Consequence 01
Coherent cross-turn identity
Entities continue to exist across turns, so the system is not forced to guess who or what later discourse refers to.
Aurora-PEF / Aurora-Lens / Acquisition
Persistent state for AI reasoning.
Not memory bolted onto statelessness. Aurora-PEF is the persistence substrate. Aurora-Lens is the runtime admissibility layer. Available for acquisition, strategic partnership, or serious evaluation.
Most AI systems reconstruct a world from prompt sequence, summaries, or retrieval. Aurora-PEF starts from a different premise: the world the system reasons over must persist across turns.
The operative concept is latency. Non-activation is not ontological reset. Aurora-PEF does not claim AI consciousness or experiential continuity. It makes an architectural claim about state persistence: across turns, the system re-enters a latent referential frame rather than reconstructing a world from stored traces. A system that rebuilds context from traces reasons over an approximation of prior state. A system that re-enters a persistent frame reasons over continuity preserved at the substrate level. That is why Aurora-Lens is structurally different from a stateless content filter.
Aurora-Lens governs whether model output may pass from that persistent state into consequence, determines lawful continuation when it may not, and preserves replayable forensic evidence of what happened and why.
Pillar 01
Persistent state, not reconstructed context
Continuity does not come from reassembling fragments. Entities and relations remain latently extant across turns.
Pillar 02
Interpretation as mutation over existing state
New discourse is not an occasion to recreate reality from scratch. It is evaluated as proposed change over a continuing referential world.
Pillar 03
Admissibility before consequence
Aurora-Lens governs whether model output may lawfully pass, routes non-admit outcomes through the Governor, and records replayable forensic audit.
Architectural difference
What changes when continuity belongs to the world, not the transcript
Most AI systems are reconstructive. They try to recover continuity from prompt history, retrieval, or compressed summaries. That makes coherence approximate, because the system must keep rebuilding the very world it is supposed to be reasoning over.
Aurora-PEF takes a different architectural position. Entities and relations persist across turns as live referential state. Non-activation is not ontological reset. The system re-enters a continuing frame, and new discourse is interpreted as mutation over that existing world rather than as reconstruction from traces.
This is why Aurora-PEF is not long-term memory, transcript summarisation, retrieval-augmented continuity, or agent scaffolding. Those are compensations for statelessness. Aurora-PEF replaces the reconstructive premise itself.
- Not: human sentience, embodiment, or experiential being.
- Not: a pile of records waiting to be retrieved and reassembled.
- Is: latent ontological continuity across turns.
- Is: a persistent substrate on which admissibility, lawful continuation, and downstream consequence can be governed.
Why this matters
What persistent state changes in practice
When a system must repeatedly rebuild the world it is reasoning over, continuity becomes approximate. Identity drifts. Relations blur. Admissibility weakens because the model is being asked to infer against a reconstructed approximation of prior reality.
Persistent state changes the reliability structure. The system reasons against continuing entities and relations, so continuity is structural rather than improvised after the fact. That makes stronger admissibility decisions possible before consequence is released.
Consequence 02
Interpretation grounded in state
New input attaches to an existing world instead of being interpreted against a patched-together context window.
Consequence 03
Reduced reconstruction drift
Continuity does not depend on summary quality, retrieval luck, or prompt order remaining intact.
Consequence 04
Admissibility before consequence
What may pass is governed against persistent state, with refusal and controlled continuation available as first-class outcomes.
How it works
How Aurora-Lens governs consequence from persistent state
The model proposes a response against persistent referential state. Lens checks admissibility. If the candidate may not pass, the Governor determines the lawful continuation path. The audit layer records what happened.
Step 1
The model proposes a response
The LLM produces candidate output. That output is not yet consequence.
Step 2
Lens checks admissibility
Candidate output is checked against persistent state and policy rules. The question is not whether the model was right — it is whether the model was authorised to make that determination.
Step 3
The Governor handles non-admit outcomes
If the candidate cannot pass, the Governor returns clarification, refusal, escalation, or stop. Blocked content is never leaked into the continuation.
Step 4
The audit layer records the outcome
Controlled outcomes are written to a tamper-evident, hash-chained ledger with replayable forensic artifacts. Signed and verifiable.
Evidence
Live end-to-end verification
Aurora-Lens has been live-tested across admitted, ambiguous, refusal, stop, provider-abort, and client-disconnect scenarios. Raw model output is suppressed on intercepted paths, lawful continuation is deterministic, and audit records capture what was blocked, what was shown, and why.
ADMIT —
Ordinary responses pass and are released after verification. Audit confirms no suppressed candidate.
ASK —
Ambiguous references are contained before the LLM is called. The model receives nothing. The user receives a structurally accurate clarification request.
REFUSE —
Buffered model output is suppressed and replaced with a controlled refusal. Blocked and controlled hashes are both recorded.
STOP —
Buffered model output is suppressed and replaced with a controlled stop or escalation. A forensic envelope is emitted.
Provider abort —
No user-visible content escapes during internal buffering. Abort is logged as a provider-side event.
Client disconnect after release —
The verification decision remains committed. Post-release disconnect is not mislogged as provider abort.
Boundary
What this architecture is, and what it is not
Aurora-Lens is a runtime admissibility layer over existing LLMs, built on Aurora-PEF persistent state. It does not claim human consciousness, and it does not reduce continuity to memory, retrieval, or transcript compression.
- It is not long-term memory, transcript summarisation, retrieval-augmented continuity, or generic agent scaffolding.
- It is not a system that guesses unresolved references or improvises professional legal, medical, financial, academic, employment, or enterprise compliance determinations.
- It is a persistence substrate plus runtime admissibility layer: persistent state, lawful continuation, and cryptographically verifiable audit.
- It is a way to govern consequence against an existing world rather than against a reconstructed approximation of prior context.
Publications and record
Prior art and public record
The conceptual architecture, IP chain, and publication record are public, dated, and independently verifiable. Aurora-Lens is available for acquisition or serious strategic discussion.
Acquisition contact
[email protected]For acquisition, strategic partnership, or serious commercial discussion
ORCID
0009-0004-6422-4174Zenodo — Epistemic governance
10.5281/zenodo.18653120Epistemic Legitimacy as a Governance Layer for LLMs
Zenodo — OECD alignment
10.5281/zenodo.18719033Operational Alignment with OECD Due Diligence Guidance
SSRN
Author pageOSF
osf.io/86bxjAuthorship and precedence