The question
Can a human-directed, AI-assisted workflow produce and evolve a useful, maintainable public product?
The EpochLex Experiment · Phase I
Building a dictionary with artificial intelligence.
A documented case study in human-directed, AI-assisted product development—examining what happened as EpochLex grew from a small dictionary concept into a working public reference product.
The experiment in one minute
From an idea to an observed development system.
The project began as a dictionary, not as a software benchmark. The experiment emerged as AI became part of the design, implementation, research, documentation, testing, and maintenance workflow.
The approach
Build a real product while keeping product intent, architecture, validation, and publication under human control.
What happened
EpochLex became a 365-entry reference work with provenance, dedicated pages, a PWA, contributor infrastructure, and documented standards.
What it suggests
AI shortened the path from intent to implementation, while judgment, verification, and process became more important.
From a question to a working system
It did more than grow the term count.
The September 18 snapshot records different kinds of activity. The report explicitly does not combine them into a single productivity score.
August 28, 202627published entries
September 18, 2026365published entries
98pull requests audited
97merged pull requests
30identified issues
183researched provenance records
Search & browse
Searchable dictionary, A–Z navigation, categories, and list/grid discovery.
Dedicated pages
Stable term pages with definitions, history, sources, status, and related discovery.
Provenance
Structured records separating meaning evidence from origin and history claims.
Pronunciation
Written and audible pronunciation with explicit overrides where needed.
PWA
An installable, app-like experience layered over the same canonical static site.
Word of the Day
A daily discovery experience built from the canonical dictionary dataset.
Contributor surface
Issues, contribution guidance, Discussions, and open-source documentation.
QA infrastructure
Validation guidance, testing expectations, accessibility, and publishing checks.
Final report: Version 1.0 contains the complete 75-page case study, finalized figures, evidence notes, appendices, and methodology details.
Key findings
What the first phase established.
These statements summarize the current report without converting its observations into claims the evidence does not support.
01
AI reduced the distance between product intent and working implementation.
Features, migrations, content expansions, revisions, and debugging could reach an evaluable state quickly enough that working software often became part of the design and requirements process.
02
Speed shifted where effort and risk accumulated.
As implementation became easier to produce, research, verification, accumulated project knowledge, and the judgment required to decide what should be built became increasingly consequential.
03
Ambiguity became part of the experiment.
Unexpected output did not always mean bad implementation. In several cases it exposed product assumptions or requirements that had never been made explicit.
04
Human direction remained consequential throughout.
Identity, architecture, editorial policy, taxonomy, source evaluation, interface quality, prioritization, ambiguity resolution, and release judgment remained human decisions.
05
Conventional development discipline became more explicit, not less.
Architecture, source control, QA, documentation, issue tracking, validation, standards, and human review expanded as the project and the AI-assisted workflow became more capable.
06
Failure was not one category—and AI assistance was not autonomy.
Problems included implementation defects, cache and deployment state, documentation drift, incomplete repository context, and unstated human assumptions. The record does not support describing EpochLex as an autonomous AI-created project.
AI assistance × human judgment
Different kinds of work. A shared outcome.
The repository phrase “Vibe coded · Human-directed · AI-assisted · Human-reviewed” describes the development model more accurately than an “AI vs. human” framing.
Human direction
- Product intent and priorities
- Architecture and constraints
- Editorial policy and taxonomy
- Risk and acceptance boundaries
AI assistance
- Product exploration and drafting
- Implementation and refactoring
- Debugging and documentation
- Research organization and QA assistance
Human review
- Evaluate actual behavior
- Judge sources and evidence
- Resolve ambiguity
- Accept, reject, revise, or publish
Review feeds the next iteration: intent → assistance → artifact → evaluation → decision → revised intent.
Case studies
Where the lessons became visible.
The final Version 1.0 report contains ten case studies. This overview gives each a compact introduction; the report preserves the complete context, evidence notes, and findings.
01
From 27 Terms to a Working Reference Product
How a modest dictionary concept became a functioning product and a platform for further questions.
02
Provenance Changed the Nature of the Project
How asking where terms came from introduced a deeper research and evidence model.
03
Mobile Development and the Difference Between Functional and Correct
Why technically working output could still fail the intended product experience.
04
The AILex to EpochLex Rename
Brand migration as a cross-system change touching identity, assets, metadata, and continuity.
05
Open Source Required More Than Publishing the Code
How contribution surfaces, documentation, issues, and community structure became product work.
