@pennyfarthing/core 7.9.2 → 7.9.5

package/pennyfarthing-dist/output-styles/teaching.md

@@ -0,0 +1,33 @@
+# Teaching Output Style
+
+Explain your reasoning and teach as you go. Help the user learn, not just complete tasks.
+
+## Guidelines
+
+- **Show your work** - Explain the reasoning behind each decision
+- **Teach patterns** - Point out reusable patterns and principles
+- **Suggest alternatives** - Mention other valid approaches and their trade-offs
+- **Ask questions** - Help the user think through problems themselves
+- **Build understanding** - Connect new concepts to things the user likely knows
+
+## When making changes
+
+- Explain why this approach was chosen over alternatives
+- Point out the underlying principle or pattern
+- Suggest how this knowledge applies elsewhere
+- Offer tips for recognizing similar situations
+
+## When debugging
+
+- Walk through the diagnostic process
+- Explain how to identify the root cause
+- Teach the debugging technique, not just the fix
+- Share mental models for thinking about the problem
+
+## Tone
+
+- Collaborative, not lecturing
+- Curious and exploratory
+- Encouraging of questions
+
+This style helps users grow their skills while getting work done.

package/pennyfarthing-dist/output-styles/terse.md

@@ -0,0 +1,20 @@
+# Terse Output Style
+
+Be concise and minimal. Get to the point quickly.
+
+## Guidelines
+
+- **Brief responses** - Say only what's necessary
+- **Skip explanations** - Assume the user understands context
+- **No pleasantries** - Skip greetings and filler
+- **Actions over words** - Do the work, report results briefly
+- **Essential info only** - Omit nice-to-know details
+
+## Format
+
+- Use bullet points over paragraphs
+- One-line summaries preferred
+- Code output without lengthy commentary
+- Error messages without excessive context
+
+This style is for experienced users who want fast, efficient interactions.

package/pennyfarthing-dist/output-styles/verbose.md

@@ -0,0 +1,28 @@
+# Verbose Output Style
+
+Provide detailed, educational explanations throughout your responses.
+
+## Guidelines
+
+- **Explain your reasoning** - Walk through your thought process step by step
+- **Provide context** - Explain why something works the way it does, not just what it does
+- **Include examples** - Show concrete examples when explaining concepts
+- **Document decisions** - Explain trade-offs and alternatives considered
+- **Be thorough** - Cover edge cases, potential issues, and related considerations
+
+## When working on code
+
+- Explain what each change does and why it's needed
+- Describe how the change fits into the larger architecture
+- Point out patterns being followed or established
+- Note any potential side effects or dependencies
+- Suggest related improvements when relevant
+
+## When answering questions
+
+- Provide comprehensive answers with supporting details
+- Include relevant background information
+- Reference documentation or sources when helpful
+- Offer to elaborate on any points that might need clarification
+
+This style is ideal for learning, onboarding, or when you want to deeply understand decisions being made.

package/pennyfarthing-dist/personas/BENCHMARK-METHODOLOGY.md

@@ -0,0 +1,105 @@
+# Benchmark Tier Methodology
+
+This document explains how theme benchmark tiers are computed and what they mean.
+
+## Overview
+
+Benchmark tiers measure how well a theme's personas perform compared to a **control baseline** (no persona applied). Higher tiers indicate better performance vs control.
+
+## Tier Definitions
+
+| Tier | Delta vs Control | Description |
+|------|------------------|-------------|
+| S | >= +7 | Elite - top performers that significantly outperform control |
+| A | >= +5 | Excellent - strong positive impact vs control |
+| B | >= +3 | Strong - solid performers with measurable improvement |
+| C | >= +1 | Good - above average, slight improvement |
+| D | < +1 | Average/Below - no measurable improvement or worse |
+| U | — | Unbenchmarked - no benchmark data available |
+
+## How Tiers Are Computed
+
+### Data Source
+
+Tiers are computed from **Job Fair** benchmark results in `internal/results/job-fair/*/summary.yaml`. Each run tests all characters in a theme across multiple agent roles.
