The data.

What an AI interaction holds, and what can be read from it.

The data set

The interaction records the process, not just the product.

An AI interaction is a sequence of turns between a person and a model, usually in a task whose purpose is to produce something else: code, an analysis, a draft, a decision. The deliverable that remains at the end is the final state of that sequence. The interaction is the path by which it was reached.

Deliberation that previously took place in meetings or in drafts, or was never written down, now takes place in the exchange with the model, where it is recorded in full. A document, a ticket, or a commit records what was done. The interaction records the deliberation that preceded it, the hypotheses tested, the approaches abandoned, the corrections applied, and therefore answers why.

As more work is produced by people and AI together, the interactions between them become the account of who contributed what. That is the data set we capture.

the task attempt refused correction approach abandoned The work product the final state
The document records the end. The interaction records the path: the attempts refused, the corrections applied, the approaches abandoned.

One example

The chat record, the parser, and attribution.

The chat record

Every AI tool keeps a record of the conversation: the prompts, the answers, their order, and which message answered which. That is what the vendor exports. It has no field for the decision, what the person did with each answer: used it, changed it, refused it. The tool never asks, so it never writes it down.

The HammerCapture parser

The parser reads the chat record and rebuilds it as a sequence of attempts. Each AI answer is one attempt. The person's next words are the context for the decision on that attempt. Where the context says what the person did, the parser derives the decision and writes it onto the record, marked as derived, with the words attached as evidence. Where the context says nothing, it writes not known. It does not guess.

Attribution

From the attempts and their decisions, HammerCapture computes how much of the work product was the person's. Each decision counts only as far as its evidence allows. Not known decisions form a not known share, which only capture at the moment the person acts can reduce.

The exported file holds the words and their order. The decisions are derived from the person's next words, each marked and carrying its evidence, or stated as not known.
Attribution of the letter, illustrative. Each decision counts as far as its evidence allows. What no evidence closes stays in the not known share, which only capture at the moment of work can reduce.