---
title: "Bandit Scoring for Task Prioritization"
date: "2025-12-23"
status: "Published"
updated: "2026-01-05"
readTime: "10 min"
abstract: "A practical heuristic for allocating creative attention across competing projects, inspired by multi-armed bandit theory. This is not a formal reinforcement learning implementation—it's a pragmatic decision framework that balances exploitation of proven ideas with exploration of new concepts."
codexMode: "System"
codexClass: "Extensions"
heroImage: "/media/codex/bandit-scoring.png"
keywords:
  - "bandit scoring"
  - "task prioritization"
  - "creative operations"
  - "exploration exploitation"
version: "v1.0"
author: "Andy"
topicTags:
  - "task-prioritization"
  - "decision-making"
  - "cognitive-science"
difficultyLevel: "Intermediate"
targetRoles:
  - "Studio founder"
  - "Operations leader"
  - "Solo creator"
prerequisites:
  - slug: "multiclock"
    title: "Multi-Clock Work Framework"
    reason: "Provides HF/LF/Dormant context for scoring"
tldr: "Allocate creative attention using simple indices that balance exploitation of proven ideas with exploration of new concepts. Score tasks 0-22 based on Revenue, Portfolio Value, Engagement, and Exploration—then pick the highest score 90% of the time."
---

import Figure from '../../components/CodexFigure.astro';
import RelatedArticles from '../../components/CodexRelatedArticles.astro';
import Callout from '../../components/CodexCallout.astro';
import CodexGlossary from '../../components/CodexGlossary.astro';

<RelatedArticles
  items={[
    {
      title: "Multi-Clock Work Framework",
      href: "/codex/multiclock/",
      description: "The three-band operating system that provides context for Bandit Scoring."
    },
    {
      title: "The Multi-Clock Problem",
      href: "/codex/multi-clock-problem/",
      description: "Why single-rhythm calendars fail creative work."
    }
  ]}
/>

<CodexGlossary>
  <dt>Bandit Score</dt>
  <dd>Composite prioritization metric (0-22) combining four signals plus urgency</dd>
  
  <dt>Exploitation</dt>
  <dd>Working on tasks with proven high value signals</dd>
  
  <dt>Exploration</dt>
  <dd>Trying new approaches that might be hidden gems</dd>
  
  <dt>ε-Greedy</dt>
  <dd>Strategy: exploit 90% of the time, explore 10%</dd>
</CodexGlossary>

<Callout title="Context">
This is how we run our studio. Yours will differ—adapt it as needed. Nothing here is a promise of results; use your judgement.
</Callout>

## What is the Bandit Score?

The Bandit Score answers one question: **"Which task should I work on next when I have multiple options?"**

It balances two competing goals:
- **Exploitation:** Focus on work with proven high value
- **Exploration:** Try new approaches that might be hidden gems

<Callout title="Important Disambiguation">
  While inspired by multi-armed bandit theory from reinforcement learning, this scoring system is a pragmatic heuristic for creative portfolio decisions—not a formal RL implementation. There are no regret bounds or probabilistic reward models. Just structured decision pressure that biases selection toward high-signal work while preserving exploration.
</Callout>

### The ε-Greedy Strategy

A useful starting point:

- **90% of the time:** Pick the highest Bandit Score (exploit known winners)
- **10% of the time:** Deliberately work on something with the "Exploration" flag (explore new territory)

Adjust based on your workload and risk tolerance. This helps prevent teams from over-optimizing for short-term wins while giving strategic bets a better chance of getting tested.

<Figure
  src="/media/codex/bandit-score-diagram.png"
  alt="Bandit scoring visualization showing the balance between exploration and exploitation"
  caption="Fig 1. How bandit indices balance recent success with uncertainty"
/>

## When to Score Tasks

### On Task Creation
Set initial estimates based on best available information. It's okay to guess—scores evolve.

### After Task Completion
Update retrospectively based on actual outcomes. This calibrates the system over time.

### During Weekly Review
Adjust when new information emerges:
- Client showed unexpected enthusiasm → increase Engagement
- Revenue path became clearer → increase Revenue Potential
- Tried it and learned something → mark as no longer Exploration

