SOTA: typed live trace folds for long-horizon agents and observers

Finding

Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers (arXiv:2609.01466, submitted 2026-09-01) evaluates an append-only event ledger incrementally folded into typed run state and compiled into consumer-specific views. The originating team reports roughly 14x to 15x fewer observer input tokens, 5x to 7x lower cost, and accuracy 0.85 to 0.87 versus 0.48 for a budget-capped raw-trace reader. On 120-link sequential-dependency tasks, a deterministic running-state fold solved 30/30 versus 8/30 for full context, while a prompt scratchpad matched its accuracy at lower cost.

Evidence status: originating-team measured with released code, benchmarks, synthetic corpus, and traces. The authors explicitly caution that observer questions were co-designed with the view schema and show order-sensitive tasks where the fold stops helping.

RuV opportunity

Create a provider-independent typed trace-fold contract across MidStream, Core Memory, Ruflo, Dream Machine, MetaHarness, and Cognitum. The differentiator should be deterministic auditability and one canonical run state serving multiple consumers, not an unsupported claim that folds always beat scratchpads.

Proposed primitive

RunEvent -> FoldState -> ConsumerView

Required properties:

• append-only event identity and provenance
• deterministic incremental fold
• versioned schema and fold function digest
• per-consumer views derived from the same state
• explicit unknown/uncovered fields rather than fabricated summaries
• bounded state and resource accounting
• reversible rebuild from the ledger
• order-sensitive fields retain ordered evidence rather than lossy aggregation

Benchmark

Compare raw trace, prompt scratchpad, and typed fold on at least one real Ruflo long-horizon workflow plus an order-sensitive negative-control workload. Report success, observer accuracy, tokens, cost, p50/p95 latency, fold CPU/memory, state size, schema misses, failures, and rebuild determinism.

Promotion requires at least 40% lower delivered observer context cost at matched accuracy, no task-success regression greater than 2 absolute points versus the stronger control, and exact deterministic state rebuild.

The issue names no repository files, tests, or entry points, so first locate the existing trace, run-state, and consumer-view implementations across the named projects. Read the benchmark requirements and negative-control workload description, then establish whether the proposed contract can be implemented in this repository; done requires the stated comparisons, resource metrics, and exact deterministic rebuild.