What happens to storage size?
A controlled 115-byte record becomes 131 bytes under both key-establishment paths with the same record cipher. Key envelopes, signatures and full-database costs have their own accounting.
Inspect the size comparison ↗Evidence / updated September 29, 2026Pick the question.
Pick the question.
Follow the evidence.
Start with the behavior you need. Each public record identifies what was measured, which implementation produced it and how to inspect the result.
Make the next technical discussion concrete.
Do implementations agree?
ACVP demonstration-service records and interoperability checks document tested algorithm behavior. The repository includes public verification instructions and raw inputs.
Inspect implementation evidence ↗Where does protection begin?
The source-bound explanation traces a recipient-specific encrypted message, independent receiver storage keys and the current lexical admission boundary.
Trace the protection path ↗What survived receiver interruption?
The retained 512-record experiment finished with all exact typed records on both hosts after a receiver-process kill and same-store reopen.
Inspect the recovery record ↗Can reads continue during a transition?
Four retained, four-CPU-limited 100,000-record runs cold-verified exact data. One met the declared read deadlines; three had misses. The public record keeps those outcomes and their conditions visible.
Inspect the migration runs ↗How does the engine perform?
Browse the measured workload comparisons, including their hardware, workload definitions and differences between the systems being compared.
Explore database measurements ↗Protection through change and recovery.
Three experiments connect the protection architecture to observable behavior. These results come from bounded Linux prototype tests, with source revisions and artifact hashes retained for review.
Revoked access stays refused after a cache hit.
A permitted reader first received the exact entity from a real warm cache. After owner revocation, the same governed read refused with NO_GRANT before cache lookup or backend access. The owner could still read the exact entity.
Test boundary and provenance
Linux root-crate debug build, revision 41137470c6af. This tests a fresh live EntityReadCache, not a recovered cache or every derived-data path. The Cargo feature named fips is not formal validation.
An interrupted write has an explicit outcome.
Five writer-process termination points produced the expected prior state, new state or refusal. Readers refused mismatches between published data, the authority receipt and the retained-floor test fixture.
Test boundary and provenance
Linux debug prototype at 80ea0cd15eba, rechecked in the 41137470c6af campaign. Process death is distinct from a power loss or whole-host rollback. The retained floor was a test oracle, not a production antirollback service.
Reject unsafe files before trusting their contents.
The bounded startup reader refused tested symbolic links, hard links, unsafe modes, invalid lengths and nonregular files. It also refused a metadata change detected after a real loader read.
Test boundary and provenance
Revision 6f3e41586185: 60 selected Linux tests passed. The privileged wrong-owner filesystem case and non-Linux case were not run. Returned root bytes remain unverified; this result does not establish authenticated bootstrap.
This public collection is a curated summary of retained internal qualification records, not the raw execution bundle. Full artifacts and the corresponding evaluation build can be reviewed through an evaluation. The summary also retains the corrupt-frame coverage gaps from the same campaign.
Application operations / Source review · 24 September 2026
Keep the context that makes a result useful.
These examples connect inspected implementation paths to a practical application. Source inspection establishes what the code expresses; it is not a fresh execution or a qualification of every deployment.
01 / Time series
Turn a missing reading into a useful next step.
A gap inside an expected, deployed measurement window can point an application toward reingestion. Beyond the observed tail, the next step can be to wait for data. Outside the declared deployment, the interval is not applicable.
The classifications use declared cadence, deployment and available samples. They do not diagnose why a sensor went silent.
Source and next check
Source-inspected: 8DB revision 7a3654189fb2, inspected 23–24 September 2026. Consumer: time_series_carrier.rs; time_series_absence.rs. Research finding D01.
Next: execute the required path on the selected build, with representative inputs, access and recovery requirements.
02 / BioTwin / Lineage
When the evidence changes, know which answers to revisit.
BioTwin keeps content and lineage fingerprints separately. A changed dependency can mark affected work as stale and place it in a dependency-ordered rerun queue. The application gets a starting point for review even when the final number has not changed.
This is the inspected BioTwin consumer. Multi-process publication and recovered-history coverage remain production qualification questions; the queue does not establish completed repair.
Source and next check
Source-inspected: 8DB revision 7a3654189fb2, inspected 23–24 September 2026. Consumer: scientific_lineage.rs. Research finding D02.
Next: execute the required path on the selected build, with representative inputs, access and recovery requirements.
03 / BioTwin / Order lattices
Keep the strongest alternatives visible.
BioTwin’s six-axis candidate ordering calls 8DB’s shared lattice operations to retain alternatives that remain incomparable under the declared ordering. The application supplies the scores and directions; the reusable core supplies the structural operation.
The adapter uses fixed-point scores. Missing inputs and rounding need an explicit policy before the result is used to decide.
Source and next check
Source-inspected: 8DB revision 7a3654189fb2, inspected 23–24 September 2026. Consumer: solution_space.rs. Research finding D03.
Next: execute the required path on the selected build, with representative inputs, access and recovery requirements.
These are separate source-inspected consumers. They do not establish a combined runtime result. Download the dated evidence summary · See what the engine and application each provide →
Match the result to the decision.
Source: a result belongs to the implementation and configuration recorded with it.
Workload: a record-cipher measurement answers a different question from a governed write or network exchange.
Reproduction: public artifact verification checks retained evidence. Fresh execution of the engine needs the corresponding evaluation build.
Negative results identify the next engineering question. For the September 16 integrated node-exchange scope and remaining qualification, use the dated capability record.