[EPIC] Cognitive actor society — specialized symbolic actors, arbitration, and measurable uncertainty #334

New issue

Open

opened 2026-09-08 02:03:34 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-08 02:03:34 +00:00

Owner

Parent architecture: #624
Builds on: implemented actor-style coordinator foundation #24; integrates with #150 semantics, #629 hypothesis branching, #631 epistemic state and canonical runtime events.

Goal

Evolve Zara's actor architecture from turn coordination into a bounded society of specialized cognitive actors whose disagreement/uncertainty is explicit system state rather than an LLM saying I am uncertain.

Conceptual topology:

perception
  -> intent actor
  -> memory actor
  -> epistemic actor
  -> security actor
  -> planner actor
  -> attention actor
  -> social-context actor
  -> uncertainty/hypothesis actor
         |
         v
      arbiter
         |
         v
  reply / ask / act / wait

This is not permission to create dozens of model agents. Most actors should be deterministic/symbolic/local state machines where that suffices.

Actor contract

Each cognitive actor has:

typed bounded mailbox;
explicit inputs/outputs;
principal/conversation/transaction correlation;
deadline/cancellation;
resource budget;
deterministic degraded/unavailable state;
no implicit capability authority;
observable health/queue metrics without private content.

Arbitration

Research a deterministic arbiter contract for conflicting proposals. Security/authorization invariants outrank preference/optimization actors. An LLM may propose or score where configured but cannot override protected policy by winning a vote.

Measurable uncertainty

Define product-owned metrics over competing typed hypotheses/evidence, such as ambiguity count/entropy/margin where appropriate. Uncertainty state can determine whether Zara:

asks a clarification;
invokes a stronger model;
waits for more evidence;
stages instead of commits;
refuses a risky side effect;
uses a cheaper/local path when confidence is structurally high.

Do not expose hidden model chain-of-thought. Expose only bounded hypotheses/evidence/decision metadata that the runtime intentionally records.

Failure isolation

One actor crash/timeout must not corrupt another actor's state or silently authorize an action. Mailboxes are bounded, stale generations are fenced, loops/deadlocks are detected, and optional actors may degrade without taking down deterministic command paths.

Tests

Cover actor crash/timeout, mailbox pressure, contradictory proposals, security veto, hypothesis entropy transitions, model escalation, cancellation, stale generation, cross-principal isolation, deterministic replay and repeated-turn resource bounds.

Acceptance

Zara can derive an explicit bounded uncertainty/conflict state from specialized actors, route/escalate/clarify accordingly, and remain safe and usable when individual cognitive actors or expensive models are unavailable.

Parent architecture: #624 Builds on: implemented actor-style coordinator foundation #24; integrates with #150 semantics, #629 hypothesis branching, #631 epistemic state and canonical runtime events. ## Goal Evolve Zara's actor architecture from turn coordination into a bounded **society of specialized cognitive actors** whose disagreement/uncertainty is explicit system state rather than an LLM saying `I am uncertain`. Conceptual topology: ```text perception -> intent actor -> memory actor -> epistemic actor -> security actor -> planner actor -> attention actor -> social-context actor -> uncertainty/hypothesis actor | v arbiter | v reply / ask / act / wait ``` This is not permission to create dozens of model agents. Most actors should be deterministic/symbolic/local state machines where that suffices. ## Actor contract Each cognitive actor has: - typed bounded mailbox; - explicit inputs/outputs; - principal/conversation/transaction correlation; - deadline/cancellation; - resource budget; - deterministic degraded/unavailable state; - no implicit capability authority; - observable health/queue metrics without private content. ## Arbitration Research a deterministic arbiter contract for conflicting proposals. Security/authorization invariants outrank preference/optimization actors. An LLM may propose or score where configured but cannot override protected policy by winning a vote. ## Measurable uncertainty Define product-owned metrics over competing typed hypotheses/evidence, such as ambiguity count/entropy/margin where appropriate. Uncertainty state can determine whether Zara: - asks a clarification; - invokes a stronger model; - waits for more evidence; - stages instead of commits; - refuses a risky side effect; - uses a cheaper/local path when confidence is structurally high. Do not expose hidden model chain-of-thought. Expose only bounded hypotheses/evidence/decision metadata that the runtime intentionally records. ## Failure isolation One actor crash/timeout must not corrupt another actor's state or silently authorize an action. Mailboxes are bounded, stale generations are fenced, loops/deadlocks are detected, and optional actors may degrade without taking down deterministic command paths. ## Tests Cover actor crash/timeout, mailbox pressure, contradictory proposals, security veto, hypothesis entropy transitions, model escalation, cancellation, stale generation, cross-principal isolation, deterministic replay and repeated-turn resource bounds. ## Acceptance Zara can derive an explicit bounded uncertainty/conflict state from specialized actors, route/escalate/clarify accordingly, and remain safe and usable when individual cognitive actors or expensive models are unavailable.