Integrity
Have the protected bytes changed relative to a trusted receipt or checkpoint?
TRUST MODEL
Evidence you can inspect.
Claims you can delimit.
Understand what a record can demonstrate, what it depends on and what remains outside its scope.
Provenance
Which recorded source and context references are associated with this event?
Reconstruction
What stored memory state can be recovered from the retained history?
THE BOUNDARIES
What the mechanism establishes.
| Mechanism | What it supports | What it does not establish |
|---|---|---|
| Cryptographic receipts | Checking protected data against a trusted reference. | That the original source is factually true. |
| Append-only history | Preserving earlier events as new ones are added. | Automatic protection from every data-loss scenario. |
| Tenant separation | Access bounded by authenticated workspace identity. | Isolation based only on a tenant field in a payload. |
| Deterministic replay | Reconstruction of retained memory state. | Identical outputs from a probabilistic language model. |
| Non-generative retrieval | Returning stored evidence and relationships. | Preventing every hallucination in a connected model. |
ACCESS & SECURITY
Identity at the boundary.
Evidence within it.
The access model uses centralized authentication, link-based email verification and JWT-based sessions. MCP and gRPC boundaries are protected, with authorization designed to fail closed.
Authenticated scope
Tenant identity is resolved from authenticated backend context.
Fail-closed access
Missing or invalid authorization must not become an empty success response.
Inspect the result
Use authoritative record identifiers to follow an event back to its evidence.
CLEAR ANSWERS
A trust model should answer the hard questions.
Does PHIROK prevent hallucinations?
PHIROK provides verifiable memory and traceable evidence. It is not advertised as a mathematical guarantee against hallucinations in a connected model.
Is this a regulatory certification?
No. Evidence and audit workflows can support review, but they do not establish EU AI Act compliance or any other certification by themselves. No certification is claimed on this site.
Are the interactive examples live production records?
No. They are clearly labeled samples. The integrity playground computes hashes in your browser using SHA-256 for its demonstration. It is independent of PHIROK’s production BLAKE3 receipt implementation.
What if the source was wrong?
An integrity check can confirm that retained bytes match their reference. It cannot turn inaccurate source material into truth. Source quality and evidence integrity are separate questions.
Does deterministic replay reproduce an AI’s reasoning?
Replay concerns retained memory state and recorded context. It does not claim access to an AI’s private internal reasoning or guarantee that a model will generate the same response again.
Which claims are deliberately excluded?
Live contradiction detection, commercial production embeddings, mathematical hallucination prevention and regulatory certification are not presented as available capabilities. PHIROK is not marketed as an answer generator.
YOUR NEXT DECISION
Inspect the evidence.
Settle the question.
A trust model is only useful when it gets used. Take it to the playground and try to break a sample record.