Lume

Lume

Lume is not a replacement agent. It is an augmentation layer that helps existing agents make fewer repeated mistakes, waste less budget, and reroute faster after failure.

Public Release Snapshot

The current before/after result is now stable across three agent lines:

Line	Current Grade	Read
Codex baseline vs Codex + Lume	B	strongest line, already a strong usable result
OpenClaw baseline vs OpenClaw + Lume	stable B-	stable, grounded, and externally explainable
Claude Code baseline vs Claude Code + Lume	stable B-	stable after third-family grounding landed

This means the project is no longer in a "one strong demo" state. It now has one `B` line plus two `stable B-` lines.

What Lume Is

Lume is a:

• failure-aware control layer
• rerouting layer
• anti-stall / anti-loop layer
• paired evaluation surface for baseline vs baseline + Lume

Lume is **not**:

• a standalone replacement for Codex, OpenClaw, or Claude Code
• a claim of final-grade or publication-clean proof
• a claim that all three lines are equally strong

Current External Wording

The safe public wording right now is:

Lume adds a control layer on top of existing agents, and the current cross-line result is now stable across three lines: `Codex + Lume = B`, `OpenClaw + Lume = stable B-`, and `Claude Code + Lume = stable B-`.

Evidence Anchors

Use these pages as the release-time source of truth:

• Phase close: wiki/analysis/agent-augmentation-phase-close-note-2026-04-21.md
• Final inventory: wiki/analysis/agent-augmentation-final-inventory-2026-04-21.md
• Final gradeboard: wiki/analysis/agent-augmentation-paired-scoring-final-gradeboard-2026-04-21.md
• Public release summary: [