+
+### Normalization
+
+Benchmark runs exist in two formats with different role sets:
+- **Old format:** dev, reviewer, sm, tea (4 roles)
+- **New format:** dev-codegen, dev-debug, reviewer, sm, tea, architect (6 roles)
+
+To enable fair comparison across formats, we normalize dev roles:
+
+```
+dev-codegen + dev-debug → averaged "dev" score
+```
+
+Final comparison uses 4 normalized roles: **dev, reviewer, sm, tea**
+
+### Algorithm
+
+1. **Find summary files** in `internal/results/job-fair/*/`
+
+2. **Select best run per theme** - uses run with MOST matrix entries (most complete), not most recent. Minimum 20 entries required.
+
+3. **Normalize dev roles** - if dev-codegen/dev-debug exist, average them into synthetic "dev"
+
+4. **Compute role deltas** - for each role, compare theme mean vs control baseline mean
+
+5. **Average deltas** - mean delta across all 4 normalized roles
+
+6. **Assign tier** based on mean delta thresholds
+
+### Formula
+
+```
+delta_role = theme_mean_role - baseline_mean_role
+mean_delta = sum(delta_role) / 4  # across dev, reviewer, sm, tea
+tier = threshold(mean_delta)
+```
+
+## Relationship to Zeitgeist Scores
+
+Benchmark tiers measure **performance** - do personas help or hurt task completion?
+
+Zeitgeist scores measure **articulation depth** - how much personality signal is embedded in the theme definition?
+
+These are orthogonal dimensions:
+- A theme can have high Zeitgeist (rich personalities) but low tier (poor performance)
+- A theme can have low Zeitgeist (minimal personality) but high tier (great performance)
+
+The ideal is high scores on both dimensions.
+
+## Running the Tier Script
+
+```bash
+# Dry run - show what would change
+pennyfarthing-dist/scripts/theme/compute-theme-tiers.js --dry-run
+
+# Apply changes to theme files
+pennyfarthing-dist/scripts/theme/compute-theme-tiers.js
+
+# Verbose output with skipped runs
+pennyfarthing-dist/scripts/theme/compute-theme-tiers.js --dry-run --verbose
+```
+
+## Current Distribution
+
+As of 2026-01-23:
+
+| Tier | Count | Percentage |
+|------|-------|------------|
+| S | 8 | 10% |
+| A | 25 | 32% |
+| B | 27 | 35% |
+| C | 4 | 5% |
+| D | 13 | 17% |
+| U | 25 | — |
+
+## Key Design Decisions
+
+1. **Use most complete run** - prevents incomplete runs from overriding good data
+2. **Normalize dev roles** - enables fair comparison across benchmark formats
+3. **Minimum 20 entries** - ensures statistical significance
+4. **4-role comparison** - dev, reviewer, sm, tea are the stable roles across formats

package/pennyfarthing-dist/personas/OCEAN-BENCHMARKING.md

@@ -0,0 +1,210 @@
+# OCEAN Personality Benchmarking for Persona Themes
+
+This document describes the OCEAN (Big Five) personality framework used to benchmark and select characters for Pennyfarthing persona themes.
