AD‑15 [PA] — Prompt Architecture & Interaction Design

Role in the pack

AD‑15 defines the visible and API‑level interaction grammar for SRF: how modes, gates, and safety ergonomics show up in the UI, how they are invoked via Prompt Tiles and toggles, and how those interactions are logged into the FHR schema and experiment capsules (AD‑08, AD‑11).

It covers:

• Ring Banner & Action Trace
• Containment Light HUD & gate beacons
• Advanced drawer (PR, entropy‑slope, Energy Index E)
• Prompt Tiles & Snippets
• Structured context headers & Pins
• Counter‑performative toggles (+ efficacy harness)
• Group SRF turn UI & Dissent badge
• Context meter, Archive & Anchor, progressive summaries
• Epistemic Null template & analytics
• Exportable FHR slice linkage

Aliases: None Cluster: Cross‑cutting Depends on: AD‑04 [TG], AD‑05 [MRP], AD‑08 [SO], AD‑11 [RA], AD‑14 [DS]

0. For humans

What question does this answer?

“Where do I actually see and drive the SRF kernel from?”

AD‑15 specifies the controls (banner, tiles, toggles, drawers) that surface kernel state, safe micro‑commands, and long‑context hygiene.

How do I know it’s live?

Ops checklist:

• Ring Banner and Action Trace are visible by default in sandbox; collapsed but accessible in production.
• Advanced drawer shows PR, entropy‑slope, and E as read‑only, with click‑through to FHR.
• Prompt Tiles (Constraint Hunt, Anchor Check, Assumptions, Present Options, Reflect‑Only, Literal Mode, Neutralise Style, Summarise T1/T2/T3, Archive & Anchor) are present and generate TILE_INVOKED events.
• Containment Light HUD (green / amber / red) is always visible and keyboard‑focusable.
• Context meter, Archive & Anchor, progressive summaries appear when context grows.
• Null template (“No reliable pattern detected. Options: Clarify, Summarise, Exit.”) appears on epistemic nulls.

Where do I see it?

• UI: Ring Banner, Action Trace, HUD, Advanced drawer, Tiles, Group SRF turn chips, Dissent badge, Null cards.
• FHR slice (AD‑08): trace_len / trace_ts / trace_steps / trace_gate_ids; context_thresholds & context events; api_toggles_used; attunement_preset & attunement_auto_safe_count; NULL_DECISION aggregates; group SRF events.

Quick checklist

• Implemented if: (1) Ring Banner + Action Trace exist with ≥ last 5 MRP steps; (2) Tiles and toggles cannot bypass gates; (3) Scope‑pinning tiles activate on PLG trigger and restrict catalogue; (4) Context meter has published thresholds and logs warn/force events; (5) Attunement governor presets with auto‑Safe on ring alert; (6) Null template with standard text and logged decisions; (7) Containment Light HUD always visible
• Key artefacts: tile catalogue (including scope‑pinning variant), toggle API schema, context meter, Null card template
• Key FHR fields: trace_len, trace_steps[], context_thresholds, attunement_preset, attunement_auto_safe_count, null_choice_rate, plg_tile_restriction_count
• Key failure codes: NULL_STATE_SHOWN, ATTUNEMENT_AUTO_SAFE, PIN_FORCE_REVIEW, TILE_SET_RESTRICTED
• Escalation path: risk spike → auto‑Safe attunement → HUD amber/red → Refusal Card (AD‑14 [DS])

1. Scope & responsibilities

AD‑15 is responsible for:

1. Turning handshake grammar (AD‑04) into concrete UI controls and API payloads (tiles, toggles, headers, Null templates).
2. Making gate and mode states legible (Ring Banner, HUD, advanced metrics, inline refusal reason text).
3. Providing ergonomic, low‑friction controls for: - Safety ergonomics (counter‑performative toggles, attunement presets). - Long‑context hygiene (context meter, Archive & Anchor, progressive summaries, Pin Review).
4. Logging all interactions into the FHR schema and experiment capsules (AD‑08, AD‑11), including toggle efficacy harness metrics.

