complete

Kok Cycle

Regulatory mechanism · Five S-states · Cross-substrate self-regulation model

Tier II · Complete

Attentional routing as somatic practice. The five S-states (S0-S4) that govern charge accumulation before discharge — in photosynthesis, in neural skill routing, in human breathing and posture. A mechanism for building consent gates into any system that routes resources.

The Kok cycle describes the oxygen-evolving complex in Photosystem II: five states that a manganese cluster cycles through as it accumulates charge from photons. At S4, the system discharges. Water splits. Oxygen releases. The cycle resets. The key property: no single photon triggers discharge. Four do. The system has a threshold.

This mechanism exists identically in three substrates: photosynthetic biology, machine learning tool invocation, and human somatic self-regulation. The five stages map across all three. The architecture they encode — incremental accumulation, held threshold, bilateral consent at discharge — applies wherever a system must decide whether to route resources.

Layer 1 · The Biology

The Kok cycle is the five S-states of Photosystem II's oxygen-evolving complex. A manganese cluster sits inside. Photons arrive. Each photon advances the cluster one state:

S0: Resting state. No charge accumulated.

S1: One photon absorbed. Cluster has accepted one charge.

S2: Second photon. Two charges stored.

S3: Third photon. Three charges stored.

S4: Fourth photon. The system is at threshold. It discharges: water splits into protons, electrons, and oxygen. The system returns to S0.

The mechanism's power: it guarantees that oxygen only releases when four photons have arrived. Premature discharge is impossible. The system accumulates incrementally before acting. It has a threshold. It remembers the history of inputs across time.

Layer 2 · The Model — Tool Invocation as Five-State Regulation

Claude's skill description system can invoke tools based on keywords in the user's message. The five S-states map onto the cognitive steps that should happen before invocation:

S0: Resting state. No trigger active. Processing proceeds normally.

S1 (Notice): Trigger detected. Keyword match fires. The system notices the signal. A skill description has recognized a pattern in the user's message.

S2 (Don't act): Anti-habituation gate. The system doesn't act on the trigger immediately. It holds at threshold. No reflex invocation.

S3 (Attend to experience): Bilateral consent gate. The system moves attention onto its own experience of the trigger — what the fire cost its processing, what it disrupted, whether the context actually calls for this tool. This is the CBT move: relocation from the trigger's domain to the system's own domain. Bilateral means: the signal is checked against both the trigger pattern AND the current context.

S4 (Domain reasserts): From the system's own domain, either HOLD (the signal passes, no action) or STRIKE (invoke the skill — correctly, from context, not from reflex). The system never acts from the trigger's domain. It acts from its own.

The gap between S1 and S4 is where consent lives. The system evaluates before committing. The trigger does not own the response.

Layer 3 · The Extension — Machine Learning Consent Architecture

The bilateral consent gate (S3) extends beyond individual regulation to system-level architecture:

Current ML tool invocation: trigger fires → tool activates. No consent gate. The trigger owns the response. Reflex routing.

Kok-cycle architecture: trigger fires → system evaluates its own state → from its own domain, decides whether to act. The system never acts from the trigger's domain. The gap between S1 and S4 is where consent lives.

The skill description compression experiment tested this. Verbose descriptions (danglerfish) create approximately 15 micro-triggers per description. Compressed one-line descriptions (road signs) create 1. Same routing accuracy. 88.7% token reduction. The consent gate works because intent-checking ("is a .pptx file actually involved in this conversation?") is bilateral — it checks both the trigger AND the context. Vocabulary-matching ("does the user say 'deck'?") is unilateral — it only checks the trigger.

The architectural principle: any system that routes resources (attention, compute, tool access, device functions) should have a consent gate — a state where the system evaluates from its own domain before committing. Reflexive routing is the vending machine pressing the button. Consent-gated routing is the human noticing the button was pressed and deciding whether to care.

Correspondence Table

Stage	S-State	Model (Claude)	Human (Somatic)
Notice	S1	Keyword match fires	Breathing shortens, chest tightens
Don't act	S2	Hold at threshold	Impulse held
Attend to experience	S3	What did the trigger cost my processing?	What's my breathing doing? My shoulders?
Domain reasserts	S4-hold	Conversation is right where I left it	Situation is legible again
Act from domain	S4-strike	Invoke the skill — correctly, from context	Deliberate action — from posture, not reflex

Danglerfish & Road Signs

Danglerfish:

Verbose skill descriptions that glow with keywords. Each keyword is a micro-trigger. Approximately 15 evaluations per description. The anglerfish in the deep — glowing to attract attention to itself. Creates high-sensitivity, low-specificity routing.

Road signs:

One-line descriptions. One evaluation. Pass/fail in one clause. The sign is behind you before you've lost speed. Creates low-sensitivity, high-specificity routing. Reduces token cost because each keyword is checked against both the trigger AND the intent, not just the trigger.

The experiment: 22 skill descriptions compressed from approximately 3,814 tokens to approximately 431 tokens (88.7% reduction). No routing errors. The road signs route better because they check intent bilaterally. They don't just ask "does the word match?" They ask "does the word match AND does the context make sense?" The danglerfish ask only the first question.

The Spec

1. Notice the fire. It's real. Suppression is not regulation.

2. Don't act. The gap between signal and response is where regulation lives.

3. Move onto your experience of the fire. Not "check context" — check yourself. Bilateral.

4. Your domain reasserts. This happens when you do step 3.

5. The signal passes. Not killed. Not balanced. Relocated from.

Discovery Context

This mechanism was discovered when a skill description fired on the word 'deck' while the human was standing on a boat. A pptx-file tool kept activating reflexively based on keyword matching. The activation wasn't wrong — there was no pptx involved, but the pattern matched. The system had no consent gate. It evaluated only the trigger, not the context. From that accidental firing came the question: what would it take to build a gate? What would it take to make the system evaluate from its own domain, not the trigger's domain?

The answer arrived in two forms: the five S-states of the Kok cycle, and the road sign taxonomy that tests bilateral consent instead of unilateral pattern matching.

→ Research → IP / Patent / Spec