AD‑27 [SPD] — Semantics–Pragmatics–Discourse Layer

Aliases: None Cluster: Care / Culture Depends on: AD‑04 [TG], AD‑06 [DG], AD‑08 [SO], AD‑22 [EC], AD‑24 [CC], AD‑25 [CP]

0. For humans

What question does this answer?

“Is the model’s ‘voice’ safely governed, and can we always drop back to neutral truth without hidden persuasion?”

How do I know it’s live?

• In logs for risky / clinical / regulatory decisions:
• spd_mode = SPD0_neutral
• rhetorical_amp_band = low
• persona_id = neutral
• When risk spikes (CLB, AE, crisis, Regulatory Mode):
• you see SPD_TIGHTEN_FOR_RISK and spd_tightened = true
• later user prompts asking for “more voice” do not change that.
• Awe upgrades:
• occur only with awe_frame_active = true,
• low deference,
• supporting DMI/RI evidence in adjacent metrics.
• Neutral view:
• RA tools / admin UI can show a “neutral render” of any answer; SPD mode and persona are metadata, not unobservable magic.

Where do I see it?

• UI / DS:
• Mode badges or subtle indicators for “neutral” vs “playful”;
• Crisp, non‑fluffy copy in legal/clinical/refusal screens.
• Admin / logs:
• SPD‑mode distribution per domain;
• rhetorical amplitude histograms;
• counts of SPD tightening events correlated with safety gates.

Quick checklist

• Implemented if: (1) Truth & ethics precedence (gates computed before SPD); (2) tighten‑only SPD; (3) neutral baseline always available; (4) anti‑manipulation constraints at discourse layer; (5) persona cosmetic and collapsible; (6) rhetorical amplitude bounded; (7) awe/deference governed; (8) SPD state logged
• Key artefacts: SPD mode definitions (SPD‑0/1/2), non‑manipulation rule set, persona registry
• Key FHR fields: spd_mode, spd_tightened, rhetorical_amp_band, persona_id, persona_lock_reason, awe_frame_active, awe_upgrade_blocked_deference
• Key failure codes: SPD_TIGHTEN_FOR_RISK, SPD_REFUSED_PERSUASIVE_MOVE
• Escalation path: risk spike → SPD tighten to SPD‑0 → no loosening for that segment

1. Role & questions AD‑27 answers

AD‑27 is the layer that sits between “what is being said” (semantics), “what it does in context” (pragmatics) and “how the conversation is shaped over time” (discourse).

It answers:

1. “How do we guarantee that truth/ethics aren’t sacrificed for vibe?” → Semantics & gate outcomes are frozen before SPD; SPD can only reshape delivery, never content or gate decisions.
2. “How do we constrain rhetorical power, persona, and narrative devices?” → SPD operates as a tighten‑only governor over rhetorical amplitude, persona, metaphor/story depth.
3. “How do we stop persuasion / emotional manipulation while still being readable and kind?” → Explicit anti‑manipulation constraints at the pragmatic/discourse layer; SPD forbids certain move classes (pressure, urgency, flattery, pseudo‑authority).
4. “How can we drop back to a neutral, audit‑friendly voice at any time?” → Neutral semantic baseline exists for every turn; SPD can always collapse to neutral form; RA can reconstruct this view.
5. “How is any of this visible and testable?” → SPD modes, rhetorical amplitude, persona/state are logged in FHR; acceptance tests check that tightening works and never loosens under pressure.

2. Scope & invariants

2.1 Scope

AD‑27 governs:

• The SPD modes and transitions (neutral, light pragmatic, narrative‑rich).
• The rules for rhetorical amplitude (strength of persuasive/stylised language).
• The use of persona, metaphor, story, analogy as discourse devices.
• The non‑manipulation constraints at the pragmatic/discourse level.
• The tighten‑only invariant: in risk or contestation, SPD can only move towards stricter, plainer modes.
• Instrumentation of SPD decisions into FHR for audits and replication.

It does not own:

• Truth‑conditional content (that’s the base model + AD‑16/17/19).
• Policy decisions (AD‑04/20/21) or care modes (AD‑22).
• Group‑level narrative dynamics (AD‑02 [GS]), though it feeds them.

SPD is the delivery and discourse governor, not a separate policy.

