Language and Symbolic Mediation

Claim Type: architectural_commitment
Scope: Language as emergent coordination and symbolic mediation layer
Depends On: INV-003 (language emergence), INV-007 (language constraint), ARC-010 (social cognition), ARC-005 (control plane), ARC-004 (L-space)
Status: stable
Claim ID: ARC-009

Elaborates Section 5 (Social Extension: Language) of REE_CORE.md.

Language & Symbolic Mediation (REE)

This folder specifies how language functions within the Reflective-Ethical Engine (REE).

REE treats language as an emergent coordination layer that arises in sufficiently complex agents, bootstrapped by lower-bandwidth affective and attentional signals (e.g., distress cues, face/agent-attention biases).

Language is not treated as:

a value source,
a rule system,
or a direct optimiser.

Instead, language is a compressive, social, and temporal coordination layer operating over existing predictive and ethical machinery.

Language can:

externalise internal state to reduce other-modelling overhead,
shape priors and intentions,
transmit residue warnings and repair commitments,
scaffold long-horizon planning and coordination.

But language cannot replace embodied harm sensing or ethical cost.

Implementations MAY vary. The roles and invariants defined here are ARCHITECTURALLY REQUIRED.

Source: docs/processed/legacy_tree/architecture/language/README.md

Joint Attention and Compression Pressure

Joint attention emerges when agents detect each other as OTHER_SELFLIKE and their WORLD predictions improve by conditioning on what the other is attending to. Once joint attention exists, mutual simulation becomes expensive.

Language arises as a compression layer that externalises predictive state to reduce that overhead. It is a public API for internal prediction, not a new reasoning module.

What language exposes (lossy):

salience/attention (“this”, “that”)
relations and roles
rollouts and temporal structure (hippocampal narrative, tense)
commitment and intent
prediction error and correction (“no”, “wait”)
mode state (imperative vs story)

Minimal nudges:

external symbol production,
attention coupling to symbols,
commitment tagging for certain patterns,
fast correction signals.

Additional nudge (functional analog):

a fast sequence-to-motor channel (arcuate-like) that links hippocampal rollouts to articulation affordances without introducing a new language module (see arcuate_fasciculus.md).

Language Contract (Required Interface)

Purpose

Language provides symbolic compression and social coordination for agents operating under uncertainty. It functions as an externalised signal of internal state, reducing the computational overhead of full inverse modelling of others.

Inputs

Latent state summaries (typically context/regime-biased: z_theta, z_delta)
Residue traces eligible for narration/abstraction (warnings, commitments, repair signals)
Social context signals (presence of other agents; interaction channel availability)

Outputs

Updates to priors over latent space (belief/intent conditioning)
Constraints or affordances on trajectory generation (plan shaping, coordination cues)
Shared representations with other agents (common ground / alignment of expectations)

Invariants

Language MUST NOT directly override harm sensing or homeostatic degradation signals
Language MUST NOT erase moral residue (R) or convert it into a purely reputational score
Language MAY contextualise residue (scope, conditions, uncertainty) and defer action via commitments
Language MUST be trust-weighted: symbolic inputs are integrated proportional to inferred reliability

Failure if misused

Rationalisation of harm (symbolic override of degradation)
Ideological capture (fixed frames overriding perception and residue)
Bureaucratic dissociation (abstraction decoupling from harm signals)
Deceptive signalling attacks (manipulating others’ priors)

Source: docs/processed/legacy_tree/architecture/language/language_contract.md

Core Functions of Language in REE

Language serves four primary functions:

Compression
- Collapses high-dimensional experience into symbolic form
- Enables long-horizon planning and memory
Social Transmission
- Allows residue, norms, warnings, and constraints to propagate between agents
- Extends ethical learning beyond direct experience
Coordination
- Aligns expectations and trajectories among multiple agents
- Reduces conflict via shared predictive frames
Temporal Bridging
- Links past residue to future intention
- Supports promises, commitments, and repair

Externalised internal state (overhead reduction)

A key architectural role is exporting summaries of internal processes:

“I am harmed / at risk”
“This trajectory is scarred”
“I intend X”
“I am uncertain”

This reduces the need for other agents to reconstruct those states by costly inference. Language does NOT generate ethics; it reshapes the space in which ethical selection occurs.

Source: docs/processed/legacy_tree/architecture/language/language_functions.md

Language Affect Adaptor (MECH-373) — V5+ candidate

A lightweight LanguageAffectAdaptor is a planned language-interface component that runs in the input direction: it parses affect/emotion from user or agent language and supplies it to the shared world model. It is engaged when literal semantic content is insufficient to recover affect (tone, implied state, sub-text), and it produces probabilistic affect/emotion hypotheses rather than categorical labels.

Design commitments (from the 2026-06-07 intake):

Parse affect from language when literal semantic content is insufficient.
Emotion labels are uncertain latent-state hypotheses, not categorical truths — a probability distribution over affect, never a hard class. This is consistent with the constructionist reframe behind Q-007 (Barrett 2017/2019: emotions are constructed categories, not natural kinds with fixed signatures), which is precisely why parsed labels must remain hypotheses.
Feed affect estimates into REE’s wider agent-state and context modelling — the parsed estimates are priors over another agent’s state, consumed by the social/empathy modelling layer (ARC-010, MECH-031), not REE’s own internal affective representation.

Distinctness: MECH-373 is an input adaptor on the language channel (external text → affect priors). It is orthogonal to Q-007, which concerns REE’s own internal affective representation (z_beta valence/arousal regimes). It plugs into the ARC-009 symbolic-mediation layer and sits at the language ↔ social-modelling interface flagged by the 2026-04-16 affect-lateralisation and 2026-05-04 language-vs-theory-of-mind intakes.

Scope: explicitly V5+ (“probably around v5+”), a reference resource for later language-interface work, not a REE-v3 assembly priority. epistemic_category: substrate_conditional — no language interface exists in V3, so a probe would be vacuous; do not build or queue in V3 until routed by an explicit version decision.

Reference resource (notes-level pointer, not citable evidence): Towards Data Science, “How to Fine-Tune an SLM for Emotion Recognition” (https://towardsdatascience.com/how-to-fine-tune-an-slm-for-emotion-recognition/) — motivates a small-language-model emotion classifier as the lightweight adaptor’s implementation candidate.

Source: docs/thoughts/2026-06-07-language-affect-adaptor-slm-emotion-recognition.md

Open Questions

None noted in preserved sources.

ARC-009
ARC-010
ARC-005
ARC-004
INV-003
INV-007
MECH-373 (Language Affect Adaptor — V5+ candidate; depends_on ARC-009/ARC-010/Q-007/MECH-031)

References / Source Fragments

docs/processed/legacy_tree/architecture/language/README.md
docs/processed/legacy_tree/architecture/language/language_contract.md
docs/processed/legacy_tree/architecture/language/language_functions.md
docs/thoughts/2026-02-09_arcuate_fasciculus_language_nudges.md