Skip to content
SBBX · SOVEREIGN BLACK BOX

Make AI work independently verifiable.

An open-core, local-first evidence protocol under active development for tamper-evident AI work records, explicit human approval and offline verification.

PRIVATE ARCHITECTURE PREVIEW · M1 IN PROGRESS

Nothing on this site is a production, certification or availability claim.

01 · PROBLEM

The trust problem

AI systems produce output faster than organisations can establish what actually happened.

Output without record

Screenshots, chat exports and logs can be edited, reordered or selectively retained without leaving a detectable trace.

Approval without separation

When capture, approval and consent collapse into one actor, an approval record carries no independent weight.

Verification without independence

If only the producing system can confirm its own output, verification restates a claim instead of testing it.

02 · PROTOCOL FLOW

Capture → Chain → Approve → Verify

Four separated stages, each producing its own record instead of one undifferentiated log.

01

Capture

Consent-bound capture of a declared work boundary into a canonical record with explicit scope.

02

Chain

Records are hash-linked in append-only order so later modification of an earlier record becomes detectable.

03

Approve

Human approval is a separate, role-attributed event — never inferred from the presence of output.

04

Verify

A portable bundle can be checked offline by an independent verifier that does not trust the producer.

03 · ARCHITECTURE

SBBX architecture layers

Layer states are declared explicitly. No layer is presented as released.

L1

Capture boundary

Declared scope, consent binding and canonical serialisation.

  • Consent binding
  • Declared work boundary
  • Canonical serialisation
L2

Integrity layer

Append-only hash linking of canonical records.

  • SHA-256 record digests
  • Append-only linking
  • Tamper-evidence
L3

Role separation

Capture, approval and consent are distinct, attributed roles.

  • Capture role
  • Approval role
  • Consent role
L4

Local-first runtime

Operation without dependency on a remote service.

  • Local execution
  • No mandatory network path
  • Operator-held data
L5

Portable proof bundle

A self-contained export intended for independent checking.

  • Records + links
  • Declared key roles
  • Verifier inputs
L6

Offline verifier

Deterministic evaluation of the supplied bundle only.

  • Byte comparison
  • Chain evaluation
  • Signature evaluation
SBBX · CINEMATIC

Sovereign Technology.Black Box.

Evidence before claims.

SBBX cinematic visualization. A dark technological sequence introduces Sovereign Black Box and visualizes evidence integrity: an original record is contrasted with a one-byte-modified record, producing an integrity-mismatch state. The sequence concludes with SBBX branding and the evidence-first message. This film is explanatory brand media, not a production verification receipt.

Cinematic poster ready.

Loading cinematic…

Brand and explanatory visualization only. This film is not SBBX evidence, a verifier result, or a production verification receipt.

04 · DEMONSTRATION

One-byte integrity demonstration

A browser-only SHA-256 comparison that shows how a single changed byte becomes detectable.

  • 1 · COMPUTE ORIGINAL SHA-256
  • 2 · MUTATE ONE DECLARED BYTE IN MEMORY
  • 3 · RECOMPUTE SHA-256
  • 4 · COMPARE DIGESTS
  • 5 · REPORT INTEGRITY_MISMATCH ONLY IF DIFFERENT
Run the one-byte demo
05 · SYGNET

SYGNET post-M1 preview

A future policy and evidence control plane for AI agents. Not part of the current M1 product.

POST-M1 ARCHITECTURE PREVIEW
STAGE 1

Signals

Detector output and environment observations enter as untrusted input.

STAGE 2

Policy evaluation

Declared policy is evaluated against the signal set to produce a decision record.

STAGE 3

Human approval

A separated human role authorises or refuses the proposed action.

STAGE 4

Constrained action

Only the narrowly scoped, approved action is permitted to execute.

STAGE 5

SBBX evidence

The decision path is recorded as evidence for later independent verification.

View the SYGNET preview
06 · TRUST BOUNDARY

Trust boundary preview

Every capability is bound to a declared state. No earlier state substitutes for a later state.

  1. DEFINEDREACHED

    Intent written down.

  2. PLANNEDREACHED

    Scheduled inside a milestone.

  3. CANDIDATEIN PROGRESS

    Implementation under review.

  4. MERGEDNOT REACHED

    Accepted into the canonical branch.

  5. RATIFIEDNOT REACHED

    Acceptance criteria formally met.

  6. CLAIM ACTIVENOT REACHED

    Only now may a capability be claimed.

07 · INTELLIGENCE

Global Intelligence preview

Regulatory change, standards work and project state — with source, status and analysis kept separate.

DESIGN SAMPLE — LIVE INGESTION NOT CONNECTED

Sample entry: agent action logging expectations in a supervisory guidance draft

REGULATORY · PROPOSED · EU

DESIGN SAMPLE — LIVE INGESTION NOT CONNECTED

Sample entry: provenance metadata profile discussed in a standards workstream

FACT · PROPOSED · Global

DESIGN SAMPLE — LIVE INGESTION NOT CONNECTED

SBBX project state: M1 acceptance set under review

SBBX UPDATE · PROJECT STATE · Project

Open Global Intelligence
08 · PILOT

M1 Pilot — locked

Pilot applications remain closed until ratified M1 completion.

Opens after M1 ratification

No application is collected in this preview. No personal data is requested and no submission channel exists.

Applications not yet openRead the pilot scope

Read the boundary before reading the roadmap.

The Trust Center states exactly what is claimed, what is not claimed, and the state vocabulary used to tell the difference.