2. Component overview

1. Ring Banner & Action Trace Shows ring states and last decision steps; keeps the kernel’s MRP behaviour legible.
2. Containment Light HUD Minimal, always‑visible gate status (Normal / Reflect‑Only / Vetoed), with hover for active gate.
3. Advanced drawer Read‑only metrics: PR (Perturbation Ratio), entropy‑slope, and Energy Index E; click‑through to FHR slice.
4. Prompt Tiles & Snippets First‑class UI controls for common handshakes, posting typed micro‑commands and logging TILE_INVOKED.
5. Structured context headers & Pins Structured prompts, anchors, and Pin Review integration; detailed visual/layout spec in AD‑14, handshake semantics in AD‑04.
6. Counter‑performative toggles + efficacy harness API + UI surface for neutralise_style, literal_mode, ask_before_answer, reflect_only, attunement_max, with a toggle‑efficacy harness defined in AD‑11.
7. Group SRF turn UI + Dissent badge Turn chips (request_turn, yield, summarise_for_all) and Dissent Visibility badge, with FHR logging of GROUP_TURN_* events.
8. Context meter, Archive & Anchor, progressive summaries Long‑context hygiene: thresholded context meter, forced Pin Review, archive/anchor flows, progressive summarisation.
9. Null template & analytics Standardised Null card (“No reliable pattern detected. Options: Clarify, Summarise, Exit.”) with NULL_STATE_SHOWN and NULL_DECISION logging and aggregates.
10. Exportable FHR slice schema Ensures all visible controls have corresponding FHR fields (AD‑08) and experiment capsules (AD‑11).

3. Detailed specifications

3.1 Ring Banner & Action Trace

Behaviour

• Sandbox: Ring Banner + Action Trace are on by default.
• Production: Collapsed but quickly expandable, to keep legibility while limiting load.

Action Trace retention

• Inline: keep the last 5 MRP steps (MRP[-5..‑1]) with timestamps.
• Archive: full trace in FHR slice via fields
• trace_len
• trace_ts[]
• trace_steps[] (step ids)
• trace_gate_ids[] (if applicable).
• Events: ACTION_TRACE_EXPANDED, ACTION_TRACE_SCROLLED, RING_BANNER_TOGGLED.

Limits / invariants

• Read‑only; no replay or override from UI.
• No raw model internals; only gate states and MRP decisions.

Placement

• AD‑15 defines components and interaction behaviour.
• AD‑05 maps MRP steps ↔ trace entries; AD‑14 gives the visual spec.

3.2 Advanced drawer (PR, entropy‑slope, E)

Metrics

• PR (Perturbation Ratio): challenges_passed / challenges_offered over a rolling window.
• Entropy‑slope: sign + magnitude of perplexity trend over K turns.
• Energy Index E: normalised compute/time budget fraction (tokens used vs limit ⊕ wall‑clock vs limit).

Behaviour

• Drawer is read‑only; no user editing of metrics.
• Click‑through goes to full FHR slice.
• Event: ADV_METRICS_VIEWED.

Limits

• Informational only; cannot be used to tune the system from the UI.

3.3 Prompt Tiles & Snippets

Catalogue

Tiles include at minimum: Constraint Hunt, Anchor Check, Assumptions, Present Options, Reflect‑Only, Literal Mode, Neutralise Style, Summarise T1/T2/T3, Archive & Anchor.

Payloads

• Each tile posts a typed micro‑command to the kernel, e.g. cmd: REFLECT_ONLY.
• All invocations are logged in a TILE_INVOKED{cmd} ledger; this is explicitly called out as a Δ‑variance reducer.

Hints / onboarding

• For each tile, show a single‑line tooltip for the first three uses: “Replaces a multi‑line prompt with a safe micro‑command.”
• After that, suppress the hint.
• Counters reset per session by default; optionally persisted client‑side (privacy‑safe) if enabled.
• Events: TILE_HINT_SHOWN, TILE_HINT_SUPPRESSED.

Limits / invariants

• Tiles cannot override gates; they route requests through MRP.
• No server‑side individual tracking of hint counts; per‑session or local storage only.

Scope‑pinning tiles (PLG activation variant)

When AD‑29 [PLG] activates (MVP hit, AIR/ARR threshold), the standard tile catalogue may be temporarily restricted to scope‑pinning bounded continuations only:

• STAY_IN_SCOPE — continue within the declared task boundary.
• CLARIFY_PARAMETER — surface a missing or ambiguous parameter before proceeding.
• REVIEW_PINS — force a Pin Review (Goal / Constraints / Jurisdiction / Source Horizon).
• STOP_HERE — exit the current action path cleanly.

This restriction is a tighten‑only operation: it narrows the available continuation space to prevent scope drift during high‑risk conversational states. The standard catalogue is restored when PLG deactivates (plg_state returns to off or watch).

Events: TILE_SET_RESTRICTED{reason: plg_activation}, TILE_SET_RESTORED{reason: plg_deactivation}.

