Skip to content
8BraidCreators of
8DB

Engineering note · Evaluator bundle · Frozen baselines · Explicit non-claims

Four Claims, One Pinned Artifact: What an Independent Evaluator Should Be Able to Run

Most product security claims arrive as documents. An evaluator should receive one pinned artifact, run it offline without source access, and get PASS, FAIL or INCONCLUSIVE against frozen baselines for a small set of bounded claims. That is how 8DB is packaged, and it is the shape every data product for high-consequence programmes should be held to.

Published
Reading time8 minutes
Independent verification and validationGovernment software evaluationSecurity evidence packageOSCAL component definitionReproducible security claimsPost-quantum databaseData at rest encryptionAssurance

An evaluator's job is to find out whether a product does what it says, under conditions the evaluator controls, without having to trust the vendor. Most products make that job harder than it needs to be. Claims are broad, evidence is a slide, the test environment is the vendor's, and the words "validated" and "compliant" appear in places where no validation or compliance has happened. The evaluator ends up assessing a narrative.

We built the evaluation package for 8DB the other way round, and this article describes its shape. It is written for the people who do this work: independent evaluators, program security staff, assessors and authorising officials, and the integrators who have to carry a product's evidence into a system package. The shape is not specific to us. It is what a product should look like when it expects to be checked.

Start from what the evaluator receives

The evaluator receives one pinned artifact. It contains a stripped runner, frozen baselines and public known-answer vectors. It does not contain source code, build manifests or repository history, so nothing in the evaluation depends on reading or trusting our code. The artifact carries a manifest hash; the evaluator verifies it before running anything, records the harness source identity and runner build hash, and does not rebuild or substitute a binary. If the version and catalogue listing do not match the pinned catalogue, the evaluation has not started.

The run happens on a host the evaluator controls, network-isolated, with commands, timestamps, operating system and CPU, standard output and error, exit codes, result JSON and artifact hashes preserved as the run record. The workbook is the authoritative record; raw machine output is attached without editing.

Claims small enough to test

The package submits four formal claims, not forty. Each has one bounded statement, named harnesses, controlled preconditions, objective acceptance criteria and explicit non-claims. Anything that does not meet that bar is not a formal claim, however true it may be.

ClaimBounded statementAcceptance
Unified, safe-by-default sealed storageIn one tested store and write path, multiple data shapes persist and return byte-correct results, persisted content is sealed so known plaintext markers are absent from raw storage, and committed data recovers after a simulated crashEvery listed harness passes and matches its frozen baseline
Offline cryptographic governanceLabel and clearance boundaries and policy or epoch changes are enforced through key derivation and decryption outcomes, with core operations executable while the evaluator network is disabledEvery listed harness passes; unauthorised or stale-policy access is rejected in the tested cases; the run is performed under approved isolation
Post-quantum-aligned authenticated sealed-transfer primitiveThe tested primitive uses ML-KEM-1024 and ML-DSA-87 operations and rejects the included tamper, wrong-key, man-in-the-middle and replay cases; the payload can carry already-sealed ciphertext without exposing application plaintextEvery listed harness passes and matches baseline; acceptance is limited to the primitive
Dataset-level integrity with compact proofsThe engine authenticates a dataset with one ML-DSA-87 signature over a SHA-384 Merkle root, validates a logarithmic inclusion proof, and rejects tampering through an independent raw verifier pathBoth harnesses pass and match baseline, including independent verification

Read the acceptance column again. There is no threshold that depends on the host, no performance number, no competitive comparison. A throughput harness exists in the catalogue, and it is deliberately excluded from acceptance because its output varies with hardware.1

Non-claims are part of the claim

Every formal claim carries the list of what it does not assert.2 Non-claims are not modesty. They are the part of the package that tells an evaluator where to stop, and they are what keeps a product PASS from being misread as a system authorisation.

Dispositions that mean something

Four dispositions, defined in advance. PASS: every required harness passes, matches the frozen baseline, no required subtest is skipped, artifact identity matches, no uncontrolled deviation. FAIL: a required harness executes under controlled conditions and fails or mismatches for a reason attributable to the product. INCONCLUSIVE: an artifact, baseline, vector, environment, identity or procedure is missing or changed, or a required subtest skipped. NOT TESTED: execution did not begin or the claim was deferred.

The claim set passes only if all four claims pass. Supporting demonstrations never compensate for a failed or inconclusive formal claim. A missing vector is INCONCLUSIVE, not PASS. These rules remove the room in which optimistic readings grow.

Evidence has classes

Not all evidence is equal, and the package says which is which.

  • Formal. Directly required for one of the four claims: named harnesses and frozen baselines. Only this class can produce PASS, FAIL or INCONCLUSIVE.
  • Corroborating. Supports technical confidence but is not the acceptance mechanism. Our NIST ACVP demonstration-server dispositions for ML-KEM-1024 and ML-DSA-87 sit here.3 Corroborating evidence explains confidence; it cannot cure a failed formal test.
  • Demonstration. Useful capability outside the current approval surface: throughput, deduplication, confidence and provenance features. Recorded separately, no effect on disposition.
  • Planned. Anything that needs a new artifact, environment or authority: a target-hardware build, a production transport, module validation, additional modalities, competitor benchmarks. Must never be described as validated.