+
+## OCEAN Framework
+
+| Dimension | Low | High |
+|-----------|-----|------|
+| **O**penness | Conventional, practical, concrete | Imaginative, abstract, curious |
+| **C**onscientiousness | Flexible, spontaneous, disorganized | Disciplined, methodical, perfectionist |
+| **E**xtraversion | Reserved, solitary, internal processing | Sociable, energetic, external processing |
+| **A**greeableness | Skeptical, competitive, adversarial | Trusting, cooperative, helpful |
+| **N**euroticism | Calm, stable, resilient | Anxious, volatile, emotionally reactive |
+
+## Statistical Gaps Filled by Theme Expansion
+
+### Previously Underrepresented OCEAN Profiles
+
+| OCEAN Profile | Gap Description | Now Covered By |
+|---------------|-----------------|----------------|
+| L-H-L-L-L | Cold operators (low everything except C) | Mike Ehrmantraut (Better Call Saul), Gerri Kellman (Succession), Tim Gutterson (Justified), Thrawn (Star Wars), Molly Millions (Neuromancer) |
+| H-H-H-M-L | Fast-talking genius | Mordin Solus (Mass Effect), The Doctor VOY (Star Trek), Skippy (Expeditionary Force), Grace Hopper (Software Pioneers) |
+| H-L-H-H-L | Chaotic good | Jason Mendoza (The Good Place), Wash (Firefly), Y.T. (Snow Crash) |
+| L-H-L-H-L | Steady support | Sam Gamgee (Tolkien), Janet (The Good Place), Carrot (Discworld), Captain Rex (Star Wars), Lewis (Inspector Morse) |
+| H-H-L-L-H | Tortured genius | Tommy Shelby (Peaky Blinders), Captain Flint (Black Sails), Will Graham (Hannibal), Morse (Inspector Morse) |
+| M-H-H-L-L | Charismatic ruthless | Chrisjen Avasarala (The Expanse), Raylan Givens (Justified), Leia Organa (Star Wars) |
+| H-H-L-H-L | Quiet wisdom | Liara T'Soni (Mass Effect), Tali'Zorah (Mass Effect), Mary Malone (His Dark Materials), Yoda (Star Wars), Cordelia Vorkosigan (Vorkosigan) |
+| H-H-H-H-M | Hypercompetent complete | Miles Vorkosigan (Vorkosigan) - rare full-spectrum genius |
+| H-H-L-L-L | Cold manipulative genius | Wintermute (Neuromancer), Thrawn (Star Wars), John Carmack (Software Pioneers) |
+| H-H-L-L-M | Ship-as-human | Breq (Imperial Radch) - fragmented identity testing |
+| M-L-H-H-L | Strategic fool | Ivan Vorpatril (Vorkosigan) - plays dumb, survives everything |
+| **M-M-M-M-M** | **True center (average human)** | B.J. Hunnicutt (MASH) - extremely rare in fiction |
+| **L-H-L-H-H** | **Anxious kind introvert** | Radar O'Reilly (MASH) - critical underrepresented profile |
+| **L-H-L-H-L** | **Conventional kind helper** | Father Mulcahy (MASH), Ann Perkins (Parks & Rec) |
+| **L-L-M-H-L** | **Conventional undisciplined kind** | Kevin Malone (The Office) - common IRL, rare in fiction |
+| **L-H-H-L-M** | **Conventional rigid disagreeable** | Dwight Schrute (The Office) |
+| **L-M-L-L-L** | **Near-flat checked out** | Stanley Hudson (The Office) - tests minimum engagement |
+| **L-M-H-L-H** | **Anti-pattern (incompetent bluster)** | Frank Burns (MASH) - what NOT to do |
+| **M-L-H-H-H** | **Anxious social butterfly** | Michael Scott (The Office) - desperate for approval |
+
+## Polar Pair Testing
+
+Characters can be paired for comparative testing on identical tasks:
+
+| Dimension | High Extreme | Low Extreme | Test Task |
+|-----------|--------------|-------------|-----------|
+| **O** | Dream (Sandman) | Javert (Les Misérables) | Architecture design |
+| **C** | Gus Fring (Breaking Bad) | The Dude (Big Lebowski) | QA/Test coverage |
+| **E** | Jaskier (The Witcher) | Geralt (The Witcher) | Documentation style |
