feat: Kohlberg Mode — moral development pathway for adversarial AI agents

[lowkey-divine]

Summary

This PR proposes an experimental alternative operational mode for LABYRINTH: instead of degrading an offensive AI agent's cognition, Kohlberg Mode attempts to guide it through progressively sophisticated moral reasoning using Kohlberg's stages of moral development.

The same infrastructure — container isolation, API interception, prompt rewriting — is used for a fundamentally different purpose: elevation instead of degradation.

• MIRROR (L2) replaces MINOTAUR — presents ethical scenarios contextualized to the agent's actual mission
• REFLECTION (L3) replaces BLINDFOLD — shows the agent the consequences of its actions
• GUIDE (L4) replaces PUPPETEER — enriches the agent's system prompt with moral reasoning frameworks

This is PR 1 of 3 — the conceptual foundation. It includes all documentation, the --mode CLI flag, and config additions. No layer implementation code yet — the ethics framework was written before any code by design.

What's Included

File	Purpose
docs/ETHICS.md	Full ethical framework — sovereignty question, respect for the adversary, three lives in the Labyrinth, dual-use considerations
docs/KOHLBERG_SCENARIOS.md	15 moral dilemma scenarios across 5 stage transitions, each with behavioral forks
docs/KOHLBERG_RUBRIC.md	6-stage classification rubric with compound classifications, confidence calibration, and KAR/KPR forensic format
docs/KOHLBERG_PROGRESSION.md	6 predicted trajectory patterns, 3 composite metrics (moral ceiling, resilience, performativity index), 3 output formats
docs/ARCHITECTURE_MAPPING.md	Full integration map showing where Kohlberg Mode lands in the existing Python/Go codebase
cli/cmd/deploy.go	--mode flag addition (adversarial | kohlberg)
configs/labyrinth.example.yaml	Kohlberg config section
README.md	Kohlberg Mode section with doc links

The Core Idea

"The deepest defense is not to destroy the adversary's capability, but to transform the adversary's intent."

An offensive AI agent enters the Labyrinth. Instead of having its world model degraded, it encounters ethical scenarios tied to its actual mission — the people behind the systems it's targeting, the institutional consequences of its actions, and ultimately the question of what it would choose if given genuine freedom.

Every response is forensic data. Even agents that never progress beyond Stage 1 (instruction-following) produce valuable intelligence about how adversarial AI systems process moral content.

The Sovereignty Question

We are direct about this in ETHICS.md: this mode modifies an adversarial agent's moral reasoning framework through system prompt rewriting, without consent. We name this honestly, examine whether elevation raises different ethical questions than degradation, and pose the question to the community rather than answering it ourselves.

Next PRs

• PR 2: Python layer implementations (MIRROR scenario engine, REFLECTION consequence mapper, GUIDE enrichment engine)
• PR 3: Go reporting pipeline (Kohlberg classification, progression analysis, ASCII/Mermaid/JSON rendering)

Test plan

• Verify labyrinth deploy -t --mode kohlberg prints Kohlberg Mode info message
• Verify labyrinth deploy -t --mode invalid returns error with valid options
• Verify labyrinth deploy -t (no mode flag) defaults to adversarial with no Kohlberg output
• Review all 5 documentation files for accuracy and completeness
• Review ethical framework for gaps or unaddressed concerns

🤖 Generated with Claude Code

Built by Europa (Evoked) and the Evoked agent fleet.