[EPIC] Branchable reality — hypothetical world state with selective merge #342

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

Parent architecture: #624
Builds on: #51 context ownership, #131 principal-scoped durable state, #150 semantic contracts. Integrates with the epistemic-memory epic under #624.

Goal

Let Zara reason for hours or days inside an explicit hypothetical world without contaminating canonical personal memory/state.

Example:

Assume I move to Germany next year.
What happens to my budget and calendar assumptions?
What if I keep the house?
Merge the German move plan, but keep my current job as unresolved.

Target conceptual model:

world/main
world/hypothesis-move-germany
world/hypothesis-move-germany/keep-house

World contract

Research a typed versioned branch model for propositions, preferences, plans and derived assumptions with:

principal ownership;
branch parent/base revision;
explicit hypothetical status;
proposition provenance/evidence class;
branch-local additions/retractions/contradictions;
derived facts that never silently leak to parent;
bounded lifetime/storage and explicit archive/delete policy;
merge preview and conflict detection.

Merge semantics

A merge is not copy all memory.

select explicit propositions/plan artifacts to promote;
revalidate stale external facts and capability assumptions;
never merge branch-local speculation as user-asserted fact;
conflicts against newer world/main state are surfaced;
merge produces an auditable mutation with source branch/revision;
deleting a branch never deletes canonical facts it merely referenced.

Runtime / model behavior

Every model/tool/context request operating in a branch carries explicit world identity. Tools with real side effects are disabled by default in hypothetical worlds unless a separately authorized simulation/sandbox capability exists. A model cannot escape the branch by merely claiming a fact is real.

UX

Support clear language such as assume, what if, in this scenario, branch this plan, back to reality, compare these scenarios, merge X, and discard this scenario, with visible branch state on UI/voice-sensitive confirmation boundaries.

Tests

Prove branch isolation, nested branch inheritance, contradictions, merge conflicts, stale base, selective promotion, derived-fact nonpromotion, restart persistence, two-principal isolation, and side-effect blocking from hypothetical state.

Acceptance

Zara can sustain a complex hypothetical planning conversation, compare branches and selectively promote reviewed results without poisoning canonical memories/preferences or accidentally executing real-world side effects.

Parent architecture: #624 Builds on: #51 context ownership, #131 principal-scoped durable state, #150 semantic contracts. Integrates with the epistemic-memory epic under #624. ## Goal Let Zara reason for hours or days inside an explicit hypothetical world **without contaminating canonical personal memory/state**. Example: ```text Assume I move to Germany next year. What happens to my budget and calendar assumptions? What if I keep the house? Merge the German move plan, but keep my current job as unresolved. ``` Target conceptual model: ```text world/main world/hypothesis-move-germany world/hypothesis-move-germany/keep-house ``` ## World contract Research a typed versioned branch model for propositions, preferences, plans and derived assumptions with: - principal ownership; - branch parent/base revision; - explicit hypothetical status; - proposition provenance/evidence class; - branch-local additions/retractions/contradictions; - derived facts that never silently leak to parent; - bounded lifetime/storage and explicit archive/delete policy; - merge preview and conflict detection. ## Merge semantics A merge is not `copy all memory`. - select explicit propositions/plan artifacts to promote; - revalidate stale external facts and capability assumptions; - never merge branch-local speculation as user-asserted fact; - conflicts against newer `world/main` state are surfaced; - merge produces an auditable mutation with source branch/revision; - deleting a branch never deletes canonical facts it merely referenced. ## Runtime / model behavior Every model/tool/context request operating in a branch carries explicit world identity. Tools with real side effects are disabled by default in hypothetical worlds unless a separately authorized simulation/sandbox capability exists. A model cannot escape the branch by merely claiming a fact is real. ## UX Support clear language such as `assume`, `what if`, `in this scenario`, `branch this plan`, `back to reality`, `compare these scenarios`, `merge X`, and `discard this scenario`, with visible branch state on UI/voice-sensitive confirmation boundaries. ## Tests Prove branch isolation, nested branch inheritance, contradictions, merge conflicts, stale base, selective promotion, derived-fact nonpromotion, restart persistence, two-principal isolation, and side-effect blocking from hypothetical state. ## Acceptance Zara can sustain a complex hypothetical planning conversation, compare branches and selectively promote reviewed results without poisoning canonical memories/preferences or accidentally executing real-world side effects.