When a vendor's marketing puts a class-four item in class-one language, the evaluator has learned something about the vendor.

Accurate security language, pre-agreed

The package includes a table of what may be said and what must not be implied, so the evaluator and the vendor are reading the same words.4 At rest: AES-256-GCM-SIV write-time sealing is exercised and raw media is scanned for known plaintext markers. Post-quantum: ML-KEM-1024 and ML-DSA-87 parameter sets match FIPS 203 and FIPS 204 and are exercised. Key handling: per-page data keys are derived on demand and not persisted in exportable form.

We hold ourselves to that table in public as well. The words "validated" and "certified" do not appear on our site attached to results from a demonstration server.

What a pass produces

A successful evaluation yields an immutable evidence bundle and a bounded claim disposition. Those feed an OSCAL component definition, control implementation narratives and downstream authorisation packages. The authority boundary is preserved: product evidence reduces duplicate work, and the authorising official still evaluates the configured system, mission, environment and residual risk. Change control follows the artifact digest. A new digest means an impact assessment and a rerun of the catalogue unless a narrower regression is approved, and a passing product evaluation does not extend to a differently configured deployment, wrapper service, protocol or hardware platform.

Why we publish the shape

Programs facing the post-quantum transition, disconnected operations and multi-source data fusion are about to evaluate a great many products. Most of the evaluation cost is in arguing about what was claimed. A package with bounded claims, named harnesses, frozen baselines, four dispositions and explicit non-claims moves that argument to before the run, where it is cheap. We would rather be measured against a hard, clear standard than described against a soft one, and we think evaluators should ask every vendor for the same.

What to bring us

  • Your evaluation criteria. If your programme has its own disposition rules, evidence classes or control mappings, send them and we will map the four claims and their harnesses onto them.
  • A request for the evaluator bundle. The pinned, source-free artifact and its baselines are available to evaluators and to the integrators supporting them.
  • The claim you think is missing. If a capability you need is in our demonstration or planned class, tell us; that is how items move into the formal class with harnesses attached.

Write to hello@8braid.com with the subject "Evaluator bundle request". You will hear back from an engineer.

Sources and further reading

Footnotes

  1. Performance is recorded as a demonstration, not a gate. No throughput figure contributes to any disposition.

  2. The non-claims, in full. The sealed-storage claim does not assert that every modality is complete or optimal, or that plaintext can never appear anywhere; it asserts what the harness tests. The governance claim does not assert a complete mandatory-access-control certification, or that a root secret, entropy source or custody process is unnecessary. The transfer claim does not validate any production transport, network, mesh or radio, and the package says so in the same paragraph that states the claim. The integrity claim does not turn a measured overhead figure into an approval threshold until the benchmark is reproducibly packaged.

  3. The package states in the same sentence that the NIST ACVP demonstration server does not issue a CAVP certificate. These dispositions are corroborating evidence only.

  4. What must not be implied, from the same table. At rest: the write-time sealing is not full-device sanitisation, not proof against every possible leak, and not a validated module. Post-quantum: exercising ML-KEM-1024 and ML-DSA-87 is not a CAVP certificate, not FIPS 140-3 module validation, and not complete CNSA 2.0 compliance. Key handling: on-demand derivation is not "no keys exist" and not "no trust anchor is needed".

Continue the technical conversation

Where could this help your work?

Bring a research question, a database workload or an application you want to build. Let’s connect the ideas in this article to an evaluation that matters to your team.

Discuss this work

Continue reading

The next layer of the argument.

Engineering note

Evaluation and Assurance

Every Number on This Site Comes With the Harness That Made It

Vendor benchmarks are the least trusted documents in technology, and they deserve it. Ours carry their quantity, their scope and the harness that produced them, and the rule is enforced in code, so a figure that cannot meet it never reaches you. Here is the rule, the incident that made us adopt it, and how to use it against us.

5 min readDoctrine enforced in code · Applied to every published figureRead article
Engineering note

Evaluation and Assurance

Entity Resolution You Can Defend in Front of a Regulator: 4.28 Million Records, Every Merge Explained, Identical on Every Run

Merging records about people, companies and vessels from hundreds of sources is where intelligence analytics earns or loses trust, and most systems hide the merge behind a score. 8DB does the opposite: every merge decision carries its evidence, every conflict is preserved rather than resolved by fiat, every gap is stated as a gap, relationships stay whole, and a rerun over 4.28 million public sanctions records reproduces the same decisions byte for byte.

6 min readPublic-corpus run · 4.28 million records · Byte-identical replayRead article
Engineering note

Evaluation and Assurance

Make Prompt Injection Harmless to Your Data: Clearance Lives in the Key, So a Hijacked Agent Decrypts Nothing It Was Not Cleared For

An agent reasoning at the tactical edge is only as trustworthy as the data it can reach. Put access control in a filter in front of plaintext and every mistake in the filter is a disclosure. Make clearance and policy inputs to key derivation and an unauthorised read yields ciphertext and nothing else. That is what 8DB does for sensemaking systems, and an independent evaluator can run the claim on an isolated host.

6 min readTested claims · Evaluator harnesses · Explicit non-claimsRead article