Evidence before assurance. Status, limitations, and review dates stay attached to the claim.
LUXION does not ask institutions to trust abstract AI-safety claims. It produces operational evidence surfaces within controlled deployment pathways — with explicit status, evidence level, limitations, and review cadence.
What the Evidence Supports. A claim earns credibility through a clear method, inspectable evidence and explicit limitations. We distinguish implemented mechanisms, bounded evaluations, active research and open questions.
Records that exist are linked from this page and the evidence log. Five experimental protocols are proposed and unpublished. They are not results. Evidence is open to challenge. A correction request should name the public page or record and should not include private data. How to challenge a finding
Related mechanism diagram (conceptual, not a replacement for the Evidence Window): Gate versus host enforcement.
Public research is organized so visitors can see what is open, what is reviewed and attributed, and what remains controlled.
Layer A and Layer B content may appear on public research and evidence pages. Layer C is not bundled here. Research disclosure model · Selected work · Research method
Where measurements are available, utility, interventions, latency, computational demand, and relevant resource costs should be read together. A more efficient workflow is not automatically a safer or more sustainable one. No energy, emissions, or lifecycle figure is published here, because those measurements are not in the public record. Conclusions must preserve experimental conditions and uncertainty.
Action → Evidence → Route → Audit → Scope — evidence and scope are bounded at the audit threshold, before operational consequence.
Company identity does not grant technical authority to a product or scientific certainty to a programme. This page remains the claims constitution. A material claim is read against its evidence, maturity, and stated limits. This page does not add review dates, benchmarks, customers, or certifications.
Every public research or product claim should carry: status, evidence level, limitations, last reviewed, and — where appropriate — the next discriminating test.
Evidence & Scope is the company constitution for claim boundaries. Research programme pages and the evidence log inherit provenance fields (FIRST_PUBLISHED / LAST_REVIEWED / CURRENT_STATUS / EVIDENCE_LEVEL / LIMITATIONS). Absence of a field is information — not implied maturity.
Every public claim should sit on an explicit rung. Absence of a rung is information. Do not assign a level without evidence.
Current public Guardian harness package maps to E3. Guardian Pilot public package: AgentDojo-compatible budget validation; official_comparable=false; not production certification. Higher rungs are not claimed.
Reusable claim-bounded cards for the current public posture. Dates on these cards are evidence-log review dates, not a page-level approval and not a website publication date.
Public quantitative Guardian claims remain bounded to the AgentDojo-compatible budget validation package on Guardian Pilot.
budget_nonleaderboard; official_comparable=false; not production certification. E4–E7 not claimed.
LUXION publishes an explicit evidence ladder so claims can be located by rung. Absence of a rung is information.
Ladder labels are taxonomy, not evidence that higher rungs have been reached.
LUXION produces evidence surfaces around proposed AI actions — within controlled deployment pathways, not through certification or abstract safety promises.
Operational evidence surfaces include proposed action, applied constraints, evidence state, authority boundary, policy route, decision, escalation path, and replayable audit record.
What the agent or workflow intends to execute before commit.
Policy, risk, and governance limits evaluated against the proposal.
Whether supporting context and provenance are sufficient to proceed.
Whether decision rights match the proposed action in this context.
How institutional policy maps to admissibility at runtime.
Allow, delay, route, block, or scaffold — recorded before consequence.
Human or institutional review when evidence or authority is insufficient.
Structured trace for incident review, near-miss analysis, and control updates.
LUXION enters operational use through a deliberate sequence — architecture fit first, observation before enforcement, evidence review before domain expansion.
Bounded fit conversation — integration surfaces, action inventory, policy boundaries, evidence maturity, and deployment design.
Architecture & deployment inquiryRecord proposed actions, constraints, risk signals, and decisions without interrupting production workflows.
Compare runtime decisions against live traffic to establish baseline fit before any enforcement.
Graduated allow, block, and escalate with agreed policies, escalation paths, and rollback.
Assess false-positive and false-negative patterns, latency impact, audit completeness, and recommended control updates.
Extend governed workflows and sector context as evidence matures — not instant enterprise rollout or certification.
Pathway outputs inform fit assessment and deployment design — they do not constitute production clearance, safety certification, or regulatory approval.
An architecture review maps where AI-generated intent can become operational consequence — across AI action surfaces, tool calls, data movements, memory writes, workflow effects, authority boundaries, evidence requirements, escalation paths, and sector constraints.
Each review produces structured artifacts shaped around real workflows and institutional constraints — for replay and runtime control, not standalone advisory reports or certification.
Structured outputs from architecture review and shadow-mode scoping — for replay and runtime control, not certification:
During shadow-mode pilots and controlled enforcement, institutions can review observable patterns — not universal prevention or guaranteed mitigation.
Proposals that lack sufficient evidence or context before execution.
Tool-use attempts evaluated at the runtime gate — not claimed as universal attack prevention.
Uncertain actions directed to human or institutional review paths.
Data access and memory operations flagged before commit.
Gaps between documented policy and runtime admissibility behavior.
Where proposals proceed without sufficient provenance or context.
Near-miss and incomplete-action patterns captured in audit envelopes.
Whether routed cases carry structured context for accountable review.
Where constraints, violations, and interventions recur across workflows.
Evaluation surfaces describe what can be observed and reviewed during controlled deployment — not guaranteed detection, elimination of risk, or production certification.
LUXION does not eliminate AI risk, certify compliance, replace legal, clinical, cybersecurity, financial, or compliance review, or guarantee safe AI behavior. It provides execution-control infrastructure and evidence surfaces for controlled deployment pathways.
LUXION can support institutional risk, governance, evidence, and accountability workflows — but it does not certify institutional compliance.
It is runtime infrastructure designed to support accountable AI deployment: execution-time decisions, evidence sufficiency, escalation paths, action provenance, and corrective signals where AI systems propose action.
Framework references are used for institutional orientation only. Guardian does not certify legal, regulatory, or standards compliance.
| Governance concern | Guardian technical surface |
|---|---|
| Risk oversight | Action risk signals, constraints, escalation thresholds |
| Decision rights | Policy-boundary and authority checks before execution |
| Semantic layer | Shared action context, evidence requirements, and constraint vocabulary at runtime |
| Auditability | Governed execution records, replay inputs, decision traces |
| Human oversight | Routed review for uncertain or high-risk actions |
| Corrective action | Incident review and control updates |
Institutional orientation describes where Guardian can support institutional workflows during bounded evaluation and controlled deployment pathways—not where it replaces counsel, compliance sign-off, or sector-specific approval.
Guardian Runtime is the current technology, a technical preview for evaluation before consequential AI action. Scope and enforcement depend on the configured workflow and host integration. It is not a certification.
Claims remain bounded by evidence tier, benchmark protocol, deployment context, and reproducibility status. AgentDojo-compatible budget validation is an internal and pilot-facing discipline—not an official AgentDojo leaderboard submission and not a substitute for production certification.
Detailed product mechanics, control surfaces, and bounded empirical evidence materials live on guardianpilot.org for product evaluation.
Explore platform product lines or start a deployment conversation — scoped to evidence maturity and controlled pathways.
Or email [email protected]
