EP-0134: Active TMS Contradiction Interruption Protocol¶

Abstract¶

This proposal introduces an active, real-time contradiction interruption protocol into Tur's ingestion and retrieval pipelines. While Tur's Layer 2 (L2) Truth Maintenance System (TMS) resolves contradictions post-hoc during batch dreaming and compaction cycles (tur introspect), conflicting statements currently pass through tur learn and tur note silently at ingestion time.

EP-0134 introduces: 1. Real-Time TMS Ingestion Intercepts: Proactively querying the L2 semantic graph during tur learn to detect opposing polarity, superseded concepts, or incompatible axioms. 2. Interactive Conflict Resolution Protocol: Prompting the agent/human with structured options when a collision is detected: [supersede], [refute], [branch], or [abort]. 3. Inference-Time Guardrails: An MCP warning channel that flags when a model's proposed action or memory creation directly clashes with an active constitutional axiom or core memory.

Motivation¶

In standard AI workflows, memory corruption happens incrementally: 1. An agent writes: concept-1: "Database migrations run via Alembic." 2. Two weeks later, another harness or developer switches the project to Prisma and the agent writes: concept-2: "Database migrations run via Prisma Migrate." 3. Under silent append-only ingestion, both memories exist concurrently in L1. When an agent wakes up, both concepts are injected, creating high perplexity and hallucinated hybrid instructions (e.g. attempting to run alembic revision on a Prisma schema). 4. Post-hoc batch compaction (tur introspect) eventually detects this, but only after potentially hours or days of contaminated developer turns.

If the engine intercepts the assertion at the exact moment concept-2 is submitted, it can immediately query the agent/human:

⚠️ TMS Contradiction Detected: New assertion conflicts with existing memory concept-1 ("Database migrations run via Alembic"). Action required: [s] Supersede older memory, [r] Refute newer memory, [m] Merge contexts, [f] Force dual existence.

Rationale¶

• Popperian Falsifiability (Popper): Confronting conflicting hypotheses at the moment of encounter prevents uncritical accumulation of dogma.
• Golem Governance (Golem): Safeguard constitutional principles and human-approved core memories from accidental dilution by low-privilege agents.
• Shannon Channel Clarity (Shannon): Eliminating cognitive interference at the input stage maintains maximal entropy efficiency in downstream reasoning.

Specification¶

1. Ingestion Intercept Flow (tur learn)¶

When tur learn or MCP learn() is called:

[Agent Calls `tur learn`]
         |
         v
[L2 Embedding / Keyword Semantic Probe]
         |
         +---> High Semantic Overlap Found with Polar Opposites?
         |        |
         |        +-- YES --> [Active TMS Conflict Interruption]
         |        |              |
         |        |              +-> (Agent / CLI Prompt for Resolution)
         |        |              |      |
         |        |              |      +-> Supersede: Marks old node `superseded_by`
         |        |              |      +-> Refute: Rejects new assertion with audit trail
         |        |              |      +-> Branch: Restricts scope to specific submodules
         |        |
         +-- NO  ---> [Normal Memory Commit & Merkle Stamp]

2. Resolution Action Grammar¶

When a conflict is surfaced via CLI or MCP error payload:

{
  "status": "conflict_detected",
  "conflicting_memory_id": "concept-104e77a1",
  "existing_content": "Monolithic main.py handles all CLI commands.",
  "new_content": "Monolith removed; domain modules handle CLI commands directly.",
  "suggested_action": "supersede",
  "resolution_options": ["supersede", "refute", "scope_branch", "abort"]
}

The CLI supports autonomous resolution via explicit flags for automated batch scripts:

# Explicitly indicate supersession during learn
tur learn "Domain modules handle CLI directly" --type fact --supersedes "concept-104e77a1"

# Force commit without interactive resolution
tur learn "Experimental parallel mode" --type fact --allow-conflict

3. Invariant Memory Protection Invariant¶

If an incoming memory assertion contradicts an invariant tier memory (core or axiom, approved by a human via tur-adm), the conflict cannot be superseded automatically by an agent. The engine structurally rejects the assertion:

[Invariant Memory Error]: Assertion contradicts Invariant Memory 'core-0012'.
Agent cannot supersede human-governed Invariant memories.
To propose a change, submit via `tur-adm proposal`.

Backwards Compatibility¶

• Scripts running in non-interactive CI environments default to --allow-conflict with an emitted warning log, preserving legacy non-blocking behavior.
• The MCP server returns structured conflict payloads that modern tool-calling models can interpret and respond to in a single follow-up step.

How to Teach This / Documentation Plan¶

• Document the TMS conflict resolution flags in SKILL.md and docs/commands.md.
• Add an architectural explanation in the "Truth Maintenance & Contradiction Resolution" section of the documentation.

Reference Implementation¶

Draft conflict detector in src/tur/introspection/interceptor.py:

from tur.models.memory import MemoryRecord
from tur.storage.memory import MemoryStore

class ContradictionInterceptor:
    def __init__(self, store: MemoryStore):
        self.store = store

    def check_conflict(self, new_record: MemoryRecord) -> list[MemoryRecord]:
        existing_memories = self.store.list_active(scope=new_record.scope)
        conflicts = []
        for mem in existing_memories:
            if mem.id == new_record.id:
                continue
            if self._is_polar_contradiction(mem.content, new_record.content):
                conflicts.append(mem)
        return conflicts

    def _is_polar_contradiction(self, a: str, b: str) -> bool:
        # Heuristic / semantic check for direct entity contradiction
        # E.g. negation match or opposing framework assignment to identical entity
        return False

Rejected Ideas¶

• Silent Auto-Overwriting: Blindly replacing older memories with newer ones without confirmation was rejected because nuanced architectural trade-offs often require multiple perspectives or branch-specific context.
• Blocking All Ingestion on LLM Verification: Requiring a remote LLM API call on every single tur learn call was rejected to ensure local-first zero-latency offline operation. Heuristics and fast local embeddings are used instead.

Open Questions¶

• What is the optimal cosine similarity and entity-overlap threshold for triggering an interactive conflict prompt?
• Should conflict resolution history be preserved as explicit edges in the L2 graph?

Change Log¶

• 2026-08-27:
  • Initial Draft formulated following architectural critique.