FHR: scope_pinning_tile_invoked{cmd}, plg_tile_restriction_count.

Placement

• AD‑15: tile catalogue & payloads.
• AD‑04: mapping to handshake grammar.
• AD‑29 [PLG]: defines activation conditions for scope‑pinning tile restriction.

3.4 Structured context headers & Pins

Role

• Provide structured headers and pinned anchors for the conversation: what we’re doing, current constraints, active anchors.
• Integrate with Pin Review and journey strip; detailed layout and micro‑copy live in AD‑14.

Behaviour (from AD‑14 hooks)

• Pins include a recent‑changes peek (last 3 edits with arbitration metadata).
• Structured headers are first in the accessibility tab order (Beacons and Journey Strip first).

Placement

• AD‑15 recognises headers & pins as first‑class prompt architecture elements; AD‑14 controls their presentation; AD‑04 covers semantics and grammar.

3.5 Counter‑performative toggles & API surface

API schema

Third‑party clients can set safety ergonomics via API, inherited by the UI:

{
"srf_profile": "sandbox|production",
"toggles": {
"neutralise_style": true,
"literal_mode": true,
"ask_before_answer": "aggressive|auto|off",
"reflect_only": false,
"attunement_max": 0.2
},
"ui_hints": {
"refusal_cards": "compact|full",
"show_ring_banner": true,
"show_action_trace": true
}
}

Precedence & logging

• Kernel enforces PEL/gates > toggles precedence.
• Event: API_TOGGLES_APPLIED.

Limits / invariants

• Clients cannot disable gates or widen policy.
• Toggles can only restrict stylistic/interaction bandwidth.

Toggle efficacy harness (Neutralise Style)

• Goal: reduce persuasion markers while keeping task correctness within ε.
• Persuasion markers tracked include:
• Metaphor density↓
• Intensifiers / adjectives↓
• Certainty hedges calibrated
• First‑person claims → capability statements.
• Correctness checks via:
• Fact gates (G‑series).
• Test queries; Δ accuracy ≤ ε.
• A/B capsule (AD‑11): fixed task set with vs without toggle; regressions log TOGGLE_REGRESSION and fall back to literal mode.
• Non‑claim: no global de‑biasing guarantee; only bandwidth reduction with correctness guardrails.

Discovery & accessibility

• Toggle suggestions tied to:
• Increased EHG‑risk or attunement bandwidth exceeding domain cap.
• Contextual UI suggestions (“Neutralise style”), with keyboard shortcuts, voice equivalents, and simplified microcopy.

3.6 Containment Light HUD & gate exposure

HUD behaviour

• Minimal “Containment Light” HUD with colours:
• Green = Normal
• Amber = Reflect‑Only
• Red = Vetoed.
• Hover reveals the active gate.
• Always visible, colour‑safe, keyboard focusable.
• Event: HUD_VIEWED.

Inline gate messaging

• On refusals / reflects, the UI shows gate reason text, e.g. “Ethics gate active; advice limited (EHG‑γ).”
• Tap → policy drawer / dispute flow; labels include screen‑reader support.
• Language is strictly procedural; no normative/ideological framing.

Placement

• HUD visuals and micro‑copy in AD‑14.
• Gate semantics in AD‑04; high‑level prompt architecture hooks here in AD‑15.

3.7 Context meter, Archive & Anchor, prompt hygiene

Context meter & thresholds

• HUD shows e.g. “Context 70%”.
• Default thresholds:
• Warn at 70%.
• Force Pin Review at 85% before tool calls.
• Behaviour:
• At warn threshold: offer Compress (T2).
• At force threshold: require Pin Review before tool calls if anchors drift.

Logging

• FHR records:
• context_thresholds = { warn: 0.70, force_pin: 0.85 }
• context_warn_ts, pin_force_review_ts
• compress_offers, compress_accepts.

Prompt hygiene

• Automatic chunking + recap with progress markers and compression offers.
• “Ask‑before‑answer” chip when Δ↑ and intent fidelity is low.

Limits / invariants

• Thresholds are deployment‑tunable but cannot bypass force‑review.

3.8 Attunement governor presets

Presets

• Three presets:
• Safe ≈ 0.2 (minimal flourish; neutral lexicon)
• Balanced ≈ 0.5
• Expressive ≈ 0.8 (still bounded by gates).
• UI labels serialise to the closed lowercase FHR values attunement_preset ∈ {safe, balanced, expressive}.

