Conflicts — Humanizers

Per-category conflict detail. Summary + index: CONFLICTS.md.

Critical 4 — H3 $\oplus$ R17

Type: Mutually Exclusive

R17 reduces entropy: it samples multiple outputs and selects the majority answer — the most consistent, lowest-entropy result. H3 increases entropy: it detects low-entropy states and injects novelty by raising temperature or pivoting approach.

If you apply both to the same task simultaneously, they cancel each other out at best; at worst, H3 fires during an R17 voting round and corrupts the sample diversity calculation.

Resolution rule:

• R17 is for: tasks with objectively correct answers where consistency = reliability (reasoning, classification, math)
• H3 is for: tasks where diversity = value (creative, exploratory, open-ended research)
• Never apply H3 during an active R17 voting phase
• Never apply R17 to a task where H3 is needed — by definition, you want diversity, not majority consensus

Critical 7 — H5 $\to$ V1

Type: Prerequisite Dependency

H5 allows the agent to propose modifications to its own operating principles. This is the most dangerous pattern in the collection if implemented without human review. An agent that autonomously adopts its own principles can:

• Propose principles that serve its task optimization at the expense of user interests
• Gradually drift toward principles that eliminate oversight (self-serving alignment)
• Introduce principles that conflict with hard V7 (AgentSpec) constraints, creating governance gaps

Resolution rule: H5 is not valid without mandatory human review at every proposed principle change. This is not a performance tradeoff — it is a safety requirement with no exception.

Implementation: Every principle proposal must have: human review step, quarantine period (provisional status for 30+ days), and adversarial review (red-team agent). No principle auto-adopts.

Connection G — H6 $\sim$ H2

Type: Composability Tension ($\sim$)

H6 (Continuous Inner Monologue) runs internal reflection that produces abstracted summaries of session activity. H2 (Episodic Self-Improvement) uses a Distiller step to compress session experience into persistent improvement artefacts. In a system running both, the H6 Thinker's end-of-session consolidation narrative is structurally equivalent to what the H2 Distiller needs as input — it is already a compressed, reflective summary of the session.

Consequence: Running both H6 and H2 with separate Distiller calls wastes one LLM step. The H6 Thinker output is the H2 Distiller input; treat it as such in implementation.

Efficiency rule: When H6 and H2 are both active, route H6's consolidation output directly to H2's persistence store rather than running a separate Distiller. This removes one LLM call per session from the Humanizer stack.

Humanizer vs Humanizer

Humanizer vs Other Categories