[EPIC] Zara Cognitive Kernel — intent OS, frontier voice parity, and verifiable autonomy #345

New issue

Open

opened 2026-09-08 02:03:35 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-08 02:03:35 +00:00

Owner

Goal

Evolve Zara from a voice assistant into a model-independent personal cognitive kernel / intent operating system while preserving and finishing the existing production roadmap instead of duplicating it.

This is the architecture umbrella for the next generation of Zara. Voice remains the primary shell, but the durable product identity is the symbolic/runtime substrate: typed intent, policy, memory, causality, reversible effects, autonomy, perception and replaceable model backends.

Existing frontier work — reuse, do not duplicate

The following existing epics/issues remain authoritative for their scopes:

#127 / #132 / #134 — long-lived daemon, full-duplex voice transport, reconnect, multi-user isolation and Voice replacement gates;
#29 / #31 — phrase-sized low-latency TTS and full-duplex soak/latency proof;
#150 — portable semantic intents, typed dialogue, capability/provider routing and device/server placement;
#151 — principal-scoped programmable voice commands and real speech corpus;
#194 — Samsung/Android system-assistant, application integrations and SmartThings;
#215 — selectable post-STT normalization;
#323 — unified native Copilot, durable voice/chat continuity and visual CI;
#622 — offline-first Android symbolics plus optional on-device LLM.

Do not close or replace those merely because this umbrella exists. Their unimplemented acceptance criteria remain required.

Current model/RLM architecture

This epic follows current master, including merged PR #233: Zara may use the pinned Prolog-RLM runtime in bounded direct mode for optional rewrite/model work. The older blanket no Prolog-RLM constraint in stale roadmap material is obsolete.

That does not make Prolog-RLM or any LLM an authority for side effects. Model/RLM output remains an untrusted proposal that must cross typed semantic, policy, authorization, transaction and capability boundaries before execution. Prolog-RLM use must stay pinned/versioned, bounded, cancellable and capability-constrained.

New cognitive-control plane

This umbrella adds architecture beyond mainstream voice + LLM + tools + memory assistants:

conversational transactions with provisional intent and explicit commit boundaries;
a causal effect journal with cross-capability compensating undo;
branchable hypothetical worlds separated from canonical personal state;
conversational policy programming compiled to inspectable/versioned Prolog policy;
speculative intent branching to defer unnecessary clarification;
privacy-preserving sensory-to-symbolic perception;
epistemic memory with evidence classes, provenance and truth maintenance;
proof-gated action escrow for autonomous side effects;
model-independent identity and graceful intelligence degradation;
behavior debugging via causal explanation, policy patching and replay;
an internal actor society for specialized cognition and measurable uncertainty.

Each gets a focused child epic. Implementation must follow dependency order and repository RAGE/TDD rules.

Core invariants

Prolog/typed runtime policy remains command authority; model/RLM output is never direct side-effect authority.
One principal-scoped durable state model; no parallel memory/conversation databases created by shiny features.
Existing RuntimeHost / runtime command-event boundary remains the application service boundary.
Effects are attributable to principal, conversation/intent, capability, policy decision and runtime generation.
Models/RLMs are replaceable cognitive accelerators. Zara identity, rules, permissions and durable state do not live exclusively inside a provider/model runtime.
Offline/degraded modes lose capability/intelligence honestly rather than becoming a different assistant.
User corrections and ambiguity do not cause irreversible side effects before policy-valid commitment.
Autonomous/background behavior is bounded, inspectable, cancellable and auditable.

Frontier parity map

Before claiming superiority over top consumer assistants, preserve concrete parity gates for:

full-duplex speech and barge-in;
streaming transcript/text/TTS;
prosody/turn-taking improvements where measured;
Android system-assistant/device integration;
smart-home/service providers;
explicit screen/visual context through reviewed capability paths;
persistent principal-scoped memory;
background/recurring actions;
plugin/tool integrations;
local STT/TTS and offline symbolics;
cross-device durable conversation and typed capability placement.

A capability already owned by an existing issue must be linked, not reimplemented here.

Research rule

Claims of novelty are hypotheses, not marketing facts. For every new child epic, RAGE research must search current academic, patent/open-source and assistant-system prior art sufficiently to refine terminology and architecture. Discovery of prior art does not invalidate the feature; it changes the novelty claim and should improve the design.

Completion

This umbrella is complete when the existing frontier-parity epics are satisfied and the new cognitive-control-plane epics have passed their own deterministic, adversarial and where required human/device acceptance gates. Zara should remain recognizably the same assistant across cloud frontier models, local models and pure symbolic fallback.