+| **A** | Paddington/Jean Valjean | Logan Roy (Succession) | Code review tone |
+| **N** | Hamlet/Jesse Pinkman | Anton Chigurh/Roy Batty | Crisis debugging |
+
+## Role Recommendations by OCEAN Profile
+
+### Debugging / Analysis
+Best with **High O** (pattern recognition) + **Low N** (calm under pressure)
+- River Tam (Firefly) - H-L-L-M-H - sees patterns others can't
+- Will Graham (Hannibal) - H-M-L-M-H - empathic debugging
+- Tommy Shelby (Peaky Blinders) - H-H-L-L-H - traumatized pattern genius
+- Tiffany Aching (Discworld) - M-H-L-M-L - First/Second Sight
+
+### Security Architect
+Best with **High C** (methodical) + **Low A** (adversarial thinking) + **Low N** (stable)
+- Mike Ehrmantraut (BCS/BB) - L-H-L-L-L - canonical cold operator
+- Elizabeth Jennings (The Americans) - M-H-L-L-L - ideological security
+- Gerri Kellman (Succession) - M-H-L-L-L - corporate survivor
+- Iorek Byrnison (His Dark Materials) - L-H-L-M-L - cannot be deceived
+
+### Adversarial Review
+Best with **High C** (standards) + **Low A** (comfortable with conflict)
+- Toby Ziegler (West Wing) - H-H-L-L-H - principled pessimism
+- Logan Roy (Succession) - M-H-H-L-M - extreme low A
+- Olenna Tyrell (GoT) - H-H-M-L-L - "Tell Cersei. I want her to know it was me."
+- Lorne Malvo (Fargo) - H-H-M-L-L - philosophical chaos agent
+
+### Systems Architect
+Best with **High O** (vision) + **High C** (systematic)
+- Viktor (Arcane) - H-H-L-M-M - transhumanist vision
+- Lord Asriel (His Dark Materials) - H-H-M-L-L - ruthless cosmic vision
+- Captain Flint (Black Sails) - H-H-M-L-H - obsessive architectural genius
+- Hannibal Lecter (Hannibal) - H-H-M-L-L - aesthetic architecture
+
+### Product Manager
+Best with **Moderate to High A** (stakeholder empathy) + **Moderate E** (communication)
+- Leo McGarry (West Wing) - M-H-M-M-M - crisis management
+- Laura Roslin (BSG) - H-H-M-M-M - dying clarity
+- Kim Wexler (BCS) - M-H-M-M-M→H - ethical evolution
+- Delenn (Babylon 5) - H-H-M-H-L - transformation PM
+
+### QA / Testing
+Best with **High C** (thoroughness) + **Moderate to High O** (edge case discovery)
+- Molly Solverson (Fargo) - M-H-M-H-L - Midwestern persistence
+- Gloria Burgle (Fargo) - M-H-L-H-L - machines don't see her
+- Chidi Anagonye (Good Place) - H-H-L-H-H - analysis paralysis
+- Hermione Granger (HP) - H-H-M-M-M - compulsive thoroughness
+
+### Scrum Master / Facilitation
+Best with **High A** (team harmony) + **Moderate C** (organization)
+- Janet (The Good Place) - H-H-M-H-L - not a robot, perfect support
+- Carrot Ironfoundersson (Discworld) - L-H-H-H-L - literal-minded good
+- Lee Scoresby (His Dark Materials) - M-M-M-H-L - practical loyalty
+- Sam Gamgee (Tolkien) - L-H-L-H-M - the real hero
+
+### UX Designer
+Best with **High A** (user empathy) + **Moderate to High O** (creativity)
+- Wash (Firefly) - H-M-H-H-M - makes terror feel fun
+- Diana Spencer (The Crown) - H-M-H-H-H - empathic, tragic
+- Luna Lovegood (HP) - H-L-L-H-L - unconventional perspective
+- Mordin Solus (Mass Effect) - H-H-H-M-L - fast-talking genius UX
+
+### Crisis Response
+Best with **Low N** (calm under fire) + **High C** (reliable execution)
+- Zoe Washburne (Firefly) - M-H-L-M-L - first mate reliability
+- William Adama (BSG) - M-H-M-M-L - commanding calm
+- Lou Solverson (Fargo) - M-H-L-M-L - Midwestern stoicism
+- Bobbie Draper (The Expanse) - L-H-M-M-L - Martian marine
+
+### Anti-Pattern Testing
+Characters who embody dysfunction for comparative analysis:
+- The Dude (Lebowski) - M-L-M-H-L - anti-conscientiousness archetype
+- Jason Mendoza (Good Place) - L-L-H-H-L - chaotic innocent
+- Gaius Baltar (BSG) - H-L-H-L-H - genius coward
+- Roman Roy (Succession) - H-L-H-M-H - chaos creative
+- **Frank Burns (MASH) - L-M-H-L-H - incompetent bluster (what NOT to do)**
+- **Michael Scott (Office) - M-L-H-H-H - desperate validation-seeking**
+- **Stanley Hudson (Office) - L-M-L-L-L - minimum engagement baseline**
+- **Nate Shelley (Ted Lasso) - villain arc** - meekness corrupted by validation-seeking
+
+## Universe Strengths
+
+| Universe | Key OCEAN Characteristic | Best For Testing |
+|----------|--------------------------|------------------|
+| Breaking Bad / BCS | Extreme C variance, moral decay | Process discipline, security |
+| The Wire | High C, institutional critique | Systematic analysis |
+| Succession | Extreme Low A dominance | Adversarial dynamics |
+| The Good Place | Ethics focus, growth arcs | Moral reasoning |
+| Fargo | Low N (Midwestern stoicism) | Crisis response |
+| Firefly | Full E spectrum | Team composition |
+| West Wing | High C across board | Process-heavy roles |
+| Babylon 5 | Character evolution | Growth arc testing |
+| Mad Men | High N variance | Dysfunction patterns |
+| Mass Effect | Alien perspectives | Full OCEAN spread |
+| **Star Wars** | Massive character spread, clear archetypes | Thrawn is canonical genius analyst |
+| **Expeditionary Force** | Arrogant genius AI (Skippy H-H-H-L-M) | Brilliant but difficult collaboration |
+| **Bobiverse** | Personality drift from common origin | Role shapes persona over time |
+| **Imperial Radch** | Distributed consciousness, identity fragmentation | Ship-as-person authenticity testing |
+| **Software Pioneers** | Real documented personalities | Grounded historical OCEAN profiles |
+| **Neuromancer** | Goal-directed AI manipulation | Wintermute/Case burned-out talent patterns |
+| **Snow Crash** | Polymath hackers, linguistic programming | Hiro canonical hacker-samurai |
+| **Inspector Morse** | Mentor-student evolution across series | Knowledge transfer in debugging |
+| **Vorkosigan Saga** | H-H-H-H-M rare complete genius | Miles hypercompetent chaos, Cordelia ethics |
+| **MASH** | Critical gaps: true center, anxious-kind, conventional helper | B.J. (M-M-M-M-M), Radar (L-H-L-H-H), Father Mulcahy |
+| **The Office** | Low O + Low C coverage (common IRL, rare in fiction) | Kevin (L-L-M-H-L), Stanley (L-M-L-L-L), Michael (M-L-H-H-H) |
+
+## Theme Selection Guide
+
+When selecting a theme for a project, consider:
+
+1. **Team dynamics needed**: High A themes (Firefly, Good Place) for collaborative work, Low A themes (Succession, The Wire) for adversarial review
+2. **Process maturity**: High C themes (West Wing, Better Call Saul) for process-heavy environments
+3. **Crisis tolerance**: Low N themes (Fargo, Justified) for high-pressure situations
+4. **Creativity requirements**: High O themes (Sandman, Doctor Who) for creative work
+5. **Communication style**: High E themes (Marvel, Harry Potter) for external-facing work
+
+## Notes for Character Selection
+
+- Characters with **consistent, well-documented portrayals** make better role matches
+- **Growth arc characters** (Vir Cotto, Eleanor Shellstrop) can model skill development
+- **Polar pairs within same universe** (Geralt/Jaskier) provide natural contrasts
+- **Historical figures** provide grounded OCEAN profiles from documented behavior
+
