The five kinds of work
ATTRIBUT doesn't score a session by how many lines changed. Instead, every session you send through the pipeline resolves to a mix of five kinds of work — creation, reduction, repair, understanding, and orchestration — and each kind is sized into expert-hours by its own formula, built from structural signal rather than the content of your code.
Expert-hours are the unit ATTRIBUT uses to value a session: an estimate of how long the work would have taken an experienced engineer to do by hand. A single session is almost never just one kind of work — it is a blend of some or all five, and that blend is what gets sized.
What "structural signal" means
None of the five kinds are read from what your code says or does. Each is read only from its structure — things like line counts, tokens, turns, tool calls, and subagent activity — and no single one of those signals is the value on its own; each kind's formula combines several of them. ATTRIBUT never inspects the meaning or quality of your code to size a session.
The five kinds of work
Kind | What it is | Reads from structure | Sized by |
Creation | Net-new capability — structure and behavior that did not exist before. | Added lines, new files, new symbols | Net-new structure turned into hours, via the creation curve |
Reduction | Deletion and simplification — often the most valuable, least visible work. | Removed lines, deleted files, dead code | Removed structure, weighted — deletion counts as real work |
Repair | Debugging and iteration — making existing code correct. | Edit churn, turn count, test runs | Iteration depth multiplied by the locality of the change |
Understanding | Reading and navigating — the work before the work. | Files opened, searches, tool calls with no edit | Breadth of exploration across the repo |
Orchestration | Coordinating subagents and multi-step plans. | Subagent spawns, tool-call depth, plan steps | Coordination span over the work graph |
For example, in one worked session: creation made up 48% of the expert-hours, repair 22%, understanding 14%, orchestration 9%, and reduction 7%. That mix is specific to that session — a different session might lean almost entirely on repair, or split evenly across all five. There is no fixed weighting; the mix is whatever the session's structure actually shows.
Why five separate formulas instead of one line count
A raw line count treats every change as equal, rewards verbosity, and can't tell a genuine new feature from copy-pasted churn. It also can't see deletions, debugging, or coordination work on their own terms, even though those are often the hardest parts of a session.
Sizing each of the five kinds with its own formula fixes that: a deletion is measured as deletion, not as "negative creation"; a debugging loop is measured by how many iterations it took and how localized the fix was, not by how many lines changed; and coordinating subagents is measured by the shape of that coordination, not folded into whichever kind of edit happened to result.
Once a session's mix of the five kinds is sized into expert-hours, that figure is carried forward into the rest of ATTRIBUT's value calculation — scaled by whether the work landed and priced at your rate.