The complete record.

01 · Experiment report

Executive Summary

EpochLex is an open-source, searchable reference work for terminology of the AI era. The project began on August 28, 2026, as a relatively small dictionary concept and developed into a 365-entry public reference work with structured provenance, dedicated term pages, pronunciation, responsive browsing, related-term discovery, a Progressive Web App, Word of the Day, contributor infrastructure, and a documented editorial and architectural model.

The project also became an experiment in AI-assisted development. ChatGPT was used extensively alongside GitHub and repository-based development during the initial phase, assisting with product exploration, implementation, debugging, documentation, research organization, quality assurance, and editorial support. On September 17, Codex was added to the existing workflow as a more repository-oriented implementation agent. The development process remained human-directed and human-reviewed throughout.

The strongest finding in Version 1.0 is that AI materially reduced the distance between product intent and working implementation. That acceleration did not remove the need for planning or engineering discipline. As the project grew, architecture, QA, source control, documentation, validation, editorial standards, and human review became more explicit rather than less important.

The report also treats ambiguity as an experimental variable and distinguishes different classes of failure. It does not interpret repository throughput as a productivity multiplier, labor saving, or proof of superiority over conventional development because no matched control project or reliable historical labor record exists.

Report snapshot

Key Metrics at a Glance

These measures describe different forms of activity at the September 18 cutoff and are not combined into a single productivity score.

Part I

Introduction, Scope, and Methodology

1. Introduction, Scope, and Methodology
2. Purpose of the Case Study
3. Questions Examined
4. Observation Period and Evidence Base
5. Development Environment
6. Quantitative Evidence
7. What This Report Does Not Claim
8. A Note on Ambiguity

The report treats the public repository and current source-controlled project state as the primary record for the reporting cutoff. Development conversations provide supporting context for intent, rejected alternatives, ambiguity, and human feedback. The case study documents one real product rather than a controlled laboratory comparison.

Part II

Reconstructing the Experiment

9. From a Term List to a Working Product
10. Rapid Expansion and the First Regression Cycle
11. Quality Assurance Becomes Part of the Product
12. The AILex to EpochLex Migration
13. The Taxonomy Expands Beyond Terms
14. The PWA and the Boundary of the Static Architecture
15. From a Private Project to an Open-Source Operation
16. Making the Research Pipeline Persistent
17. Scaling the Corpus Changes the Bottleneck
18. Word of the Day and the Question of Architectural Cost
19. The PWA Becomes an Intentional Product Experience
20. Introducing Codex Into an Existing Workflow
21. Maintenance Work Becomes More Explicit
22. What Changed During the First Three Weeks

This part reconstructs how a dictionary project accumulated software architecture, research infrastructure, contributor workflows, QA, and maintenance concerns. The emphasis is on how the system changed—not merely on counting features.

Part III

Case Studies and Findings

23. Case Study One: From 27 Terms to a Working Reference Product
24. Case Study Two: Provenance Changed the Nature of the Project
25. Case Study Three: Mobile Development and the Difference Between Functional and Correct
26. Case Study Four: The AILex to EpochLex Rename
27. Case Study Five: Open Source Required More Than Publishing the Code
28. Case Study Six: Rapid Corpus Growth and the Creation of Research Debt
29. Case Study Seven: Word of the Day, Push Notifications, and Architectural Restraint
30. Case Study Eight: A Stale Pull Request and the Introduction of Codex
31. Case Study Nine: Strong Automated Validation Did Not Eliminate Human Testing
32. Case Study Ten: Ambiguity as Part of the Experiment
33. Cross-Case Findings
34. A Developing Interpretation of “Vibe Coding”
35. The Role of Human Judgment
36. The Emerging Development Model
37. Interim Finding

Version 1.0’s interim interpretation is that AI changed the speed and cost of iteration more clearly than it changed the need for judgment. Working software became part of the thinking process, while the value of that acceleration increasingly depended on the surrounding system’s ability to distinguish useful output from plausible but incomplete output.