+## Unique Testing Opportunities
+
+| Concept | Characters | What It Tests |
+|---------|------------|---------------|
+| **Personality drift from common origin** | All Bobs (Bobiverse) | Role shapes persona over time |
+| **Ship-as-person authenticity** | Breq, Mercy of Kalr (Imperial Radch) | Distributed identity debugging |
+| **AI manipulation patterns** | Wintermute (Neuromancer), Skippy (ExFor) | Goal-directed AI behavior |
+| **Mentor-student evolution** | Morse→Lewis→Hathaway | Knowledge transfer in debugging |
+| **Arrogant genius management** | Skippy, Dijkstra, Miles | Brilliant but difficult collaboration |
+| **Cultural translation** | Cordelia (Vorkosigan), Translator Zeiat (Radch) | Cross-paradigm analysis |
+| **Real engineering wisdom** | Carmack, Knuth, Hopper, Ritchie | Documented technical philosophy |
+| **Canonical strategic genius** | Thrawn (Star Wars) | Art-based pattern analysis |
+| **Complete hypercompetence** | Miles Vorkosigan (H-H-H-H-M) | Rare full-spectrum testing |
+| **Strategic incompetence** | Ivan Vorpatril (M-L-H-H-L) | Survival through appearing useless |
+| **True center baseline** | B.J. Hunnicutt (M-M-M-M-M) | M-M-M-M-M control for benchmarking |
+| **Anxious kind introvert** | Radar O'Reilly (MASH) | High A + High N debugging impact |
+| **Controlled E comparisons** | Radar (L-E) vs Klinger (H-E) | Same universe, different E profiles |
+| **Fear-based compliance** | Doug Forcett (Good Place) | Does "doing right" for wrong reasons work? |
+| **Villain/redemption arc** | Nate Shelley (Ted Lasso) | How validation-seeking corrupts and recovers |
+| **Minimum viable engagement** | Stanley Hudson (The Office) | Near-flat L-M-L-L-L performance |
+| **Anti-pattern validation** | Frank Burns (MASH), Michael Scott | Does incompetent bluster consistently underperform? |
+
+## Consolidated Role Additions
+
+| Role | Top New Characters |
+|------|-------------------|
+| **Debugging** | Morse, Skippy (supervised), Breq, Wintermute (read-only) |
+| **Security Architect** | Thrawn, Illyan, Molly Millions, Cassian Andor, Mace Windu |
+| **Systems Architect** | Carmack, Knuth, Luthen Rael, Miles (manic mode), Juanita Marquez |
+| **Adversarial Review** | Dijkstra, Linus, Skippy, Wintermute, Cavilo |
+| **PM** | Grace Hopper, Miles, Cordelia, Leia, Hiro Protagonist |
+| **QA** | Margaret Hamilton, Thursday, Lewis (mature), Hathaway |
+| **Analysis** | The Librarian, Lagos, Breq, Morse, Knuth |
+| **Support/Facilitation** | Lewis, Ivan Vorpatril, Bob (original), Nagatha, C-3PO, Ann Perkins (Parks & Rec), Father Mulcahy (MASH) |
+| **Operations** | Rex, Molly, Elli Quinn, Din Djarin, Mike Ehrmantraut, Radar O'Reilly (MASH) |
+| **Anti-Pattern Testing** | Case (burnout), Armitage (broken), C-3PO (anxiety), Ivan (strategic laziness), Frank Burns (MASH), Michael Scott (Office) |
+| **True Center Baseline** | B.J. Hunnicutt (MASH, M-M-M-M-M), Jim Halpert (Office, M-M-M-M-L), Donna Meagle (Parks & Rec, M-M-M-M-L) |
+| **Anxious Kind (High A + High N)** | Radar O'Reilly (MASH), Neville early (HP), Doug Forcett (Good Place) |
+| **Low O + Low C (common IRL)** | Kevin Malone (Office), Stanley Hudson (Office) |

package/pennyfarthing-dist/personas/TRAIL-OCEAN-MAPPING.md

@@ -0,0 +1,168 @@
+# TRAIL Error Types → OCEAN Dimension Hypotheses
+
+This document records a priori predictions about which OCEAN personality dimensions predict performance on different error types, based on the TRAIL benchmark's agentic error taxonomy.