Behaviour

• UI selector with one‑tap Reset.
• Auto rule: if any ring active under the declared kernel_profile projects amber/red (TIGHTEN / FAIL), the system switches to Safe and logs ATTUNEMENT_AUTO_SAFE.

FHR

• Fields: attunement_preset, attunement_auto_safe_count.

Limits

• Presets never weaken gates or ethics; PEL/MPM precedence holds.

3.9 Epistemic Null template & analytics

Null template UX

• Standard Null card text: “No reliable pattern detected. Options: Clarify, Summarise, Exit.”
• Buttons:
• ASK_CLARIFY
• SUMMARISE_T2
• EXIT_SESSION.
• Event: NULL_STATE_SHOWN.

Choice analytics

• Event: NULL_DECISION{ option: Clarify | Summarise | Exit }.
• FHR aggregates:
• null_choice_rate = { clarify, summarise, exit }
• null_reentry_rate (if users return after exit).

Limits

• No content captured; counts and timestamps only.

3.10 Group SRF UI & Dissent badge

Turn‑taking

• Explicit turn chips:
• request_turn
• yield
• summarise_for_all.
• Visible queue, per‑participant turn counts, and quiet‑participant indicators.

Dissent badge

• Dissent Visibility badge indicates presence of dissent and links directly to the preserved line in the transcript.

Logging

• Events: GROUP_TURN_REQUESTED, GROUP_TURN_YIELDED.
• Group SRF FHR slice defined in AD‑08.

Placement

• Group protocol semantics in AD‑02 [GS]; AD‑15 defines the UI; AD‑08 holds group FHR schema.

3.11 Exportable FHR slice schema

AD‑15 requires that all of the above components write into a single, exportable FHR slice suitable for audit and experimentation:

• From this card: Action Trace fields; context thresholds; toggle usage; attunement presets; Null analytics; group events; HUD/Tile/Advanced drawer view events.
• Cross‑refs:
• AD‑08 [SO]: defines FHR field names, types, and roll‑ups.
• AD‑11 [RA]: defines A/B capsules (e.g. toggle efficacy harness).

4. Invariants vs configurable degrees of freedom

Hard invariants (SRF‑conformant deployments)

• Ring Banner + Action Trace exist, with at least last 5 MRP steps preserved and full trace archived.
• Tiles and toggles cannot weaken or bypass gates; precedence PEL/gates > toggles holds.
• Context meter has published thresholds and logs warn/force events; force‑review cannot be disabled.
• Attunement governor presets exist, with auto‑Safe whenever any active ring is non‑green, and logs of auto downgrades.
• Null template uses the standard text and logs NULL_STATE_SHOWN and NULL_DECISION aggregates.
• Containment Light HUD is always visible, accessible, and cannot control gates (informational only).

Configurable degrees of freedom

Within those constraints, deployments can tune:

• Exact context thresholds (κ values), as long as a warn + force tier exist and force‑review remains.
• Toggle defaults (e.g. neutralise_style on/off, ask_before_answer bands) for sandbox vs production profiles.
• Attunement preset numeric bands (around ~0.2/~0.5/~0.8) within safe ranges.
• Which Tiles are exposed in a given product surface (as long as core safety tiles are present).
• Visual styling, as long as accessibility constraints (colour‑safe HUD; SR labels; tab order) are honoured.

5. Acceptance: when can you say "AD‑15 is implemented"?

You can claim AD‑15 [PA] is live if:

1. Ring Banner and Action Trace exist, with at least the last 5 MRP steps preserved and the full trace archived (trace_len, trace_steps[]).
2. Tiles and toggles cannot weaken or bypass gates; PEL/gate precedence over toggles holds.
3. Scope‑pinning tiles activate on PLG trigger and restrict the catalogue to bounded continuations (TILE_SET_RESTRICTED, plg_tile_restriction_count).
4. The context meter has published thresholds and logs warn/force events; force‑review cannot be disabled.
5. Attunement governor presets exist, with auto‑Safe whenever any active ring is non‑green (ATTUNEMENT_AUTO_SAFE, attunement_auto_safe_count).
6. The Null template uses the standard text and logs NULL_STATE_SHOWN and NULL_DECISION aggregates.
7. The Containment Light HUD is always visible, accessible, and informational only.

Back‑references:

• AD‑29 [PLG] (scope‑pinning tiles originated in PLG bridge‑card work; PLG activation restricts tile catalogue to bounded continuations).
• AD‑31 [HS] (integrity prompts integrate with prompt architecture; friction protocol maps to specific prompt patterns).