ADR-0056: Temporal-interval-aware conflict detection

• Status: Accepted
• Date: 2026-08-27
• Authors: IBEX Harness team
• Milestone: 3.C.3 Temporal-interval-aware conflict detection

Context

Milestone planning text referred to “ADR-0042” for this decision. ADR-0042 is already taken (self-hosted OpenAI-compatible adapter). This record is ADR-0056.

Near-duplicate candidates from ADR-0055 must not all escalate to an LLM. Sequential facts (“prefers Python” in March → “switching to Go” in June) are supersedes, not contradicts. Temporal columns ship in ADR-0047; relationship types in ADR-0049.

Options considered

1) Similarity-only → LLM

Escalate every near-dup. Cheap to implement; wastes LLM budget and misclassifies sequential facts.

2) Interval-overlap-first (chosen)

1. Same subject/attribute and non-overlapping half-open intervals and new.valid_from > candidate.valid_from → supersedes, zero LLM.
2. Intervals overlap (or validity missing/unreliable) → escalate to LLM classifier.
3. No overlap and subjects differ → no conflict.

3) Subject extraction

1. Secondary LLM call — accurate but costs latency/budget on the write path.
2. spaCy CNN (en_core_web_md, already admitted for PII / ADR-0054) — free, <5ms, self-hosted.

Decision: spaCy. Weak fixtures trigger follow-up, not an initial LLM subject extractor.

Decision

1. Half-open intervals [valid_from, valid_until); NULL valid_until = open (+∞), per ADR-0047.
2. Overlap iff a_from < coalesce(b_until, +∞) AND b_from < coalesce(a_until, +∞).
3. Auto-supersede only when non-overlapping, newer valid_from, and subject keys match.
4. Missing valid_from on either side → escalate (never silent skip).
5. Escalation is a pluggable classifier (ConflictClassifier protocol). This milestone ships the seam + metrics; production LLM worker may land later. Tests inject a fake classifier for overlap fixtures.
6. Status taxonomy: use existing CHECK values (superseded, merged_into). Do not invent pending_review without a migration — contradicts outcomes record a contradicts edge and leave review routing to a later write/API milestone.
7. Pipeline order: … → near_dedup → conflict. Supersession persist requires the new memory id (post-insert helper); the stage records pending actions on WriteContext.
8. Metrics: ibex_memory_conflicts_total{outcome} and ibex_memory_conflict_llm_calls_total (llm_call_made counter per detection run).

Consequences

• Sequential-fact fixtures cost zero LLM calls.
• Overlapping same-subject claims still escalate.
• Full write HTTP API remains a later milestone; integration tests drive service + persist helpers directly (same pattern as 3.C.1 / 3.C.2).

Addendum (2026-08-27) — Persist and escalation durability

1. Applying pending_supersede_targets after the new memory receives its DB id is owned by ADR-0057 / milestone 3.C.5.
2. ESCALATE_PENDING is WriteContext-only in 3.C.3. Durable storage is an explicit deferral to 3.C.5: org-scoped table memory_conflict_escalations (not pending_review on memories.status, not a fake relationship type). Tracking: #620.
3. Until 3.C.5 ships, a conflict-resolution worker has nothing queryable to poll for escalations — that gap is intentional and ticketed, not silent.