White paper
Deterministic Computing Architecture
Deterministic computing architecture ensures every output is a pure function of its inputs. This paper presents the hash-chain verification method, the content-addressed caching layer, and the...
Abstract
Ten determinism properties of the RJV platform are published as measured assertions in the evidence register. This paper records each one with the test that settles it, the result of executing those tests at commit b378ee4, and the three classes of claim the platform does not make. Twelve tests were executed on a throwaway PostgreSQL cluster: twelve passed, none failed.
Method
Each published assertion was taken from the register by its urn and resolved to a test in the delivered source. The test was read to confirm it asserts the claim rather than a neighbouring one — a test whose name suggests a property and whose body checks something else is not evidence for the property. The tests were then executed together against a cluster built from an empty database by the repository's own migration applier, running as an unprivileged role with no access to production, and the cluster was destroyed at the end of the run.
Assumptions
- A test demonstrates behaviour for the case it runs. None of the twelve is a proof, and none generalises to inputs nobody constructed.
- The database's privilege boundary is the boundary. The application connects as a role that does not own the tables, so a constraint enforced there holds against a compromised application process and not only against a bug.
- Chain order is given by the entry links, not by timestamps. The host clock is a declared impurity.
- The consistent rewrite case depends on an anchor having been recorded. Without an anchor, a chain rewritten from its first entry verifies.
Results
| Claim | Test | What it asserts |
|---|---|---|
| An operation is refused when the caller holds no current grant for that operation on that resource. | authority::tests::deny_by_default_without_grant | A caller with no grant is denied by the authority layer, not by the caller. |
| The same, at the execution boundary, including a revoked grant. | execution::db_tests::missing_or_revoked_grant_denies_before_any_effect | A revoked grant denies before any effect reaches the store. |
| A changed entry in the evidence chain is detected on verification. | evidence::chain_db_tests::tampered_entry_hash_and_content_are_detected | Altering an entry's link or content fails verification and identifies the entry. |
| A truncated chain, and a consistently rewritten one once an anchor exists, are detected. | evidence::chain_db_tests::anchored_verify_catches_truncation_and_a_consistent_rewrite | Removing the tail is caught; a rewrite is caught because the recomputed head no longer matches the anchor. |
| The role the application connects as cannot change or delete a chain or audit entry, or grant itself authority. | interface::routes::tests::app_role_appends_audit_and_evidence_but_cannot_rewrite_or_self_grant | Append is permitted; alter, delete, and granting itself is refused by the privilege boundary. |
| Balances are derived from posted entries rather than written. | execution::db_tests::journal_post_is_idempotent_and_fails_closed; interface::routes::tests::billing_snapshot_reads_own_wallet_spend_and_balance | A posting is idempotent and fails closed; the billing read derives spend and balance from entries. |
| A passkey assertion can be used once, and one assertion cannot create two sessions under concurrency. | interface::routes::tests::db_passkey_sign_in_one_assertion_yields_one_session_under_concurrency | Two concurrent presentations of one assertion yield exactly one session. |
| The metrics surface refuses an anonymous caller and a caller without the metrics grant, and carries no identifier of an account, session or path. | interface::routes::tests::db_metrics_is_gated_and_exposes_four_families_without_identifiers | Both refusals hold; labels carry no account, session or request path identifier. |
| A record that is a draft, archived or not public is not served and does not appear in the sitemap. | interface::routes::tests::db_sitemap_honours_crawl_policy_and_publication_state | Publication state governs both the address and discovery. |
| Every response, including an error and an unknown address, carries a content policy forbidding executable script and framing, restricting assets to the same origin, requiring transport security and preventing content type guessing. | interface::routes::tests::security_headers_and_no_store_on_every_response; security_headers_cover_the_fallback_and_every_error_response | The headers are present on a normal route, on the fallback and on error responses. |
Twelve tests executed, twelve passed, none failed, none skipped, at commit b378ee4.
Limitations
Three classes are outside the results and are not claimed anywhere on this site. Reproducibility of arbitrary computation: the platform does not pin floating point accumulation across compiler versions, does not schedule on logical ticks, and does not claim that an arbitrary query returns identical bytes on a later run. Model output: sampling and serving infrastructure are outside the platform's control, so the text a model produces is not a reproducible artefact here. Outside services: an operation that calls an external service crosses a boundary the chain cannot see into, and the record states that the call was made and what was returned.
One mapping is weaker than its grade and is stated rather than smoothed. The derived balance claim has no test named after it. The two tests cited cover the posting being idempotent and failing closed, and the billing read deriving spend and balance; neither is a test that shows a direct balance write being refused. The register grades the claim as measured, and the mapping does not fully carry that grade.
Citations
- Evidence register, /research/evidence: ten active assertions, recorded 24 October 2026, grade measured, tier self tested.
- Source
b378ee4:src/authority/mod.rs,src/execution/mod.rs,src/evidence/mod.rs,src/interface/routes.rs. - Run:
cargo test --lib -- --test-threads=1with the eleven named tests, against a scratch cluster created by the repository's verification target and destroyed after the run.