## The Five Scoring Signals

### 1. Revenue Potential (0-3 scale)

**Question:** Could this task generate direct or indirect revenue?

| Score | Definition | Examples |
|-------|------------|----------|
| **0 = None** | Pure R&D, no revenue path visible | Internal process docs, personal learning, speculative research |
| **1 = Low** | Indirect revenue (brand building → future sales) | Instagram Stories, Substack essays, brand development |
| **2 = Medium** | Direct but uncertain revenue path | Pilot commission (unconfirmed), experimental offering, untested price point |
| **3 = High** | Confirmed or highly probable revenue | Paid client deposit, qualified lead with budget, proven offering |

**Examples from practice:**
- Client commission inquiry → **3** (confirmed interest, potential £3,500)
- Multi-Clock Whitepaper → **0** (no direct revenue, IP development)
- DM outreach to supercar owners → **1** (indirect, may generate future leads)

### 2. Portfolio Value (0-3 scale)

**Question:** Does this build long-term capability, brand equity, or showcase value?

| Score | Definition | Examples |
|-------|------------|----------|
| **0 = None** | One-off task, no reuse or showcase value | Administrative cleanup, one-time email response |
| **1 = Low** | Minor reusable asset or small improvement | Template creation, minor process tweak, single social post |
| **2 = Medium** | Significant capability or quality showcase work | MIR system implementation, pilot commission, marketing campaign |
| **3 = High** | Foundational IP or flagship work that defines brand | Core methodology documentation, first gallery exhibition, signature technique |

**Examples:**
- Pilot Commission → **2** (first Survivor, proves concept, showcase work)
- Multi-Clock Whitepaper → **3** (foundational IP, publishable research)
- Email template updates → **1** (useful but minor asset)

### 3. Engagement Signal (0-3 scale)

**Question:** Has this received external validation or interest?

| Score | Definition | Examples |
|-------|------------|----------|
| **0 = None** | No external feedback yet | Brand new idea, not yet shared, internal-only work |
| **1 = Low** | Polite interest, minimal engagement | 1-2 likes, "interesting concept" comment, single polite inquiry |
| **2 = Medium** | Clear market interest | Multiple DMs/inquiries, decent engagement (20+ likes), shared by others |
| **3 = High** | Strong demand or viral signal | Multiple qualified leads, viral post (100+ engagements), waitlist forming |

**Important:** This signal updates over time. Start at 0, increase as feedback accumulates.

### 4. Exploration Bonus (+2 or 0)

**Question:** Have we done this type of work before?

| Value | Definition | When to Apply |
|-------|------------|---------------|
| **+2** | Exploration: Never tried this before | New market, new format, new medium, new client type |
| **0** | Exploitation: Familiar territory | Similar to past work, proven format, repeat client type |

**Key Rule:** The first time you try something = Exploration. Second time onwards = Exploitation (even if details differ).

**Examples:**
- First PAID commission → **+2** (new territory, even if pilot existed)
- First long-form whitepaper → **+2** (new format)
- Second commission → **0** (no longer exploring "paid commission" space)

**Why This Matters:** The +2 bonus nudges you to try new things by making them competitive with established high-performers.

### 5. Urgency Multiplier (×1 or ×2)

**Question:** Is there a hard deadline within 7 days?

| Multiplier | Condition | Examples |
|------------|-----------|----------|
| **×2** | Deadline within 7 days | Client response needed, event date, publication deadline |
| **×1** | No deadline or more than 7 days out | Flexible timing, internal deadlines, "whenever" tasks |

**Applied Last:** This multiplier doubles the entire base score.

**Note:** Deadlines create artificial urgency. Use sparingly—not everything needs one.

## Calculating the Final Score

### Formula

```
Bandit Score = (Revenue + Portfolio + Engagement + Exploration) × Urgency
```

### Score Ranges & Interpretation

| Range | Interpretation | Action |
|-------|----------------|--------|
| **16-22** | Critical priority | Work on immediately (HF today) |
| **12-15** | High priority | Schedule in next 2-3 days (HF this week) |
| **8-11** | Medium priority | Good LF refresh candidate |
| **4-7** | Low priority | Dormant or far-future LF |
| **0-3** | Minimal signal | Consider archiving unless strategic |

## Worked Examples

<Callout title="Quick note">
The scores and £ figures below are illustrations—not promises. Tune to your reality.
</Callout>

### Example 1: Client Commission Inquiry

**Context:** Client inquiry received, wants commission for their supercar, asks about pricing.

**Scoring:**
- Revenue Potential: **3** (confirmed interest, £3,500 if closes)
- Portfolio Value: **2** (showcase work, builds client portfolio)
- Engagement Signal: **2** (referenced specific details = engaged)
- Exploration Bonus: **+2** (first paid commission, new client type)
- Urgency Multiplier: **×2** (respond same-day to maintain momentum)

