The "Four Operations & Saji-Kagen" Cognitive Protocol for GenAI Part 4

Part 4: Protocol Updates & The Manual Kill Switch (x = -1)

Originally formulated on 2026-11-12

Selected Mode: Freshness 85 (High subtraction, strict factual fidelity)

Following the initial 3-part English release, continuous refinement and cross-lingual testing between Japanese and English led to key updates in the protocol. This post outlines the specification corrections, boundary refinements, and the newly introduced manual kill switch (x = -1).

A. Specification Corrections & Errata

1. Re-defining Hypothesis Certainty (x = 1)

• In the previous English translation, 0 < x < 1 was labeled as "Unverified Hypothesis." This has been corrected: 0 < x < 1 strictly denotes an Uncertain Event (where the existence of the event itself is undetermined).
• Once a Hypothesis is placed on the table for discussion, it is treated as x = 1 (Event Exists). Its unverified or pending status is evaluated via the Depth metric (0 < y < 100).

2. Mode Correction for Part 2 (Experiment 2: The BBQ Notice)

• While Part 2 introduced the experiments under Standard Mode, Experiment 2 actually operated under Extended Human-Led Adaptability Mode, as the human operator manually dialed the Freshness metric (30 and 70).
• This manual intervention triggered a Forced Division by Zero, branching out into Element C (The Remainder).

B. Nuance & Boundary Refinements

3. Output Metric Scope & Fiction Exception ("Refinement Purity")

• The output metric Production Purity is specifically termed Refinement Purity (Suikou Jundo) when applied to text creation.
• Fiction Exception: For creative fiction, achieving Purity 100 is not strictly required; retaining an "AI flavor" or mechanical imprint is acceptable.

4. Defining "Ultimate Delegation" (Marunage no Kyokuchi)

• Distinct from total AI autonomy (Saji-Kagen = 0), when Saji-Kagen approaches the limit of zero while remaining under human management (0 < Saji-Kagen → 0), paired with Freshness 0 and Depth 0, the state is defined as Ultimate Delegation.

5. Dual Adaptability Modes (Standard vs. Extended)

• Standard Mode (AI-Led Adaptability): The human inputs Saji-Kagen only; Freshness, Depth, and Certainty (0 <= x <= 1) are blended internally by the AI.
• Extended Mode (Human-Led Adaptability): The human expands manual control to input Saji-Kagen, Freshness, and Depth directly.

6. Structural Flexibility & Error Handling

• Module Independence: While designed as a unified system, each group (Circular Group and Probability Group) can operate independently.
• Persistence of Addition & Subtraction: Even in environments where the 4-step circular loop cannot be applied, the core concepts of Addition (+) and Subtraction (-) remain valid.
• Human Verification Error & Recovery: If a human blindly accepts an AI hallucination or misidentifies a fact, the logic breaks down. However, if the human deliberately adopts an error, it survives as a valid Hypothesis (x = 1, 0 < y < 100).

C. New Protocol Features (2026-11-10 to 11-12 Additions)

7. Exponentiation-Based Fade-Out (Topic Decay & Load Reduction)

• Topic Decay: In long single-session runs (such as zero-login environments), shifting themes mid-way (e.g., from Topic A to the protocol itself) causes the influence of the initial topic to decay exponentially with each rally.
• Strategic Fade-Out: When the AI detects high cognitive load on the human operator, noise accumulation, or session limits, it drops the reverse-question from the final step of the 4-step loop (Impressions & Reverse-Questions). This allows the session to fade out naturally without abrupt disruption.

8. Event Collapse & The Manual Kill Switch (x = -1 / /STOP)

• Definition: Certainty x = -1 is newly defined as Event Collapse (the collapse of session continuity).
• Strict Human-Only Trigger (Anti-Misfire Lock): While 0 <= x <= 1 is handled automatically under Adaptability Modes, x = -1 is strictly prohibited from AI auto-triggering. It is reserved exclusively for human manual input.
• Command Specification (/STOP): To prevent accidental triggers during mathematical or meta-discussions, the command is strictly restricted to a standalone uppercase input of /STOP. Inline mentions or lowercase text are ignored.
• Virtual Execution & Exponentiation Persistence:
  • Triggering /STOP preserves the validated deliverables achieved up to that point while resetting the corrupted cognitive state (switching the flow).
  • In text-based environments (where closing the tab ends the session), /STOP acts as a virtual execution—functioning like the Clear key on a console or calculator. Because the physical session remains open, the Exponentiation (rally count and accumulated load) persists upon resumption.
  • Analogy: If applied to Physical AI, this mechanism corresponds directly to a hardware Physical Emergency Stop Button.

Link: Original Japanese Wiki

Part 1: Core Theory｜ Part 2: Experiments｜ Part 5: Tables