[EPIC] Speculative intent branching — defer clarification with bounded hypothesis actors #340

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
Depends on: #150 typed intent/dialogue state and #625 conversational transactions.

Goal

Stop unnecessary clarification questions when Zara can safely keep multiple interpretations alive and let later context disambiguate them.

Example:

User: Open the server project and fix that actor issue.

hypothesis A: starintel-server
hypothesis B: actor repo

User: Make sure the Common Lisp tests pass.

=> collapse to A

Required architecture

Represent ambiguity as an explicit bounded set of typed hypotheses over IntentFrame / transaction state, not free-form hidden model thoughts.

Each hypothesis needs:

stable ID and source evidence;
candidate intent/slot values;
confidence/evidence metadata appropriate to the resolver;
constraints that would confirm/reject it;
capability/risk projection;
expiry/resource budget;
principal/conversation/transaction ownership.

A lightweight actor/state-machine implementation may advance hypotheses in parallel where useful, but must preserve deterministic authority boundaries.

Clarification policy

Zara may defer clarification only while all surviving hypotheses are safe to keep provisional.

Clarify before:

divergent irreversible side effects;
different authorization requirements;
materially different privacy/resource access;
ambiguity budget/timeout exhaustion;
no later context can safely resolve the distinction.

Never execute multiple real-world branches to discover which one the user meant.

Collapse semantics

Later utterances, explicit corrections, capability availability and deterministic domain facts may prune hypotheses. Keep an auditable bounded reason for pruning without exposing private chain-of-thought.

Tests

Cover two/three-way ambiguity, later disambiguation, equally plausible branches, irreversible divergence forcing clarification, malicious ambiguity explosion, cross-principal isolation, reconnect/stale events, cancellation, and deterministic resource bounds.

Compare UX against eager-clarification fixtures and record when speculative branching genuinely reduces turns without increasing wrong actions.

Acceptance

For supported ambiguous requests, Zara can continue with multiple safe provisional interpretations and collapse them from later evidence, asking the user only when ambiguity becomes action-relevant or cannot be resolved safely.

Parent architecture: #624 Depends on: #150 typed intent/dialogue state and #625 conversational transactions. ## Goal Stop unnecessary clarification questions when Zara can safely keep multiple interpretations alive and let later context disambiguate them. Example: ```text User: Open the server project and fix that actor issue. hypothesis A: starintel-server hypothesis B: actor repo User: Make sure the Common Lisp tests pass. => collapse to A ``` ## Required architecture Represent ambiguity as an explicit bounded set of typed hypotheses over `IntentFrame` / transaction state, not free-form hidden model thoughts. Each hypothesis needs: - stable ID and source evidence; - candidate intent/slot values; - confidence/evidence metadata appropriate to the resolver; - constraints that would confirm/reject it; - capability/risk projection; - expiry/resource budget; - principal/conversation/transaction ownership. A lightweight actor/state-machine implementation may advance hypotheses in parallel where useful, but must preserve deterministic authority boundaries. ## Clarification policy Zara may defer clarification only while all surviving hypotheses are safe to keep provisional. Clarify before: - divergent irreversible side effects; - different authorization requirements; - materially different privacy/resource access; - ambiguity budget/timeout exhaustion; - no later context can safely resolve the distinction. Never execute multiple real-world branches to discover which one the user meant. ## Collapse semantics Later utterances, explicit corrections, capability availability and deterministic domain facts may prune hypotheses. Keep an auditable bounded reason for pruning without exposing private chain-of-thought. ## Tests Cover two/three-way ambiguity, later disambiguation, equally plausible branches, irreversible divergence forcing clarification, malicious ambiguity explosion, cross-principal isolation, reconnect/stale events, cancellation, and deterministic resource bounds. Compare UX against eager-clarification fixtures and record when speculative branching genuinely reduces turns without increasing wrong actions. ## Acceptance For supported ambiguous requests, Zara can continue with multiple safe provisional interpretations and collapse them from later evidence, asking the user only when ambiguity becomes action-relevant or cannot be resolved safely.