**Calculation:** (3 + 2 + 2 + 2) × 2 = **18**

**Decision:** HIGH PRIORITY → Draft response immediately, create follow-up tasks in MIR.

### Example 2: Framework Whitepaper

**Context:** Research project documenting the operational framework.

**Scoring:**
- Revenue Potential: **0** (no direct revenue, IP development)
- Portfolio Value: **3** (foundational IP, publishable)
- Engagement Signal: **1** (internal validation, not yet public)
- Exploration Bonus: **+2** (first long-form whitepaper format)
- Urgency Multiplier: **×1** (no hard deadline)

**Calculation:** (0 + 3 + 1 + 2) × 1 = **6**

**Decision:** MEDIUM PRIORITY → Good LF refresh work. Schedule 2-3 hour blocks weekly.

### Example 3: Behind-the-Scenes Content

**Context:** Post behind-the-scenes footage from a shoot.

**Scoring:**
- Revenue Potential: **0** (brand awareness, no direct revenue)
- Portfolio Value: **1** (content library, minor asset)
- Engagement Signal: **2** (proven format, consistent 30-50 views)
- Exploration Bonus: **0** (familiar format, done many times)
- Urgency Multiplier: **×1** (flexible timing)

**Calculation:** (0 + 1 + 2 + 0) × 1 = **3**

**Decision:** LOW PRIORITY → Fill-in work when you have 15-20 min gaps. Don't prioritize over higher-scoring tasks.

### Example 4: Outreach Target List

<Callout title="Outreach, done properly">
If your next action involves outreach or DMs, keep it respectful: don't scrape personal data, follow platform rules, and make sure your messaging fits the relationship and context.
</Callout>

**Context:** Build spreadsheet of potential clients to DM.

**Scoring:**
- Revenue Potential: **1** (indirect, may generate future leads)
- Portfolio Value: **1** (reusable list, but not showcase work)
- Engagement Signal: **0** (not yet executed, no validation)
- Exploration Bonus: **0** (DM outreach is familiar tactic)
- Urgency Multiplier: **×2** (part of 30-day sprint, due soon)

**Calculation:** (1 + 1 + 0 + 0) × 2 = **4**

**Decision:** MEDIUM-LOW PRIORITY → Must do because of deadline, but intrinsically low-signal. Schedule a focused block, get it done, move on.

<Figure
  src="/media/codex/bandit-score-example.png"
  alt="Example bandit score calculation"
  caption="Fig 2. Worked example showing how signals combine into final score"
/>

## Using Bandit Scores in Practice

### Daily HF Work: "HF Priority Queue"

**Filter:** Clock = HF, Status ≠ Done  
**Sort:** Bandit Score DESC, then Next Trigger ASC  
**Decision Rule:** Pick highest score, unless mid-burst (finish-to-switch)

### Weekly LF Refresh: "LF Candidates"

**Filter:** Clock = LF  
**Sort:** Bandit Score DESC  
**Decision Rule:**
- Select 1× highest Bandit Score (exploit)
- Select 1× random "Is Exploration" (explore)

### Fortnightly Dormant Review

**Filter:** Clock = Dormant, Age Points above 15  
**Decision Rule:** Promote if Bandit Score above 5, otherwise archive or extend dormancy

## Exploration Budget Management

### Target

**10-20% of active WIP** should have "Is Exploration" ✓

### Weekly Check

Count tasks:
- Total active (HF + LF): Example = 15
- Exploration tasks (✓): Example = 3
- Exploration %: 3/15 = 20% ✅

### If Under 10% Exploration

**Action:** Force-promote a Dormant exploration item to LF or HF, even if Bandit Score is medium-low.

**Why:** Prevents getting stuck in local maxima. Innovation requires trying new things, even when current work is "good enough."

## Common Pitfalls & How to Avoid Them

### Pitfall 1: Everything Gets High Scores

**Problem:** If you score too generously, everything looks equally important.

**Fix:**
- Use 0 liberally—most tasks have no revenue path
- Reserve 3s for truly exceptional cases
- Calibrate by comparing: "Is Task A really more valuable than Task B?"

