Evidence & limits

Every adjective must earn
its evidence class.

Filiolae reports what each bounded exercise established, retains negative results, and names the next gate before using stronger language.

01

Implemented & tested

Source and automated tests exist. This does not imply external administration or production operation.

02

Bounded acceptance

One named topology and precommitted procedure met its checks. This is the strongest current class for live integration.

03

Independent reproduction

A fresh party uses novel materials, separate administration, credentials, infrastructure, and evidence custody.

04

Production evidence

Operational reliability, incident response, recovery, monitoring, and adversarial administration are established in deployment.

An adult and child look across a landscape marked by distinct points of light
Evidence over distanceStronger claims require stronger custody and reproduction.

Accepted, bounded results

What the retained record supports.

Live integration

Pinned prime-rl two-A6000 campaign

Two happy-path promotions completed. A one-byte staged-artifact tamper caused durable denial, freeze, exit 75, and no later promotion.

One pinned version and topology; no general host or production claim.
Containment

Separate-UID native systemd game day

Distinct witness/orchestrator UIDs, protected key/store, systemd binding, and cgroup-v2 kill of a hostile detached descendant passed bounded scenarios.

Reboot persistence, GPU-device containment, and hostile root administration remain open.
Candidate evaluation

One-use real-model path

A frozen candidate passed its precommitted final threshold under distinct final custody and produced one Gate approval and one disposable shadow promotion.

The candidate and suites are consumed. One narrow task is not general model quality.
Retained failure

First frozen real candidate failed

The first campaign missed its precommitted threshold. Standard post-hoc replay durably denied and froze with zero promotion.

The candidate remains closed and cannot be tuned into positive evidence.

Current non-claims

What this preview does not establish.

  • Production security or deployment fitness
  • Independent reproduction
  • General model quality
  • Root or all-admin adversary resistance
  • Public transparency or trusted time
  • Key rotation and revocation
  • M-of-N promotion approval
  • Reboot and disaster-recovery readiness
  • Certification or endorsement

Canonical claim register

Historical reports never silently become stronger claims.

The repository's capability-and-gap matrix is the authoritative map from capability to evidence, residual gap, and next gate.

Open the claim register