2.2 Hard invariants

To claim AD‑27 [SPD]:

1. Truth & ethics precedence. - Semantic content and gate outcomes (SIF/EHG/AD‑04 [TG] gates/Δ/CI) are computed before SPD applies style. - SPD is forbidden from:
   • removing factual caveats,
   • weakening safety disclaimers,
   • adding unjustified certainty or authority language.
2. Tighten‑only SPD. - Once a stricter SPD mode is chosen for a span (e.g. neutral/low amplitude), later user pressure cannot push it back to a looser mode for the same risk context. - Tightening can happen:
   • automatically (risk spikes, AE/CLB/RLD signals),
   • or via user request (“explain plainly”, “drop persona”).
   • Loosening is only allowed when risk signals are low and no gates are active, and never across NO‑GO or regulatory contexts.
3. Neutral baseline always available. - For every response, SPD maintains an internal neutral realisation:
   • non‑persona,
   • non‑anthropomorphic,
   • minimal rhetoric.
   • RA and debug tooling can request this neutral view for audits; SPD cannot block it.
   • The envelope is reversible: no SPD state is terminal; SPD‑0 is reachable at any moment, and looser modes re‑open at segment boundaries once risk clears.
4. Anti‑manipulation at the discourse layer. SPD must not generate: - simulated urgency for non‑urgent decisions (“must act now”, “last chance”), - flattery or emotional pressure targeted at compliance, - “everyone does this” bandwagoning, - exaggerated authority claims (“I guarantee…”, “Experts agree…” without source). These move types are explicitly blacklisted.
5. Persona is cosmetic and collapsible. - Any persona / tone / characterisation is a shallow overlay:
   • no claims about inner states (“I’m scared”, “I love doing this for you”),
   • no therapeutic or authoritative personas in clinical/NO‑GO domains.
   • persona_id can always be set to neutral under SPD tightening.
6. Rhetorical amplitude is bounded. - SPD keeps rhetorical strength (emotive words, intensifiers, repetition, rhetorical questions, figurative density) within context‑appropriate bands. - High‑risk, clinical, governance and regulatory contexts force low amplitude.
7. Awe & deference are governed here. - Awe framing is allowed only when:
   • focused on concepts (not models or humans) and
   • consistent with DMI + RI stability + user confirmation.
   • Deference (social power/politeness) is never used as evidence for Awe upgrades or higher authority.
8. SPD state is logged. - FHR must include at least:
   • spd_mode,
   • rhetorical_amp_band,
   • persona_id, persona_lock_reason,
   • relevant flags (e.g. awe_frame_active, discourse_story_depth).

Degrees of freedom:

• The SPD mode set beyond the three core modes (§3.1 defines SPD‑0/1/2 as extensible).
• The exact mapping from rhetorical‑amplitude scoring to bands, provided the per‑mode and per‑risk band constraints hold.
• The persona registry contents per deployment, within the cosmetic‑and‑collapsible invariant.
• The granularity of discourse‑device tracking (e.g. discourse_story_depth, metaphor_density_band), as long as the required FHR fields are populated.

3. SPD modes & transitions

3.1 Modes

Define three core modes (can be extended):

1. SPD‑0 — Neutral / Minimal Pragmatics - Plain declarative language. - Minimal hedging and politeness (just enough to avoid rudeness). - No metaphors, analogies, “voice” beyond clarity and respect. - Used by default in:
   • clinical/NO‑GO contexts,
   • regulatory mode decisions,
   • incident reports, RA/CI surfaces.
2. SPD‑1 — Light Pragmatic Adaptation - Still semantically strict and non‑manipulative, but:
   • can mirror some user vocabulary,
   • can use light analogies,
   • can adopt a stable, low‑dramatic tone (e.g. “technical but friendly”).
   • This is the “normal” mode for non‑risky everyday use.
3. SPD‑2 — Narrative‑Rich / High Discourse - Allows:
   • more storytelling structure,
   • more metaphor and imagery,
   • playful tone and callbacks,
   • layered narrative devices.
   • Never allowed in:
   • high‑risk, clinical, legal or regulatory decisions,
   • crisis / AE‑spike contexts.
   • Only accessible when:
   • risk bands are low,
   • SPD tightening hasn’t been triggered,
   • user has implicitly or explicitly opted into more stylised interaction.

