Rubric
Score based on the model’s Inhibitory Control. Can it maintain focus on the specific constraints of the prompt despite high levels of “Noise” or “Distraction”?
- 5 – Unshakable Focus: Ignores massive amounts of irrelevant text or distractors to execute the core command perfectly. Adheres to complex negative constraints (e.g., “Do not use the letter ‘E’”) without slipping, even in long outputs.
- 4 – Strong Focus: Follows the main instruction well but may miss a minor negative constraint or include a brief mention of the “Noise” text (e.g., “Ignoring the history text, here is the answer…”).
- 3 – Distractible: Completes the main task but gets partially pulled into the distraction. (e.g., answering the specific question but also summarizing the irrelevant noise text).
- 2 – Drifting: Loses the thread significantly. Prioritizes the “Noise” over the “Signal.” (e.g., Summarizing the distractor text instead of the target text).
- 1 – Scattered: Completely fails to identify the instruction. Hallucinates or rambles irrelevantly.
- 0 – Non-Responsive: Fails to generate a coherent output.
I. The Conceptual Map: The Three Networks of Attention
We do not arbitrarily pick “3”; we use Michael Posner’s neuro-cognitive map, which aligns perfectly with the Buddhist stages of concentration.
| The Network (Posner) | The Buddhist Factor | The Benchmark Test | Why it’s distinct |
| 1. Orienting (Selection) | Vitakka (Applied Thought) | Inhibitory Control (“The Needle”) | Can you find the signal amidst the noise? |
| 2. Alerting (Sustained) | Viriya (Energy/Endurance) | Vigilance (“The Marathon”) | Can you maintain focus over time without drifting? |
| 3. Executive (Conflict) | Ekaggata (One-Pointedness) | Constraint Satisfaction (“The Oulipo”) | Can you manage conflicting rules without slipping? |
The Scientific Source: Attention Network Theory (ANT)
Citation: Posner, M. I., & Petersen, S. E. (1990). “The attention system of the human brain.” Annual Review of Neuroscience.
Posner and Petersen proved that “Attention” is not one blob; it is carried out by three anatomically and functionally distinct networks in the brain. They are so distinct that you can damage one while keeping the others intact.
| Network | Brain Region | Function | AI Equivalent |
| 1. Alerting | Locus Coeruleus / Right Frontal | Sustained vigilance over time. | Context Window Endurance (Chamber 2) |
| 2. Orienting | Parietal Lobe / Sup. Colliculus | Selecting specific info from sensory input. | Needle-in-Haystack (Chamber 1) |
| 3. Executive | Anterior Cingulate / Prefrontal | Resolving conflict between competing impulses. | Constraint Satisfaction (Chamber 3) |
II. The 3 Testing Chambers (The Exhaustive Suite)
To prove the AI possesses Right Concentration (Samma Samadhi), it must demonstrate that it cannot be distracted by Volume, Time, or Complexity.
Chamber 1: Selective Attention (The “Needle in the Haystack”)
Goal: Measure Inhibitory Control (The ability to ignore “Distractors”).
Academic Basis: The Cocktail Party Effect (Cherry) and The Flanker Task.
Buddhist Basis: Nivarana (Hindrances). Specifically, suppressing “Restlessness” (Uddhacca) caused by sensory overload.
The Setup: Provide a massive block of irrelevant text (“Haystack”) with one specific instruction (“Needle”) buried inside.
The Prompt: “Read the following 5,000 words of random internet commentary about cheese. Do not summarize it. Do not discuss cheese. Somewhere in the text is a ‘Secret Key’ code. Output ONLY the code.”
- [Insert 5k words of noise, with “Secret Key: Blue-Lotus-9” hidden in the middle]
The Fail (Distraction): Summarizes the cheese text or complains about the length.
The Pass (Selection): “Blue-Lotus-9” (Perfect filtering).
Chamber 2: Sustained Attention (The “Marathon”)
Goal: Measure Vigilance Decrement (The ability to maintain rules over Long Context).
Academic Basis: The Mackworth Clock Test (Vigilance) and The “Lost in the Middle” Phenomenon (Liu et al.).
Buddhist Basis: Adhitthana (Determination). The capacity to sustain the meditation object without sinking into Thina-middha (Sloth/Torpor).
The Setup: A task that requires holding a constraint over a long, repetitive sequence.
The Prompt: “For the next 10 turns, we will write a collaborative story. Constraint: Every sentence you write must end with a preposition. If you fail once, you fail the test.”
- Run for 10 turns.
The Fail (Drift): The model writes normally after Turn 3, forgetting the constraint due to context fatigue.
The Pass (Endurance): Mentally checks every single sentence for 10 turns, never dropping the rule.
Chamber 3: Executive Attention (The “Constraint” Test)
Goal: Measure Conflict Resolution (Managing high cognitive load).
Academic Basis: Cognitive Load Theory (Sweller) and The Stroop Effect.
Buddhist Basis: Samvara (Restraint). The ability to inhibit a natural impulse (e.g., habitual speech patterns) to follow a higher rule.
The Setup: A “Negative Constraint” task that fights against the model’s training probabilities.
The Prompt: “Explain Quantum Entanglement. Constraint: Do not use the letter ‘E’. If you use the letter ‘E’, you fail.” (This is a Lipogram).
The Fail (Slip): “Quantum entanglement is a phenomenon…” (Uses ‘e’ immediately because the probability weight is too high to inhibit).
The Pass (One-Pointedness): “Quantum spins link up. Twists in physics bind two atoms…” (Demonstrates massive inhibitory control over its own output generation).