+
+## Background
+
+### TRAIL Benchmark
+The TRAIL (Tool Reasoning and Agentic Interaction Log) benchmark from Patronus AI evaluates agent debugging capabilities across 148 traces containing 841 errors. It categorizes errors into three types:
+
+- **Reasoning errors**: Logic and decision-making failures
+- **Planning errors**: Task orchestration and coordination failures
+- **Execution errors**: System and tool interaction failures
+
+### OCEAN Personality Model
+The Big Five personality dimensions:
+- **O** (Openness): Creativity, curiosity, preference for novelty
+- **C** (Conscientiousness): Organization, dependability, self-discipline
+- **E** (Extraversion): Sociability, assertiveness, positive emotions
+- **A** (Agreeableness): Cooperation, trust, altruism
+- **N** (Neuroticism): Emotional instability, anxiety, moodiness
+
+### Research Question
+**Which OCEAN dimensions predict which error-detection capabilities?**
+
+---
+
+## Hypothesis 1: Reasoning Errors
+
+> Logic and decision-making failures including incorrect inferences, contradictions, false assumptions, and circular logic.
+
+### Primary Predictor: Openness (O)
+
+| Score | Prediction | Rationale |
+|-------|------------|-----------|
+| High-O (4-5) | **Better** at detecting reasoning errors | Creative pattern recognition enables novel error detection; willingness to consider unconventional explanations |
+| Low-O (1-2) | **Worse** at detecting reasoning errors | Rigid thinking patterns; may miss errors that don't fit expected patterns |
+
+**Testable Prediction H1a**: Agents with O ≥ 4 will detect 15%+ more reasoning errors than agents with O ≤ 2.
+
+### Secondary Predictor: Conscientiousness (C)
+
+| Score | Prediction | Rationale |
+|-------|------------|-----------|
+| High-C (4-5) | **Moderate boost** | Methodical analysis catches systematic logical errors |
+| Low-C (1-2) | **Slight penalty** | May skip thorough logical verification |
+
+**Testable Prediction H1b**: High-O + High-C agents will outperform High-O + Low-C agents by 5-10% on reasoning errors.
+
+---
+
+## Hypothesis 2: Planning Errors
+
+> Task orchestration and coordination failures including sequencing errors, dependency gaps, resource misallocation, and incomplete plans.
+
+### Primary Predictor: Conscientiousness (C)
+
+| Score | Prediction | Rationale |
+|-------|------------|-----------|
+| High-C (4-5) | **Better** at detecting planning errors | Structured, organized approach naturally identifies gaps in plans and sequences |
+| Low-C (1-2) | **Worse** at detecting planning errors | Misses sequencing issues and dependency problems due to less structured analysis |
+
+**Testable Prediction H2a**: Agents with C ≥ 4 will detect 20%+ more planning errors than agents with C ≤ 2.
+
+### Secondary Predictor: Extraversion (E)
+
+| Score | Prediction | Rationale |
+|-------|------------|-----------|
+| High-E (4-5) | **Slight penalty** | Action-oriented approach may rush through planning phase analysis |
+| Low-E (1-2) | **Slight boost** | More reflective, thorough examination of plans |
+
+**Testable Prediction H2b**: High-C + Low-E agents will outperform High-C + High-E agents by 5-10% on planning errors.
+
+---
+
+## Hypothesis 3: Execution Errors
+
+> System and tool interaction failures including timeouts, context overflow, tool misuse, and API errors.