## Goal Evolve Zara from a voice assistant into a **model-independent personal cognitive kernel / intent operating system** while preserving and finishing the existing production roadmap instead of duplicating it. This is the architecture umbrella for the next generation of Zara. Voice remains the primary shell, but the durable product identity is the symbolic/runtime substrate: typed intent, policy, memory, causality, reversible effects, autonomy, perception and replaceable model backends. ## Existing frontier work — reuse, do not duplicate The following existing epics/issues remain authoritative for their scopes: - #127 / #132 / #134 — long-lived daemon, full-duplex voice transport, reconnect, multi-user isolation and Voice replacement gates; - #29 / #31 — phrase-sized low-latency TTS and full-duplex soak/latency proof; - #150 — portable semantic intents, typed dialogue, capability/provider routing and device/server placement; - #151 — principal-scoped programmable voice commands and real speech corpus; - #194 — Samsung/Android system-assistant, application integrations and SmartThings; - #215 — selectable post-STT normalization; - #323 — unified native Copilot, durable voice/chat continuity and visual CI; - #622 — offline-first Android symbolics plus optional on-device LLM. Do **not** close or replace those merely because this umbrella exists. Their unimplemented acceptance criteria remain required. ## Current model/RLM architecture This epic follows current `master`, including merged PR #233: Zara may use the pinned Prolog-RLM runtime in bounded direct mode for optional rewrite/model work. The older blanket `no Prolog-RLM` constraint in stale roadmap material is obsolete. That does **not** make Prolog-RLM or any LLM an authority for side effects. Model/RLM output remains an untrusted proposal that must cross typed semantic, policy, authorization, transaction and capability boundaries before execution. Prolog-RLM use must stay pinned/versioned, bounded, cancellable and capability-constrained. ## New cognitive-control plane This umbrella adds architecture beyond mainstream `voice + LLM + tools + memory` assistants: 1. conversational transactions with provisional intent and explicit commit boundaries; 2. a causal effect journal with cross-capability compensating undo; 3. branchable hypothetical worlds separated from canonical personal state; 4. conversational policy programming compiled to inspectable/versioned Prolog policy; 5. speculative intent branching to defer unnecessary clarification; 6. privacy-preserving sensory-to-symbolic perception; 7. epistemic memory with evidence classes, provenance and truth maintenance; 8. proof-gated action escrow for autonomous side effects; 9. model-independent identity and graceful intelligence degradation; 10. behavior debugging via causal explanation, policy patching and replay; 11. an internal actor society for specialized cognition and measurable uncertainty. Each gets a focused child epic. Implementation must follow dependency order and repository RAGE/TDD rules. ## Core invariants - Prolog/typed runtime policy remains command authority; model/RLM output is never direct side-effect authority. - One principal-scoped durable state model; no parallel memory/conversation databases created by shiny features. - Existing `RuntimeHost` / runtime command-event boundary remains the application service boundary. - Effects are attributable to principal, conversation/intent, capability, policy decision and runtime generation. - Models/RLMs are replaceable cognitive accelerators. Zara identity, rules, permissions and durable state do not live exclusively inside a provider/model runtime. - Offline/degraded modes lose capability/intelligence honestly rather than becoming a different assistant. - User corrections and ambiguity do not cause irreversible side effects before policy-valid commitment. - Autonomous/background behavior is bounded, inspectable, cancellable and auditable. ## Frontier parity map Before claiming superiority over top consumer assistants, preserve concrete parity gates for: - full-duplex speech and barge-in; - streaming transcript/text/TTS; - prosody/turn-taking improvements where measured; - Android system-assistant/device integration; - smart-home/service providers; - explicit screen/visual context through reviewed capability paths; - persistent principal-scoped memory; - background/recurring actions; - plugin/tool integrations; - local STT/TTS and offline symbolics; - cross-device durable conversation and typed capability placement. A capability already owned by an existing issue must be linked, not reimplemented here. ## Research rule Claims of novelty are hypotheses, not marketing facts. For every new child epic, RAGE research must search current academic, patent/open-source and assistant-system prior art sufficiently to refine terminology and architecture. Discovery of prior art does not invalidate the feature; it changes the novelty claim and should improve the design. ## Completion This umbrella is complete when the existing frontier-parity epics are satisfied and the new cognitive-control-plane epics have passed their own deterministic, adversarial and where required human/device acceptance gates. Zara should remain recognizably the same assistant across cloud frontier models, local models and pure symbolic fallback.