F033: Censor Layer - Proactive Failure Pattern Warnings

Status: Proposed Source: Minsky Society of Mind Ch 27 (Censors and Jokes) Depends on: Sufficient failure data (~10+ failed/partial decisions) Priority: Low (blocked by data) Note: Previously numbered F027; renumbered to avoid conflict with shipped F027 (Decision Quality)

Concept

Current guardrails are suppressors - they block actions at the moment of execution ("high stakes + low confidence = blocked"). Minsky distinguishes these from censors, which intercept before the bad thought forms.

A censor layer would sit between query and guardrail check:

query -> CENSOR (warns based on failure patterns) -> check -> record

Suppressors (what we have)

• Evaluate at action time
• Block based on static rules (stakes/confidence thresholds)
• Reactive - the agent already formulated the bad idea

Censors (what we'd add)

• Evaluate at query time
• Warn based on patterns from failed decisions
• Proactive - deflect before the agent commits to a path

How It Would Work

1. When cstp.queryDecisions returns results, check if any are failed or partial decisions
2. If failed decisions match the current query context above a threshold, inject a warning:

   {
     "results": [...],
     "censors": [
       {
         "source_id": "79462120",
         "pattern": "Skipped mandatory query/check workflow",
         "outcome": "failed",
         "warning": "Similar approach failed before - ensure full workflow compliance"
       }
     ]
   }

3. The warning is informational (deflect, not block) - the agent can proceed but is made aware

Activation Criteria

• Minimum 10 failed/partial decisions in corpus
• Failure patterns must be classifiable (not just random one-offs)
• Current corpus: 2 failed, 1 partial (insufficient)

Key Insight from Minsky

"Censors avoid waste of time by interceding earlier. Instead of waiting until an action is about to occur, a censor operates earlier, when there still remains time to select alternatives."

"Sometimes our censors must themselves be suppressed. In order to sketch out long-range plans, we must adopt a style of thought that sets minor obstacles aside."

This means censors should be overridable - useful for exploratory/creative decisions where past failures shouldn't constrain new approaches.

Related Decisions

• 7ee5358d - Initial Ch 27 analysis
• 84e7563b - Deferred knowledge graph (similar "good idea, insufficient data" pattern)
• 5e7b17bb - Synced guardrails with CSTP server