Trie-Trunk: One Million Events

Measured synthetic fixture, not extrapolated: 30 days, five resources, five
branches, 15 Definition fields, adaptive integers level 16, collection indexing
enabled. The fixture cycles through 250 object identities per resource; this is
not one million distinct objects. Messages and mapped values repeat.

Event Pods: 44,779,478 bytes
Tantivy: 80,163,220 bytes
Tries: 3,572,325 bytes
Indexes: 28,528,984 bytes
Summaries: 4,875,617 bytes
Journals & Other: 6,491,505 bytes
Entire Datastore: 168,411,129 bytes
Filesystem Allocated: 179,720,192 bytes
Expanded JSONL: 428,597,753 bytes
Expanded JSONL gzip-6: 25,441,667 bytes
Median Daily Summary Transfer Across Resources: 153,578 bytes

Every event was read back; UUID uniqueness and counts were verified. Tantivy
catch-up completed before accounting. All datastore files, journals and metadata
are included after close. Write PACs are uncompressed; other PACs use default
gzip. Whole-file gzip JSONL has no equivalent indexes, summaries, replay lookup
or independently retrievable hourly segments. These are storage measurements,
not ingestion-throughput or query-latency claims.

The bulk fixture used an 8 GiB native-memory budget per resource. The initial
1 GiB run hit staging backpressure. This override is a limit, not measured RSS;
the results do not establish unpaced ingestion under the default budget.

Source JSONL SHA-256: 9ca9311cc7a62c962aa6488b90c5575b0cbd5766df1af87a578c3a3f553a8ad9
Evidence and reproduction inputs:
https://github.com/AuthenticityLabs/trie-trunk/tree/main/benchmarks/storage/2026-10-24-million-events
See the repository README for reproduction commands. Bulk exports are generated
locally rather than committed. Data is synthetic; production results will vary.