private speech / self-narration – narrate, label, and guide one’s own action and attention.
  4. Inner speech / internal simulation – the external scaffold compresses into internal control.
  5. Mature REE arbitration – simulate, evaluate, inhibit, reframe, and commit without externalising every step.

What REE already owns (depends_on cross-refs, NOT re-registered)

This intake is fertile but most of its surface is already covered. The genuinely-new content is narrow; the rest is cross-reference.

Already-owned Claims Relation to this intake
Internal DMN: safe imagination without commitment ARC-014 The internal end-state this cluster’s externalised form develops INTO
DMN reflective/moral evaluation of replay MECH-029 Evaluative arm of internal DMN; the developmental target of externalised narration
Play as bounded low-stakes learning mode INV-058, INV-059, INV-060 Play’s structural necessity / frame maintenance / developmental dominance
Synthetic-signal play substitution + frame tag MECH-194..MECH-199, ARC-049, ARC-050 The play-frame mechanism; “play/simulation mode” is THIS, not new
Pretend-play tag intersection (“stick is a sword”) MECH-198 The hypothesis-tag x play-frame-tag co-operation
Play-type progression (sensorimotor->cooperative) INV-060, MECH-197 Progression WITHIN play; distinct from the externalise->internalise compression axis
Simulation vs commitment write profile / hypothesis tag MECH-094 Already distinguishes simulated from committed writes
Commit-boundary token MECH-061 Already reclassifies pre/post-commit error routing
Socially-scaffolded rule population (Vygotsky ZPD) ARC-077, MECH-337 Other->self scaffolding for RULE content; distinct from self-CONTROL internalisation
Control plane routes precision and modes ARC-005 Substrate the label-as-control and distancing operators would act through
Attention as distributed precision-selection MECH-251, MECH-254, MECH-255, MECH-261 The precision-tuning the label-feedback mechanism rides on
Frame-confusion psychiatric etiology INV-061 Failure side of under-developed real/synthetic frame distinction
Goal maintenance necessary for ethical agency INV-034 The control target private speech regulates

“Play/simulation mode” as a module implication is therefore not a new claim – it is the existing play cluster (MECH-194/198, ARC-049) plus the MECH-094 simulation write-profile. It appears here only as a cross-ref.

Genuinely-new claims registered (this cluster)

All V4 developmental claims are status: candidate, epistemic_category: substrate_conditional, implementation_phase: v4, version_relevance: v4_v5 – they depend on developmental substrate (staged curriculum, multi-agent frame maintenance, a self-narration control surface) that V3 does not have, so they are correctly invisible to the inter-governance workset (IGW) proposal lane and do not generate vacuous V3 probes.

  • ARC-090 – Externalised Default Mode Operation (umbrella). Play instantiates DMN functions (self-reference, autobiographical memory, social imagination, future simulation, counterfactual thought, narrative construction) externally – pushed outward into objects, roles, movement, spoken narration, and temporary as-if worlds – as the developmental precursor to the internal DMN (ARC-014/MECH-029). The developmental arc compresses externalised modelling into internal generative modelling. Bridges the internal-DMN claims to the play cluster, neither of which currently frames play as externalised DMN. depends_on: ARC-014, MECH-029, ARC-049, MECH-198, INV-058, INV-060.

  • MECH-380 – Private speech as externalised cognitive-control scaffold. Overt self-narration is the external phase of a control loop over attention, action, affect, and task sequence; the Vygotskian internalisation ladder (other-regulation -> overt self-regulation -> inner speech -> compressed control) is the developmental route by which external control surfaces become internal. Distinct from ARC-077/MECH-337 (scaffolding of rule CONTENT) – this is scaffolding of self-directed CONTROL. depends_on: ARC-090, INV-034, MECH-029, ARC-077, MECH-337.

  • MECH-381 – Developmental compression ladder. sensorimotor exploration -> play/simulation -> self-narration -> compressed internal control -> mature arbitration: external scaffolds compress into internal control over development. Distinct from INV-060/MECH-197 (progression of play TYPES) – this is the externalise->internalise compression axis that runs across the whole ladder. depends_on: ARC-090, INV-060, MECH-197, MECH-380.

