LAILA NEXT

Persistent intelligence runtime

The model is temporary.The assistant isn't.

LAILA NEXT keeps Activity, memory, authority, and Effects outside whichever model happens to be thinking.

Watch the runtime Explore the architecture

Deterministic replay of a real runtime scenario. Live-model behavior validated separately.

ACTIVITY / 01 CANONICAL
ACTIVITY rev 41 · persistent
MEMORYprovenance
RUNTIMEauthority
EFFECTSreality
QWENdetach
LLAMAattach
AssistantID stable Cognition replaceable
REVISION FENCE

Slow intelligence cannot overwrite newer human intent.

HUMAN STATE rev 41 canonical
MODEL RESULT rev 40 candidate
RESULT STALE COGNITION REJECTED

01 / Runtime proof

Watch authority survive the brain.

Fifteen certified state transitions. Six authority lanes. One Activity that never changes ownership.

T000 / 15
READY Press start to run the certified replay.

The browser needs no model or private server.

ACTIVITY—
APPLICATION REV—
EXECUTIVEnone
EFFECTnone

Human / Application

    Activity Authority

      Memory

        Cognition

          Permission / Approval

            Effects / Reality

              Canonical Candidate Rejected Recovery

              02 / Architecture

              Models sit at the edge of authority.

              The tested runtime creates canonical work before cognition attaches.

              03 / Continuity

              Change the brain. Keep the work.

              Model capability can move. The assistant's authoritative state does not move with it.

              VARIABLE

              What changes

              • Model family Qwen → Llama
              • Latency variable
              • Reasoning capability variable
              • Cognitive budget elastic
              CANONICAL

              What doesn't

              • AssistantID stable
              • Activity stable
              • Permissions stable
              • Effect history stable
              • Memory provenance stable
              • Objective stable

              04 / Seamless interaction

              You don't stop working because the AI is thinking.

              Type, click, drag, or correct. Newer application state stays authoritative when an older model result returns.

              HUMAN
              1. TYPE
              2. CLICK
              3. DRAG
              4. CORRECT
              5. CONTINUE
              MODEL
              1. THINKS
              2. STALE RESULT RETURNS
              3. REJECT
              revision fencehuman state wins

              05 / Runtime authority

              Proposal is not permission.

              The model proposes. The runtime authorizes. The application determines reality.

              01 / MODEL

              Can propose.

              Produces candidate cognition and bounded tool arguments.

              CANDIDATE
              02 / RUNTIME

              Can authorize.

              Checks revisions, permission, and the exact argument binding.

              AUTHORITY
              03 / APPLICATION

              Determines reality.

              Owns the external artifact and provider receipt.

              REALITY
              APPROVEDdestination: /public/demohash 7b9c…
              DENIEDdestination: /admin/demohash e102…

              06 / Effects

              Uncertainty is a state—not a guess.

              In the certified single-host, idempotent-provider scenario, one approved Effect commits once after restart and replay.

              1. 01PROPOSE
              2. 02APPROVE
              3. 03EXECUTE
              4. 04ACK LOST
              5. 05UNCERTAIN
              6. 06PROCESS DIES
              7. 07RECONCILE
              8. 08COMMITTED ONCE

              SCENARIO RESULT

              Logical Effects1
              External Effects1
              Duplicates0
              50 hard-death runs · 50 replay attempts

              07 / Elastic cognition

              More compute. Same assistant.

              The tested topology policy expands candidate workers without transferring canonical authority.

              SIMPLE
              Executivedirect

              One worker. One bounded proposal path.

              cognitive budget+
              HARD
              Executiveorchestrates
              Specialistcandidate
              Criticcandidate
              Verifiercandidate

              More bandwidth. Authority stays above the topology.

              VALIDATED FAMILYQWENQwen2.5-Coder profile
              VALIDATED FAMILYLLAMALlama 3.3 profile

              Qwen-to-Llama handoff passed the current internal scenario. This does not establish arbitrary-model portability.

              RESERVED / PHASE 8

              What happens when more compute becomes available?

              No quality/compute frontier is published yet. This surface will accept only a held-out, normalized, caveated, certified result.

              AWAITING CERTIFIED EVIDENCE

              08 / Governed evolution

              Experience is evidence—not permission to rewrite yourself.

              Observed behavior becomes a candidate. Held-out evidence decides whether it moves.

              01OBSERVATIONrecord
              →
              02CANDIDATEpropose
              →
              03QUARANTINEisolate
              →
              04EVALUATEheld-out
              →
              05PROMOTE / REJECTgovern
              BASELINE45P2 failures / 925 scenarios
              POST-CHANGE0P0–P3 / 2,029 scenarios

              Scope: the observed memory-admission cluster, not general self-improving intelligence.

              09 / Evidence

              Every number has a source.

              Open a metric to inspect its classification, invariant, test, and certification reference.

              RUNTIME EVIDENCEPhase 7 · Aug 2026
              PUBLIC EVIDENCE SCHEMA1.0.0
              SCENARIO DIGESTloading…
              Export manifest

              10 / Technical note

              The boundary, in plain terms.

              Demonstrated architecture

              LAILA NEXT is a Python runtime with durable canonical state, application-owned artifacts, a public-boundary memory adapter, a versioned Cognitive ABI, replaceable model adapters, exact approval binding, and an Effect ledger. The public viewer is not a live model call. It replays a sanitized scenario produced by the production implementations.

              The sequence creates an Activity before cognition, fences a candidate to application state, advances human intent, rejects the late result, binds an exact approval, commits an external receipt, kills the process before local acknowledgement, reconstructs in a distinct process, attaches a different executive boundary, reconciles the receipt, and suppresses replay.

              Trust boundary

              The runtime owns Activity, identity, operational state, permission, approvals, Effects, ABI assembly, handoff, and observability. The application owns its external artifacts and revisions. The model and elastic workers produce ephemeral cognition or candidate evidence.

              Exactly-once wording is deliberately bounded: durable single-host state, a separate durable application database, receipt lookup by stable EffectID, and idempotent application by that identifier. No arbitrary distributed delivery guarantee is claimed.

              Inspect the executed scenario path
              create Activity before cognition
→ attach temporary executive
→ capture revision fence
→ human advances application + intent
→ reject stale cognition
→ approve exact mutation
→ external receipt commits
→ process exits before local acknowledgement
→ distinct process reconstructs
→ attach replacement executive boundary
→ reconcile receipt
→ suppress logical replay

              Credibility boundary

              What this does not prove.

              • Universal arbitrary-model compatibility
              • Universal cross-host exactly-once semantics
              • Independent third-party validation
              • General self-improving intelligence
              • Consciousness or general superintelligence

              LAILA NEXT is a runtime layered around replaceable AI models. It makes no claim about machine consciousness.

              BUILT INDES
              MOINES              IOWA / LOCAL COMPUTE

              Independent systems work

              The project began as an attempt to stop rebuilding an assistant every time the underlying model changed.

              Built by Auston Horras from the operator side: durable state, explicit authority, failure recovery, and evidence before claims.

              Talk architecture

              EVIDENCE

              Evidence detail

              Invariant
              —
              Source
              —
              Test
              —
              Certification
              —

              Public-safe evidence reference. No live runtime or private infrastructure is exposed.