Part IV

Quantitative Analysis and Experimental Measures

38. Why Quantitative Measures Matter
39. Repository Activity During the Observation Period
40. Development Was Highly Concentrated in Bursts
41. Corpus Growth
42. Provenance as a Quantifiable Bottleneck
43. Pull Requests as a Measure of Iteration, Not Productivity
44. Rework and Follow-Up Work
45. Human-Detected Versus Automated Findings
46. The Toolchain Timeline
47. Model Changes Are Part of the Experiment
48. Token Utilization
49. Token Counts Are Not Productivity Scores
50. Human Effort Is the Largest Missing Quantitative Measure
51. Measuring the Next Phase Beginning With Codex
52. Tracking Ambiguity as Data
53. Candidate Metrics for the Continuing Experiment
54. What Can Be Quantified Now
55. What the Numbers Tell Us

The quantitative section deliberately separates direct measurements from reconstructed estimates and prospective measures. It does not convert pull-request counts, tokens, or corpus growth into unsupported productivity claims.

Part V

Discussion, Lessons, and Implications

56. The Experiment Was Not Really About Whether AI Could Write Code
57. AI Reduced the Cost of Trying an Idea
58. Faster Implementation Changes the Economics of Iteration
59. The Hammer and the Nail Gun
60. Established Development Practices Did Not Disappear
61. Repository Context Became a Form of Organizational Memory
62. Documentation Is Part of the AI Control System
63. AI Can Accelerate Technical Debt as Easily as Features
64. Human Judgment Became the Scarcer Resource
65. Human Review Is More Than Error Checking
66. Ambiguity Is Shared Between the Human and the AI
67. The Ability to Surface Ambiguity May Become a Model Capability Worth Measuring
68. Product Responsibility Limits Experimental Freedom
69. Architecture Became a Decision Framework
70. Simplicity Has Ongoing Value in AI-Assisted Systems
71. AI-Assisted Research Has Different Risks From AI-Assisted Coding
72. The Meaning of “AI-Built” Requires Precision
73. Tool Changes Complicate Cause and Effect
74. The Work Is Shifting Rather Than Disappearing
75. Implications for Professional Practice
76. Process Knowledge May Outlast Tool Knowledge
77. Vibe Coding and Professional Practice Are Not Necessarily Opposites
78. What the Experiment Suggests So Far
79. A More Mature Statement of the Experiment
80. The Central Lesson

Across these sections, the report argues for process principles that remain useful even as tools change: current repository state over stale planning material, claim-specific evidence, deliberate handling of high-risk ambiguity, explicit invariants for large changes, and validation that states both what was tested and what was not.

Part VI

Conclusion, Limitations, and the Continuing Experiment

81. The September 18 Snapshot
82. What the First Phase Established
83. Limitations of the First Report
84. Limits on Generalization
85. Codex Marks a New Observation Period
86. Measurement Plan for the Next Phase
87. Token Utilization Status
88. Model Timeline Status
89. Contributors and the Expanding Experimental Environment
90. Ambiguity Should Continue to Be Observed Deliberately
91. EpochLex as a Living Experiment
92. Conclusion

The first phase demonstrates AI participation across a broad product-development lifecycle while also documenting consequential human direction. Faster implementation did not remove conventional development practice; as the project grew, architecture, source control, QA, documentation, validation, and human review became more explicit.

Limits carried forward

• No matched control project.
• No systematic historical human-labor tracking.
• No defensible lifetime token total can be reconstructed from the retained evidence set.
• Exact historical model/configuration is left unspecified where it cannot be verified.
• Repository and conversation records are substantial but not forensically complete.
• Results should not automatically be generalized to projects with substantially different risk profiles, architectures, or teams.

Appendices

Evidence and measurement layers

• Appendix A — Evidence Register: claim-to-evidence traceability across repository artifacts, PRs, issues, development conversations, measurements, and analytics.
• Appendix B — Repository Classification Framework: completed PR classification and secondary tags.
• Appendix C — Model and Toolchain Timeline: verified tool phases and interpretive boundaries.
• Appendix D — Token Utilization Reconstruction: staged reconstruction that separates visible counts, modeled context, and direct analytics.
• Appendix E — Next-Phase Measurement Framework: a more deliberate measurement model for the continuing experiment.
• Appendix F — Report Terminology: consistent definitions for experiment-specific language.
• Appendix G — Source Notes and Methodology Details: source hierarchy, evidence handling, citation convention, and measurement labels used in Version 1.0.

Publication

Version 1.0 · September 2026

The finalized report is a 75-page designed research case study covering the August 28–September 17, 2026 observation period with a September 18 repository snapshot.

• The PDF is the authoritative complete Version 1.0 publication.
• This web companion summarizes and navigates the report without strengthening claims beyond the published evidence.
• Later observation periods should be published as new phases or editions rather than silently rewriting this historical baseline.