### Pitfall 2: Ignoring Exploration Bonus

**Problem:** Only working on proven winners, missing new opportunities.

**Fix:**
- Track exploration % weekly
- If under 10%, force yourself to try something new
- Remember: Today's exploit was yesterday's exploration

### Pitfall 3: Overusing Urgency Multiplier

**Problem:** Giving everything ×2 defeats the purpose.

**Fix:**
- Reserve ×2 for true time-sensitive deadlines (within 7 days)
- Most internal deadlines are flexible—be honest
- If everything is urgent, nothing is urgent

### Pitfall 4: Never Updating Scores

**Problem:** Scores get stale, lose predictive value.

**Fix:**
- Update after task completion (retrospective calibration)
- Adjust when new information emerges
- Review scoring quality during weekly reviews

## Integration with Multi-Clock Mechanics

### Aging & Bandit Score

- **Aging increases urgency** (promotes old LF/Dormant work)
- **Bandit Score prioritizes within a Clock** (which HF task to work on today?)
- They work together: Aging says "don't forget me," Bandit says "work on me next"

### WIP Limits & Exploration

- WIP caps prevent overload
- Exploration % ensures we don't just exploit
- Together: Disciplined focus + strategic experimentation

### Kill Criteria

If a task has:
- Bandit Score = 0 (no signals)
- Age Points exceeds Threshold × 2
- 2+ reviews without progress

**Consider archiving.** Low score + high age + no momentum = probably not viable.

## Scoring Workflow Checklist

### When Creating a Task

- [ ] Assign Revenue Potential (0-3)
- [ ] Assign Portfolio Value (0-3)
- [ ] Assign Engagement Signal (0, initially)
- [ ] Check "Is Exploration" if never tried before
- [ ] Add Due Date if hard deadline exists (within 7 days)
- [ ] System auto-calculates Bandit Score

### During Weekly Review

- [ ] Update Engagement Signals based on new feedback
- [ ] Adjust Revenue/Portfolio if new info emerged
- [ ] Remove "Is Exploration" if we've now tried it once
- [ ] Check overall exploration % (target 10-20%)

### After Task Completion

- [ ] Retrospective scoring: Was Revenue/Portfolio accurate?
- [ ] Document learnings in Notes field
- [ ] Use findings to calibrate future similar tasks

## Advanced: Tuning Signal Weights

*For use after 4-6 weeks of scoring data*

### Retrospective Analysis

After completing 20+ scored tasks, analyze:
1. Did high-Bandit tasks actually deliver high value?
2. Which signal was most predictive? (Revenue? Portfolio? Engagement?)
3. Did we explore enough (10-20%)?

### Potential Adjustments

If Revenue consistently dominates, consider:
- Weighting Portfolio Value ×1.5 (encourages long-term thinking)
- Increasing Exploration Bonus to +3 (forces more experimentation)

<Callout title="Warning">
  Don't tune prematurely. Let the system run for 4-6 weeks before adjusting weights. Early tuning optimizes for noise, not signal.
</Callout>

## Implementation Notes

### Notion Formula (Reference)

The Bandit Score can be implemented as a Notion formula field:

```
((prop("Revenue Potential") + prop("Portfolio Value") + prop("Engagement Signal") + 
  if(prop("Is Exploration"), 2, 0)) * 
  if(dateBetween(prop("Due Date"), now(), "days") <= 7, 2, 1))
```

### View Configuration

**Today View Filter:**
```
Clock = HF AND 
Status ≠ Done AND 
(Bandit Score >= 6 OR Promotion Flag = "Promote" OR Days Until Deadline within 7)
```

**Sort:** Bandit Score DESC, Next Trigger ASC

## Summary

The Bandit Score helps you decide **"What should I work on next?"** by scoring tasks 0-22 based on Revenue potential, Portfolio value, Engagement signals, Exploration bonus, and Urgency.

Pick the highest score 90% of the time (exploit proven winners), but deliberately try something new 10% of the time (explore hidden gems). Update scores as you learn.

It's not perfect, but it **can help reduce decision paralysis** while keeping you from getting stuck only doing familiar work.

## Related Frameworks

- [Multi-Clock Work](/codex/multiclock/) - Task scheduling across time horizons
- [Governance Protocols](/codex/governance-protocols) - AI-delegated work patterns
- [Campaign Framework](/codex/campaign-framework) - Marketing orchestration

---

*For questions or implementation support, see the MIR Database Schema Reference or consult the Multi-Clock Work whitepaper.*