Evaluation

MALT evaluation studies verified read latency, write amplification and the costs of maintaining authenticated relations. Nested Roots can still propagate updates to affected ancestors. Measurements must include payloads, relation materialization, indexes, commitments and publication metadata within the declared boundary.

Current Evaluator

The active malt-evaluation repository uses malt-eval-next: explicit state/run/request plans, capability-checked workers, raw artifacts and reproducible offline analysis. Its Core adapters measure in-memory authentication; runtime and Gateway workers supply specific product capabilities. A Core result does not include network, payload delivery or durable storage unless a separately declared worker measures them.

Non-MALT Git replay uses malt-eval-replay and its own accounting contract. Phase-0 registration and the earlier manuscript-specific dispatch commands are retired; historical artifacts retain their original identities.

Implemented infrastructure does not establish a completed paper campaign. Current reports remain publication_eligible=false until the required release/E0 and claim evidence is implemented and satisfied. The Evaluation and Benchmark Protocol links the current commands, capability matrix and historical sources.

What Comparisons Must Report

Read measurements include end-to-end latency, acceptance and failures, retrieval rounds, fetched bytes, evidence size, and proof generation and verification. Conditional success latency must remain distinguishable from the distribution that preserves failed or timed-out requests.

Write measurements include source changes, derived structures and actual storage: payload and metadata blocks, ArcTable records, commitments, root/index writes and the declared retention history. Attempted writes and newly retained bytes are different quantities.

The Evaluation Story explains these research dimensions. They are intended comparisons, not claims of measured speedups.

Layouts and Historical Results

The current runtime supports flat-v1, hybrid-v1 and rooted-v1 UnixFS layouts. Experiments must identify the selected strategy. The historical v0.0.5 MALT-flat result label is not a current runtime layout identifier.

HAMT is a directory/map-relation baseline. Large-file range comparisons must also identify their chunk layout and payload-verification work. MALT range evidence authenticates element-index bounds, count, an optional metadata CID and ordered chunk CIDs. The client verifies its CID-bound application geometry, hashes fetched chunks and assembles the requested byte interval.

The frozen harness record remains historical provenance. Its schemas, commands and results must not be relabeled as current measurements.