Co-Creation Audit Trail

Inspiration

Long AI collaboration sessions can produce useful ideas, revisions, and decisions—but the most important human choices often disappear inside the conversation.

While working on a long-form creative project with multiple AI tools, I repeatedly faced the same problem: I could remember that an important decision had been made, but not where it had been made, what evidence supported it, or whether it came from me or from the AI.

I wanted a way to preserve human agency in AI-assisted creation—not by recording everything, but by identifying meaningful turning points, verifying their evidence, and allowing the human to decide what becomes part of the official record.

What it does

Co-Creation Audit Trail turns a long human–AI collaboration log into a reviewable decision history.

The user pastes a conversation log, and the application:

1. Uses AI to extract candidate turning points.
2. Returns a verbatim evidence quote for each candidate.
3. Verifies in code whether the full quote exists exactly in the original log.
4. Shows the matching line numbers.
5. Classifies the result as explicit, inferred, or unsupported.
6. Lets the human approve, reject, or edit each item.
7. Exports only the approved decisions as Markdown.

The AI proposes the record, but the human decides what becomes official.

How I built it

I built the application as a lightweight web prototype using GPT-5.6 and Codex.

GPT-5.6 is used to analyze the collaboration log and identify possible turning points. The model is instructed to provide a verbatim evidence quote for every candidate.

Instead of trusting the model's claim, the application compares the returned quote against the original source text in code. An item is marked as an exact match only when the complete quote appears as a continuous string in the source log.

Codex helped me inspect and improve the implementation, including:

• exact contiguous quote verification
• source line-number detection
• explicit, inferred, and unsupported classifications
• human approval and rejection controls
• Markdown export
• support for both English and Japanese logs

Challenges

The main challenge was separating AI interpretation from verifiable evidence.

An AI can produce a convincing explanation of why a decision was made, even when the supporting quote is incomplete, paraphrased, or absent. I therefore designed the system so that semantic interpretation and source verification are separate steps.

Another challenge was defining what counts as a meaningful turning point. Not every revision should become part of the permanent record. The application therefore treats extracted items as candidates, not facts, and requires human review before export.

What I learned

I learned that provenance in human–AI collaboration is not only about tracing model outputs. It is also about preserving the human's decisions.

As AI becomes more involved in creative and knowledge work, it becomes easier to produce more—and harder to remember who decided what, why it changed, and which choices were ultimately accepted.

A useful collaboration record should therefore be:

• selective rather than exhaustive
• grounded in the original source
• transparent about inference
• controlled by the human

What's next

Future versions could support multiple uploaded logs, decision timelines across sessions, links between related turning points, and comparison between AI suggestions and final human decisions.

The long-term goal is to create a practical provenance layer for human–AI collaboration—one that helps people use AI without losing sight of their own judgment.

Built With

• api
• codex
• css
• gpt-5.6
• next.js
• openai
• react
• tailwind
• typescript