White paper

Post-Probabilistic Security Systems

Post-probabilistic security moves from confidence scores to deterministic verification via tamper-evident audit chains. This paper presents the threat model, chain structure, and measured overhead...

Published
21 October 2026
Status
Working document
Author
RJV Technologies Ltd

The register is the argument

RJV's position is that a security claim is worth exactly the evidence attached to it, and that a probability is not evidence. The platform's control register is the working instance: 140 requirements across eleven domains, each with an owner, the implementation files that carry it, and the negative tests that exercise it. Twenty-five are marked verified. One hundred and fifteen are not.

DomainRequirements
Agents20
Application18
Data14
Infrastructure14
Architecture12
Identity12
Supply chain12
Authorisation10
Cryptography10
Detection10
Recovery8

Three hundred and 329 negative tests are named against those requirements. A negative test is the only kind that proves a control: it removes or violates the mechanism and shows the system refusing, rather than showing the happy path working.

Where the register overstates itself

Auditing the register against its own entries finds a defect that matters more than any single control. The 25 requirements marked verified cite between them eight distinct test names. Twenty-two of the data and infrastructure requirements point at two tests that exercise database role separation and request locking; four recovery requirements — encrypted backups, copies isolated from production, restoration tested against the deployment procedure, integrity verified on restore — are marked verified with no negative test named at all.

That is not a claim that the controls are absent. It is a claim that the record does not support the status it assigns them, and a register that grades an untested requirement as verified has destroyed the information a reader needs. The honest state is the one this page states: 25 requirements carry a test that exists, of which a subset carry a test that actually exercises them, and 115 carry none. Until the mapping is repaired requirement by requirement, the register's verified count is an upper bound, not a measurement.

The two tiers, applied

An enforceable property is one the system decides on every request and either holds or does not. The platform publishes eight such properties as measured claims, each settled by its own check — an unauthorised operation refused, an altered chain entry detected, the application role unable to write a balance or rewrite the record, a not public record not served. Those are in the evidence register with their grade and date, and they are the part of security that a reader can settle by inspection.

Everything else is an estimate: adversary capability, the likelihood that an unobserved attack class applies, the chance that a control fails in a way nobody anticipated, and the impact of a tail event. Estimates are recorded internally with their assumptions, the reasoning behind the ranking and the date the judgment is due to be revisited. They inform where money is spent. They are not published as findings, and no composite score, probability of compromise or certification appears on this site, because a reader cannot audit an estimate whose assumptions are not beside it.

What a control claim has to carry

A published control claim carries four things or it is marketing: the mechanism, the deployment it applies to, the test that fails without it, and the date. "The platform is secure" carries none of the four. "The application's database role cannot alter the evidence record, tested on 24 October 2026 against this release" carries all four, and a reader who distrusts it can ask for the test.

This is also why the register is public rather than summarised. A summary would let RJV choose which entries to show. The register lets a reader find the four recovery requirements marked verified without a test, which is the entry the register's author would least like to display.

Limits

Tier one can be complete and the system still insecure, because the properties made enforceable are the ones someone thought of; coverage is a decision, not a derived truth. The register lists what has been examined and not what exists. Estimates are only as good as the people holding them, and grading them does not improve them. And the boundary can be drawn dishonestly in either direction: a property placed among the estimates because testing it is inconvenient is an evasion, and an estimate promoted to enforceable without a mechanism is a claim dressed as a control.

The register itself, with the requirement text, the owning domain, the implementation files and every named negative test, is published in the repository that delivers this platform. The count above is measured from it, not asserted about it.

Read the security and assurance record