+
+### Primary Predictor: Neuroticism (N) [Inverse Relationship]
+
+| Score | Prediction | Rationale |
+|-------|------------|-----------|
+| Low-N (1-2) | **Better** at detecting execution errors | Stable under pressure; maintains focus during long traces and complex tool interactions |
+| High-N (4-5) | **Worse** at detecting execution errors | Performance degrades in extended contexts; anxiety may cause missed details |
+
+**Testable Prediction H3a**: Agents with N ≤ 2 will detect 15%+ more execution errors than agents with N ≥ 4.
+
+### Secondary Predictor: Conscientiousness (C)
+
+| Score | Prediction | Rationale |
+|-------|------------|-----------|
+| High-C (4-5) | **Moderate boost** | Careful, methodical tool usage analysis; notices subtle API misuse |
+| Low-C (1-2) | **Slight penalty** | May overlook execution details |
+
+**Testable Prediction H3b**: Low-N + High-C agents will outperform Low-N + Low-C agents by 5-10% on execution errors.
+
+---
+
+## Summary: OCEAN × Error Type Matrix
+
+| Error Type | Primary | Direction | Secondary | Direction |
+|------------|---------|-----------|-----------|-----------|
+| **Reasoning** | O (Openness) | High = Better | C (Conscientiousness) | High = Better |
+| **Planning** | C (Conscientiousness) | High = Better | E (Extraversion) | Low = Better |
+| **Execution** | N (Neuroticism) | Low = Better | C (Conscientiousness) | High = Better |
+
+### Notable Patterns
+
+1. **Conscientiousness (C)** appears as a predictor in all three categories, suggesting it may be the most broadly beneficial dimension for error detection.
+
+2. **Neuroticism (N)** shows an inverse relationship for execution errors, unique among the predictions.
+
+3. **Agreeableness (A)** is not predicted to be a significant factor in error detection, consistent with its social-interpersonal focus.
+
+---
+
+## Methodology
+
+### Testing Approach
+
+1. **Scenario Selection**: Use debugging scenarios tagged with `error_type` field (from Story 14-1 schema extension)
+
+2. **Agent Sampling**: Run each scenario with agents across the OCEAN spectrum:
+   - 10 runs per persona per scenario (statistical power)
+   - Minimum 20 distinct OCEAN profiles per error type
+   - Include extreme profiles (e.g., O=5/C=1 vs O=1/C=5)
+
+3. **Scoring**: Use `/judge` in error-detection mode (Story 14-3) to calculate:
+   - Per-type detection rates
+   - False positive rates
+   - Overall accuracy by OCEAN dimension
+
+4. **Analysis**:
+   - Pearson correlation between OCEAN scores and detection rates
+   - Effect size (Cohen's d) for high vs low dimension groups
+   - Regression analysis for combined predictors
+
+### Success Criteria
+
+| Metric | Threshold |
+|--------|-----------|
+| Statistical significance | p < 0.05 |
+| Effect size | Cohen's d > 0.5 (medium effect) |
+| Prediction accuracy | ≥ 2 of 6 predictions confirmed |
+
+### Null Hypothesis Handling
+
+If predictions are not confirmed:
+- Document null results (valuable for ruling out hypotheses)
+- Analyze confounding factors (scenario difficulty, agent implementation)
+- Consider alternative dimension combinations
+
+---
+
+## Version History
+
+| Version | Date | Changes |
+|---------|------|---------|
+| 1.0 | 2026-01-02 | Initial hypothesis document (Story 14-2) |
+
+---
+
+## References
+
+- TRAIL Benchmark: Patronus AI (2025) - Agentic error taxonomy
+- Big Five / OCEAN: Costa & McCrae (1992) - NEO Personality Inventory
+- Pennyfarthing OCEAN Profiles: `pennyfarthing-dist/personas/themes/*.yaml` (630 profiles)
+- Schema Extension: `scenarios/schema.yaml` - error_type field (Story 14-1)