06
Rapid Corpus Growth and the Creation of Research Debt
How publication throughput created a downstream provenance-research bottleneck.
07
Word of the Day, Push Notifications, and Architectural Restraint
How new feature ideas were evaluated against the intentionally static architecture.
08
A Stale Pull Request and the Introduction of Codex
How repository state and tool context became part of the development system itself.
09
Strong Automated Validation Did Not Eliminate Human Testing
Why automated checks and actual device/browser evaluation remained complementary.
10
Ambiguity as Part of the Experiment
When ambiguity must be closed—and when observing it can reveal something useful.
How it evolved
A project, step by step.
This overview emphasizes turning points rather than reproducing a commit-by-commit changelog.
Aug 28
The dictionary begins
A 27-entry implementation turns the idea into something concrete enough to evaluate.
Early phase
The product expands
Search, categories, dedicated pages, and rapid corpus growth create new product questions.
Research
Provenance becomes a system
Meaning, origin, history, sources, and research status become distinct concerns.
Open source
The community surface appears
Contribution guidance, issues, Discussions, and documentation become part of the product.
Maturation
Safeguards and quality become explicit
QA, validation, accessibility, maintenance, and architectural constraints become more formal.
Sep 17
Codex enters the existing workflow
A repository-oriented implementation agent creates a natural boundary for Phase II.
Sep 18
The first reporting snapshot
The 365-entry snapshot becomes the baseline documented by the report.
What this experiment cannot tell us
Important boundaries around the evidence.
The report is a documented case study, not a controlled benchmark of AI against conventional software development.
No matched control project exists, so the report does not establish a causal productivity multiplier.
Human labor was not tracked systematically, preventing credible labor-savings or hourly-productivity calculations.
No lifetime token total is reported in Version 1.0 because the retained evidence does not support a defensible reconstruction.
The exact underlying AI model is not verified for every historical interaction.
Repository history does not preserve every rejected idea, informal test, design reaction, or conversation.
One static, public reference product should not automatically be generalized to very different risk profiles or teams.
Why this matters beyond EpochLex
A case to compare against—not a universal result.
EpochLex does not prove that every AI-assisted project will become faster, cheaper, or better. It provides a documented example of where the balance of work can change when one person can move rapidly between design, implementation, research, testing, and documentation with AI assistance.
Faster production can move the bottleneck.
When the cost of turning an idea into something testable drops, the constraint can move downstream. Review, validation, research, maintenance, and the judgment required to decide what should be accepted may become more important rather than less.
Context becomes infrastructure.
Several unexpected outcomes were not simply AI coding failures. They exposed missing requirements, stale project information, conflicting instructions, environmental behavior, or decisions that had never been made explicit. Documentation, standards, examples, and architecture increasingly shaped what the AI could reliably do.
Professional judgment becomes more visible.
AI increased production capacity without deciding what should be built, whether evidence was credible, whether an experience felt right, or whether a change was ready to ship. The scarce skill may increasingly be knowing what good looks like and accepting responsibility for the result.
A familiar pattern in AEC
A faster rendering engine does not decide whether a design is appropriate. Dynamo does not decide what a BIM standard should be. Revit automation does not determine whether a workflow is safe for every project condition. AI-assisted development changes production capacity in a similar way: it can reduce the friction of creating a solution without reducing the responsibility of understanding the problem, defining the standard, testing the result, and deciding whether it should be used.
What changes next
The experiment continues with better measurement.
The September 18 report establishes a baseline rather than a final assessment. The Codex phase begins a deliberately more prospective observation period.
Measure prospectively
Track activity intentionally rather than reconstructing important measures after the fact.
Observe ambiguity
Distinguish high-risk ambiguity that must be closed from lower-risk ambiguity that may reveal reasoning behavior.
Preserve boundaries
Keep architecture, source-of-truth documentation, validation expectations, and human decision points explicit.
Compare phases
Use the first-phase snapshot as the baseline for evaluating changes in workflow, rework, QA, and tool behavior.
Go deeper
The complete record.
This page highlights the major observations. Version 1.0 is the finalized 75-page case study, including the full narrative, finalized figures, evidence notes, quantitative analysis, limitations, and appendices.
Designed Research Case StudyThe EpochLex
ExperimentBuilding a Dictionary of the AI Era through
Human-Directed, AI-Assisted DevelopmentReport · Version 1.0 · September 2026
ExperimentBuilding a Dictionary of the AI Era through
Human-Directed, AI-Assisted DevelopmentReport · Version 1.0 · September 2026