Semantic structure
Warm Index
A bounded index for locating structured state.
ANLS packets
Versioned semantic payloads with explicit templates, coordinates, and slots.
Deltas and lineage
Changes remain connected to predecessor state and transformation evidence.
Reconstruction receipts
Proof that an expected boundary view can be rebuilt from governed inputs.
Privacy classification
Publication and access decisions travel separately from semantic representation.
Synthetic fixture byte counts · 80 records
| Representation | Aggregate bytes | Savings vs literal | Interpretation |
|---|---|---|---|
| Literal canonical JSON | 27,975 B | 0.0000% | Uncompressed literal baseline |
| Canonical JSONL | 27,925 B | 0.1787% | One full canonical record per line |
| zlib level 9 | 2,961 B | 89.4155% | Smallest in every measured cohort |
| Structural template store | 27,272 B | 2.5130% | Larger than literal on the heterogeneous cohort |
| ANLS coordinate-plus-delta | 14,833 B | 46.9777% | Queryable semantic structure with exact reconstruction |
Fixture and measurement boundary
The public fixture contains 80 project-generated synthetic records: 64 repeat-heavy billing, claim, and Microsoft 365 migration-assessment records plus 16 heterogeneous governance notes. It contains no customer data and performs no live database read or host storage scan.
Every total includes the payload plus all template, dictionary, index, envelope, or compression-manifest bytes needed for exact reconstruction. The benchmark does not measure wall-clock latency or establish production throughput.
The purpose of ANLS is queryable semantic structure, lineage, and governed reconstruction—not universal compression superiority. zlib remained smaller in every measured cohort.
Bounded compact-state evidence · tradeoffs kept beside results
| Experiment | Observed result | Required interpretation |
|---|---|---|
| 10K observations | V002 hot-state bytes were 96.45056% lower than the earlier V001 expanded hot store; exact reconstruction and tested query equality passed | A project-internal representation comparison—not total machine memory, network bandwidth, or an optimized conventional-system comparison |
| 50K observations | V002 hot-state bytes were 96.657398% lower than the earlier V001 expanded hot store; exact reconstruction and tested query equality passed | The compact hot-plus-cold total was only about 14.74% smaller than the SQLite reference, while replay was about 9.7× slower |
| 80 synthetic records | 14,833 B versus a 27,975 B literal baseline: 53.02% of literal size, or 46.9777% smaller | Ordinary zlib was substantially smaller; the result supports reconstructible structure, not compression superiority |
| Custom incremental adapter | Changed-index updates were cheaper than rebuilding the complete in-memory index in the tested scenarios | Durable serialization, fsync, recovery, rollback, and a conventional incremental implementation were excluded |
What the measurements support
The evidence supports a narrow engineering idea: invariant structure can be shared, attributable differences can remain explicit, and an expanded boundary view can be reconstructed when its dependencies are present.
The measured adapters did not outperform ordinary preindexed controls for lookup or hydration, and the compact prototype did not outperform SQLite for replay. The next question is usefulness: when does compact, reconstructible structure improve perception and cooperation rather than merely reduce one kind of duplication?