AtoAxis keeps AI, human and Unattributed changes separate. Coverage tells you how much of the observed change set has evidence instead of turning uncertainty into a confident zero.
Author:Jitwisut Thobut, Founder & developer of AtoAxis. Reviewed: . Status: Invite-only Early Access.
The three attribution buckets
AI: a hook or client explicitly marks the activity as AI-assisted and can provide an agent name and line counts.
Human: the editor reports hand-typed keystrokes or characters and the corresponding measured changes.
Unattributed: the system has observed a change but cannot prove which source produced it.
The buckets are mutually exclusive in the canonical totals. Unattributed is evidence that is missing, not a synonym for human work and not a reason to guess.
Where evidence comes from
AI attribution comes from explicit fields supplied by a supported hook or client, such as an agent name, is_ai_generated flag and line counts. Human attribution uses editor activity fields such as keystrokes and characters. A compatible editor can send time without sending every attribution field, so missing fields remain visible.
Line coverage is calculated from the same canonical totals: attributed AI and human line changes divided by attributed plus Unattributed line changes. The response also exposes the denominators and an explicit status. With no activity, the percentage is unavailable rather than a misleading 0%.
Worked example — not customer data
Illustrative change set: 40 AI + 35 human + 25 Unattributed = 100 lines
Attributed line coverage: (40 + 35) / 100 = 75%
Data coverage is a separate question from the AI share. A result can have a high AI percentage among known lines while still having incomplete evidence for the whole period. Read both the line totals and the coverage status.
How to compare agents
Compare agents only over the same date range, project filters and evidence status.
Check whether a tool has a bundled hook or a provider-specific adapter before expecting automatic attribution.
Do not treat line counts as correctness, maintainability, time saved or business value.
Keep Unattributed visible when a tool reports activity without attribution metadata.
Privacy boundary
Attribution carries labels, identifiers and numeric counts. It does not require sending source code, prompts or AI response text. Provider usage can add token metadata, but the text of the transcript remains outside the usage event.
Review and verification
Use the AI coding tracker page for supported hooks and the dashboard analytics response for the denominators behind a result. Review owner: AtoAxis product team. Last reviewed: 20 September 2026.
How should you read AI attribution and coverage together?
AtoAxis separates AI, human and Unattributed changes. Coverage reports how much of the change set has evidence, so missing evidence never becomes a confident zero.
A 30-second tour recorded from this website: the time-ring overview, the Human + AI split, the AI coding tracker and planned pricing.
Key takeaways
AI evidence:
An agent name, an is_ai_generated flag and line counts from a hook.
Human evidence:
Editor activity such as keystrokes and characters.
Unattributed:
Changes without enough evidence; not a synonym for human work.
Line coverage:
Attributed lines divided by attributed plus Unattributed lines.
How do you compare two agents fairly?
Illustrative example, not customer data: one agent shows 60% of known lines with 90% coverage, another 60% with 40% coverage. The first figure rests on far more evidence. Compare agents only over the same range, with both hooks installed for the whole period.
How to read coding activity metrics
Signal
What it supports
What it cannot prove
Editor heartbeats
Recorded coding activity
All working time
Attributed changes
Reported source of changes
Code quality or time saved
This comparison separates what the data supports from conclusions it cannot establish.
Recommendation
Read the line totals, the coverage status and the denominators before quoting an AI share. Install the supported hook for each agent you use, and treat a gap in coverage as a reason to add instrumentation rather than to guess.
Trade-offs to weigh
On the other hand, a high AI share with low coverage says little. Coverage tells you how much of the period had evidence at all, so read the two numbers together before drawing a conclusion.
“A metric that cannot be established should say so instead of rendering as zero.”
— Jitwisut Thobut, Founder & developer, AtoAxis
Sources and review
Written and reviewed by Jitwisut Thobut, Founder & developer of AtoAxis. Reviewed . Product claims describe our implementation; examples are illustrative, not customer research.
84% of developers use or plan to use AI tools in development, yet 46% distrust the accuracy of AI output and 33% trust it (Stack Overflow Developer Survey 2025). Source.
AtoAxis