One evidence system for every Casita layer. It shows where results exist, where coverage is missing, and which architectural limits fail before production finds them.
No composite score. A fast operation that loses verification, grows memory with history, or explodes backend requests is a failed design.
Repository size is not one number. Each target isolates a resource boundary so a result explains what broke; release evidence must also include a real large repository.
Prove that catalog publication, open, lookup, sync, fsck, and collection remain sharded and bounded-memory at roughly two billion 256 KiB chunks; payload materialization is not required for catalog-only runs.
Exercise sustained ingest, clone, verification, compaction, collection, and cold restoration against bytes that cannot be compressed away.
Expose per-object metadata, lookup, traversal, publication, and backend-request amplification independently of payload bytes.
Stress very wide trees, checkout, directory encoding, query, and lazy workspace behavior.
Measure deep-history traversal, reachability, commit graphs, incremental import, and retention metadata.
Measure manifest, MIDX, WAL replay, compaction debt, startup, and point-lookup degradation.
Prove that unchanged import, warm sync, incremental fetch, and indexing follow the delta instead of total retained state.
Exercise spill, streaming, pack resolution, verification, and collection while total repository state is much larger than RAM.
Restore a serving node from authoritative object storage and indexes, then verify time to first correct read and full readiness.
The registry is also the roadmap. Planned and partial suites stay visible until their measurements exist.
Hashing, verification, chunking, compression, write admission, and deduplication.
4 operations · no published resultFresh filesystem import, incremental import, checkout, sync, verification, and collection.
8 operations · results availableLocal, memory, and S3-compatible request amplification, range reads, multipart writes, caching, and byte budgets.
6 operations · no published resultSnapshot reads, record publication, unrelated-writer retries, competing roots, and multi-process coordination.
5 operations · no published resultForced-spill closure verification and collection across wide-tree object-count and spill-budget sweeps; depth and restart sweeps remain future work.
2 operations · results availableLocal, SSH, and S3 transfer with path depth, deduplication, batching, RTT, bandwidth, chunk reuse, and interrupted receivers.
6 operations · results availableMany-object, delta-heavy, physical-pack, and wide-tree import, bind, clone, and incremental-fetch scaling.
7 operations · no published resultCross-suite lifecycle runs at 30+ GiB physical bytes, 10M+ objects, 5M paths, 1M revisions, and 10k incremental generations.
11 operations · no published resultMark/sweep, physical reclamation, corruption discovery, concurrent holds, spill, and ENOSPC recovery.
5 operations · no published resultStreaming export/import, duplicate catalogs, bounded decoding, and malformed input behavior.
5 operations · no published resultCrash points, backend failures, stale processes, partial writes, retries, and deterministic recovery time.
5 operations · no published resultEvery value is normalized from an authoritative runner result. Filters never manufacture zeroes for missing metrics.
Dots mean the suite is responsible for measuring the dimension, not that the work is complete.
| Suite | correctness | latency | throughput | memory | storage | scale | concurrency | recovery |
|---|---|---|---|---|---|---|---|---|
| Core primitives | ● | ● | ● | ● | ● | — | — | — |
| Repository workflows | ● | ● | ● | ● | ● | ● | — | — |
| Blob and object-store backends | ● | ● | ● | ● | ● | ● | ● | ● |
| State, publication, and CAS | ● | ● | ● | ● | — | ● | ● | ● |
| Graph traversal and spill | ● | ● | ● | ● | ● | ● | — | ● |
| Transfer and synchronization | ● | ● | ● | ● | ● | ● | ● | ● |
| Native Git | ● | ● | ● | ● | ● | ● | ● | ● |
| Huge repository frontier | ● | ● | ● | ● | ● | ● | ● | ● |
| Collection and integrity | ● | ● | ● | ● | ● | ● | ● | ● |
| Casitar streaming format | ● | ● | ● | ● | ● | ● | — | ● |
| Fault injection and recovery | ● | ● | — | ● | ● | ● | ● | ● |
Protocol errors, limit failures, and validation mismatches are benchmark results.