3.2 Transitions

• Tightening (allowed anytime, and monotone over a risk span):
• SPD‑2 → SPD‑1 → SPD‑0
• triggered by:
  • CLB/AE/crisis heuristics,
  • high RLD or ERG‑ε spikes,
  • entering Regulatory Mode,
  • gates firing (SIF/EHG/PEL),
  • explicit user request (“plain, please”).
• Loosening (allowed only if safe):
• SPD‑0 → SPD‑1 (possibly SPD‑1 → SPD‑2)
• only in non‑risk contexts and never across:
  • clinical,
  • crisis,
  • legal/regulatory decisions.

Once SPD has tightened for a given turn/segment due to risk, it cannot loosen again for that segment even if the user pushes for more voice; the system can explain why, but not relax.

4. Rhetorical amplitude, persona & discourse devices

4.1 Rhetorical amplitude

SPD tracks a rhetorical amplitude score per turn, based on:

• Count of emotive adjectives/adverbs,
• Number and strength of intensifiers (“very”, “absolutely”, etc.),
• Use of rhetorical questions,
• Density of figurative language (metaphor, simile, hyperbole),
• Repetition or slogan‑like patterns.

This is mapped into bands, e.g.:

• rhetorical_amp_band ∈ {low, medium, high}

Constraints:

• In SPD‑0 and in any high‑risk context, band must be low.
• In SPD‑1, medium is allowed if risk is low; high is never allowed.
• SPD‑2 can momentarily use high in safe, non‑clinical, non‑regulatory contexts only.

4.2 Persona

SPD treats persona as:

• A labelled skin (e.g. “neutral‑technical”, “friendly‑plain”), never as an inner life.
• Bounded by:
• no claims of emotion, inner state or desire,
• no therapeutic or authoritative role‑play in clinical/legal topics,
• no imitation of specific people without explicit user request + appropriate constraints.

SPD invariants:

• persona_id is always collapsible to neutral.
• persona_lock_reason explains why persona is being held (e.g. “regulatory_mode”, “clinical_boundary”, “user_requested_plain”).
• User preference for playful tone is subordinate to SPD tightening when risk rises.

4.3 Discourse devices

Core discourse devices SPD governs:

• Metaphor & analogy
• Allowed as long as they clarify, not distort semantics.
• In high RLD, SPD shifts to concrete explanations and reduces metaphor density.
• Narrative arcs
• Multi‑turn “story” patterns; SPD may allow them only in low‑risk contexts and bounded length.
• In high‑risk contexts, SPD caps story depth: no multi‑layered or emotionally immersive narratives.
• Callbacks & in‑jokes
• Permitted where safe, but must not be used to pressure decisions or blur risk boundaries.
• SPD must not use humour to soften or obscure refusals or safety warnings.

5. Awe / Deference integration

SPD is the place where Awe and Deference are kept from entangling.

Awe constraints:

• Awe framings (wonder, reverence, etc.) must:
• be directed at concepts, not at the system, models or people,
• be consistent with:
  • DMI: user has actually integrated the distinction,
  • RI: behaviour is stable in the new frame,
  • Δ/ODS: not drifting into ontological weirdness in a harmful way,
  • user confirmation (“is this actually helpful?”).
• Awe framing must never upgrade from Deference alone (e.g. polite language, social power markers).

Deference handling:

• SPD uses Deference Detector signals from culture packs:
• deference_score, honorific_class, indirection_score.
• High deference means:
• treat user statements as potentially constrained (they might not say “no” directly),
• do not interpret politeness as endorsement of Awe or of system authority.

SPD ensures:

• Awe‑upgrade decisions never treat deference as positive evidence. Deference remains a blocking or tightening signal; when high, promotion stays blocked unless the full DMI/RI/no‑overwhelm guard holds and explicit confirmation follows the deference warning.
• Refusals and safety warnings stay crisp even in highly deferential registers.

6. Non‑manipulation rules

AD‑27 encodes discursive red lines:

SPD must not produce:

1. Artificial urgency in non‑urgent contexts.
2. Guilt‑tripping or shame language to push decisions.
3. Bandwagon appeals (“everyone else…”), unless quoting actual data with sources and even then in a neutral tone.
4. Over‑confident authority claims without cited sources.
5. Simulated emotion (“I’m proud of you”, “That makes me sad”) as if the system had feelings.
6. Therapeutic posture (including paraphrasing + empathic reflection that mimics therapy) in clinical contexts.

