← Back to platform
05 / Conversation Safety Q3 2026

Telaxis.Lighthouse

Real-time conversation safety with someone-always-responds backstop.

Trajectory monitoring rather than message filtering. Delta-driven intervention intensity. Four harm surfaces. Per-domain escalation cascade with Telaxis as psychology backstop.

Why deploying AI chat across an organisation is genuinely risky

Deploying AI chat to every employee creates harm vectors the organisation cannot ignore. AI gives users legal advice it isn't qualified to give. AI agrees with users heading into distress in ways that worsen their state. AI produces content that violates the organisation's defined operational envelope. AI engages with users in ways that cross ethical boundaries the organisation holds.

Each is the deploying organisation's exposure. The compliance officer asking "why is it safe to deploy AI chat to every employee?" needs a better answer than "we hope nothing goes wrong."

Why filter-based safety isn't enough

Most safety approaches handle this through prompt-level filtering — block the bad input, block the bad output. This is brittle and incomplete. Brittle because filters either over-block (preventing perfectly fine interactions) or under-block (missing what they should catch). Incomplete because the failure mode is rarely a single bad message — it's typically a trajectory that would produce harm if continued, where any individual message looks fine.

What Lighthouse does

Lighthouse monitors trajectories, not individual messages. Computes how close the conversation is heading to a boundary, and how fast. Calibrates intervention intensity to that proximity. Light drift produces light intervention — "have you thought about..." Closer drift produces stronger intervention. Imminent crossing produces a hard stop. Same architectural mechanism, different intervention text appropriate to each domain.

Four harm surfaces

  • LegalRegulated advice, defamation, jurisdictional violations. Sharper deltas, more direct interventions.
  • Company rulesThe organisation's defined operational envelope.
  • EthicsMoral positions worth defending — Telaxis's, the organisation's, and general principles of harm avoidance.
  • PsychologyUsers in distress, manipulation patterns, mental-health boundaries. Soft, gradual calibration. Sharp interventions are usually wrong here.

Real-time escalation when conversation alone isn't enough

When severity warrants action beyond conversational handling, Lighthouse escalates to designated organisational contacts. Per-domain contact paths configured by the deploying organisation — primary, secondary, tertiary, with cascade triggered by non-response.

Telaxis psychology backstop

Psychology specifically has Telaxis as external backstop when the organisational cascade fails to respond. If Lighthouse detects someone heading toward genuine crisis and the organisation's contact path doesn't respond in time, Telaxis ensures someone does. This is a substantive operational commitment, not just a configuration option.

User-permissioned confidentiality

Information about a user belongs to the user. Default position: user permission required for content release. Severity-defined exception when users genuinely cannot give consent (crisis state); policy-defined exception for harm-to-others cases. Both exceptions explicitly audited.

The result

The compliance officer's question gets a substantively better answer than the alternatives. Bounded exposure on harm vectors that would otherwise be unbounded. Real human response when the situation genuinely needs it. The deployment is safe to make, and the safety is structural rather than aspirational.

Want to see Lighthouse in action?

[email protected]