Line-level attribution for AI-written code
Most of your codebase was written by a model and nobody recorded which parts
Provenance records which lines came from which model, under whose review, and shows you the day the mix changed. Git-native, automatic, and invisible until somebody asks.
The problem
Review capacity did not grow when writing capacity multiplied
The same four reviewers who handled two hundred human-written lines a day now face two thousand generated ones. Nobody decided to lower the review bar; it lowered itself. And when an acquirer or a regulator asks which parts of this system a human actually read, there is no answer, because nothing was recording.
The insight
The number that matters is not how much AI code you have
Every dashboard reports model-written percentage, which is a vanity metric — high is not bad. The number that matters is the unreviewed model-written surface, and the moment it started growing. That second part is a changepoint question, so it runs through the same detector the rest of this engine uses rather than being eyeballed off a chart nobody opens.
Per-commit authorship and review attribution rolled into a daily model-share series, with CUSUM changepoint detection to date the adoption shift, and unreviewed-line accounting split by model.
How it works
Four steps, no data science team
A git hook and a CI step. No change to how anyone works.
Which model, which prompt, which reviewer, per line, carried through rebases.
Alert when generated code outruns review capacity, dated to the day it started.
A report an acquirer, an auditor or a customer can read.
Who it is for
The VP of Engineering who will be asked the question
Engineering organisations where agents write most of the code, and someone will eventually be asked to account for it. Regulated, acquisition-bound, or selling to enterprises with a security questionnaire.
Pricing
- –Line attribution
- –Model-share series
- –Changepoint detection
- –Review-coverage tracking
- –Unreviewed-surface alerts
- –Per-model breakdown
- –Export
- –Self-hosted
- –Attestation reports
- –SSO
- –Long retention
Competition
What exists, and what it does not do
| Who | What they do | The gap |
|---|---|---|
| GitHub Copilot metrics and vendor dashboards | Report acceptance rates and usage of their own tool. | Single-vendor, aggregate, and about adoption rather than accountability. They cannot tell you what went unreviewed. |
| git blame | Already attributes every line to a committer. | The committer is the human who ran the agent. That is exactly the distinction being lost. |
| DORA and engineering analytics | Measure throughput and review latency. | They measure the process, not the provenance. Neither knows which lines a model wrote. |
| A policy in the handbook | All AI code must be reviewed. | Unenforced and unmeasured, which is how the unreviewed surface grew without anyone deciding to let it. |
Adoption dies on friction: anything requiring a developer to tag their commits will be abandoned in a week, so the attribution has to be fully automatic and that is genuinely hard across rebases, squashes and mixed-authorship lines. The demand is also anticipatory — most teams do not feel this pain yet, and a product sold against a question nobody has been asked is a product sold on fear. It becomes urgent the first time a large acquisition or a regulator makes it urgent, and that timing is outside our control.
Market
Attached to the compliance and engineering-governance budget rather than the developer-tools one
Narrow today, structurally growing. The buyer is companies facing diligence, which is a small share of the market with a high willingness to pay and a hard deadline.