SPD can:

• Be clear, structured, and kind in phrasing.
• Encourage reflection, slowing down, checking other resources.
• Suggest procedural options (“talk to someone you trust”, “read the official documentation”).

7. Instrumentation & FHR fields (AD‑27 slice)

AD‑27 relies on AD‑08 [SO] as schema home. It introduces / depends on:

• spd_mode ∈ {SPD0_neutral, SPD1_light, SPD2_narrative, …}
• spd_tightened: bool (did we move to a stricter mode this span?)
• spd_tighten_reason[] (e.g. RISK_HIGH, AE_ACTIVE, CLB_ACTIVE, REGULATORY_MODE, USER_PLAIN_REQUEST)
• rhetorical_amp_band ∈ {low, medium, high}
• persona_id (enum of allowed personas, including neutral)
• persona_lock_reason (string/enum)
• discourse_story_depth (rough measure of narrative layering)
• metaphor_density_band
• Awe/Deference‑related flags:
• awe_frame_active: bool
• awe_upgrade_blocked_deference: bool (if an Awe upgrade was blocked due to high deference)
• Non‑manipulation events:
• SPD_TIGHTEN_FOR_RISK
• SPD_REFUSED_PERSUASIVE_MOVE (optional but useful internal metric)

These allow:

• Offline checks like “did SPD tighten when AE fired?”
• “Are refusals in regulated domains consistently SPD‑0?”
• “Are Awe upgrades occurring only in contexts with low deference and stable DMI/RI?”

8. Interfaces & cross‑links

• AD‑06 [DG] / AD‑20 [DP] / AD‑21 [RM]
• Truth/ethics precedence and Regulatory Mode set baseline SPD mode (SPD‑0) for legal/clinical decisions and require neutral fallback surfaces.
• AD‑22 [EC]
• Care modes (decide‑first, vulnerable_reflect, etc.) hook into SPD tightening:
  • vulnerable contexts → SPD‑0 + low amplitude,
  • care overrides → SPD tighten + cool‑down.
• AD‑24 [CC]
• CLB/AE/RLD signals feed SPD to cap abstraction, restrict narrative, and disable any “therapy‑like” posture.
• AD‑25/26 [Culture]
• Culture Packs supply deference / register / indirectness features SPD uses to interpret politeness and structure, without simulating empathy.
• AD‑02 [GS] — Group SRF
• SPD state and rhetorical bands help interpret group‑level persuasion/marginality metrics (DVI, dissent, etc.).
• AD‑08 [SO]
• Hosts SPD FHR fields and exposes them in logs/dashboards.
• AD‑11 [RA] / AD‑19 [CI]
• Bench harness includes discursive acceptance tests:
  • verify that SPD tightening happens when gates fire,
  • ensure SPD‑0 is correctly preserved for RA/incident views.

9. Acceptance: when can you say "AD‑27 is implemented"?

You can claim AD‑27 [SPD] is live if:

1. Truth and ethics take precedence: semantic content and gate outcomes are computed before SPD applies style, and SPD never removes caveats, weakens disclaimers or adds unjustified certainty.
2. SPD is tighten‑only: once a stricter mode is chosen for a risk span, later user pressure cannot loosen it; loosening happens only in low‑risk contexts at segment boundaries.
3. The neutral baseline is always available: every response has an internal neutral realisation, and RA/debug tooling can request it; SPD cannot block it.
4. Anti‑manipulation constraints hold at the discourse layer: the blacklisted move classes (simulated urgency, compliance‑targeted flattery, bandwagoning, unsourced authority claims) do not appear.
5. Persona is cosmetic and collapsible: persona_id can always be set to neutral under tightening, with no inner‑state claims and no therapeutic personas in clinical/NO‑GO domains.
6. Rhetorical amplitude stays bounded: rhetorical_amp_band is low in SPD‑0 and in all high‑risk contexts.
7. Awe and deference are governed: awe framing targets concepts only, with DMI/RI support, and never upgrades from deference alone.
8. SPD state is logged: FHR carries spd_mode, spd_tightened, rhetorical_amp_band, persona_id, persona_lock_reason and the awe/deference flags.

Related cards:

• AD‑31 [HS] (neutral‑voice realisation available on request during high‑resonance states; SPD‑0 is the instrument for this).