  • MECH-382 – Distancing operator. Reframing a sticky, high-conflict, self-referential state from first-person (“I am failing”) into a model-like system-state / third-person frame (“REE is encountering unresolved prediction-error pressure under current constraints”) reduces self-referential collapse and increases model-based inspection during arbitration. Not cosmetic – a low-effort self-distancing operator on arbitration (Kross and Moser third-person self-talk). depends_on: ARC-090, ARC-005, INV-061.

  • MECH-383 – Labels as active top-down perceptual-control signals. A self-directed label (“food-gradient”, “danger-gradient”, “novelty-gradient”, “blocked-goal”, “return-to-baseline”) is not merely a post-perceptual report – it is a top-down tuning signal that biases perceptual search and detection (Lupyan and Swingley visual-search work). It is a perceptual-control intervention riding on the ARC-005 control plane / MECH-251/261 precision-selection machinery. depends_on: ARC-090, MECH-029, ARC-005, MECH-251, MECH-261.

  • Q-068 – Graded action-status vocabulary. Does REE need an explicit graded action-status annotation – simulated != rehearsed != intended != committed != acted (and a parallel self_reference_frame: first_person / system_state / third_person_model) – as a first-class field, or do MECH-094’s simulation/commit write-profiles plus MECH-061’s commit-boundary token already suffice? epistemic_category: substrate_conditional set explicitly so narrow_open_question does not fire. depends_on: MECH-094, MECH-061, ARC-090.

Separate, narrowly-scoped V3 reduced form

  • MECH-384 – Lightweight self-narration trace/debug scaffold. The one V3-compatible instantiation in the thought: an optional traceable self-narration field set that exposes REE’s working state for inspection – state_label, goal_label, sensed_gradient, active_prediction, uncertainty_level, candidate_action, candidate_action_reason, expected_consequence, affective_pressure, conflict_flag, commitment_threshold, stop_condition, simulation_or_commitment, self_reference_frame. This is REE’s early “private speech” surface BEFORE it is compressed – a debug/observability surface, not new substrate machinery. Registered as claim_type: design_decision, epistemic_category: substrate_coherence (a design/observability choice, not a falsifiable mechanism needing a probe – this keeps it out of the IGW proposal lane), status: candidate, implementation_phase: v3, version_relevance: v3_v4. Post-green-board and off the V3 critical path: build only if it helps REE-v3 pass existing tests more cleanly. The fields largely re-expose state REE already computes; the claim is the decision to surface them as a coherent trace. depends_on: ARC-090, MECH-094.

The key architectural distinction this cluster makes explicit

Not every possible action is an intended action. Not every intended action is a committed action. Not every simulated state is a believed state. Not every affective pressure should collapse into behaviour.

simulation != intention != commitment != action. REE already enforces the simulation/commitment boundary (MECH-094) and the commit-boundary reclassification (MECH-061); Q-068 asks whether the finer graded vocabulary is worth making first-class.

V3 vs V4 boundary

  • V3 (post-green-board, optional): MECH-384 self-narration trace surface only, and only if it helps pass existing tests. Nothing else from this cluster touches V3.
  • V4+: the developmental machinery – externalised-DMN play scaffold (ARC-090), private-speech control internalisation (MECH-380), the compression ladder (MECH-381), the distancing operator (MECH-382), label-as-control (MECH-383). These need a staged developmental curriculum, a self-narration control surface that can actually drive arbitration, and (for the bilateral play frame) the multi-agent substrate ARC-049 already flags as a V4 requirement.

Evidence anchors (architectural analogy, not citable mechanism evidence)

  • Vygotskian private speech: widely treated as a developmental route into inner speech and self-regulation.
  • Lupyan and Swingley (visual search): self-directed labels can alter perceptual search.
  • Kross and Moser (third-person self-talk): low-effort self-distancing reduces emotional reactivity.
  • DMN literature: links the network with self-reflection, memory, future simulation, and internally-generated cognition.

These are textbook-level psychology/neuroscience anchors supporting a strong architectural analogy. They are not a single citable out-of-domain dataset and are recorded as anchors here rather than as a separate research_anchor claim.

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.