🌱 Green IT
The factory treats energy, tokens, and accessibility as delivery constraints, not just as a report written afterwards.
No claim without evidence. Every statement on this page is tied to a mechanism in the model and to an audit agent that checks it. Statements that no audit can verify are reworded or removed.
Principle
The cheapest token, build minute, and byte is the one that is never spent. Green IT in the factory has two halves:
- Sober delivery: the factory spends as few tokens and CI minutes as possible to ship a change.
- Sober, accessible product: the software it ships is light, fast, and usable by everyone.
Tokens are a proxy for energy, not a measure of it. The factory reports tokens, retries, CI minutes, and page weight. It does not publish carbon figures it cannot measure.
Sober delivery
| Lever | Mechanism | Where it is defined |
|---|---|---|
| Right-sized model per stage | Each agent declares the model tier suited to its task; frontier reasoning is concentrated in Think | Cost Management |
| Narrow the work before generating | The Think → Plan → Code funnel fixes architecture and file scope before any code is written | Cost Management |
| Small context windows | Instructions are short and scoped to narrow file globs; the coding agent loads only the files of the current step | Persistent Context |
| Never regenerate what exists | Think binds proven assets from the Forge instead of generating them again | Forge |
| Zero-token checks first | Compilers, linters, type checkers, and tests run before any AI review | Deterministic Controls |
| Bounded retries | One atomic step per turn and a hard retry cap stop runaway loops and escalate to a developer | Code |
| Fix once, for every squad | A repeated mistake becomes a context change in Learn, so no squad pays for it twice | Learn |
| Attributed consumption | Tokens and cost are recorded per work item, stage, and squad, so waste is visible and reviewed weekly | Cost Management |
Sober, accessible product
The factory carries non-functional requirements from Plan into Prove, where they are checked like any other acceptance criterion.
| Property | Reference standard | How it is checked |
|---|---|---|
| Accessibility | WCAG 2.2 AA (and the EN 301 549 / RGAA obligations that build on it) | Accessibility rules in E2E tests, plus an on-demand static audit of templates and pages (a11y-audit) |
| Performance | Declared budgets per project: page weight, request count, load metrics | Performance tests (k6) and an on-demand audit of the built output (web-footprint-audit) |
| Eco-design | RGESN (French eco-design framework) and GR491 | On-demand audit (web-footprint-audit): third-party calls, fonts, images, unused JS and CSS, caching, motion |
| Reuse over rebuild | Dependency and Forge leverage | Weekly dependency-leverage and dead-code audits |
Automated tools cover only part of accessibility. Screen-reader and keyboard journeys that need a person are listed as manual-check-needed in the audit report, never reported as passed.
How the statements are kept honest
Three read-only audit agents from the coding-pal catalog are available to any project. Only token-audit is scheduled today (weekly, on coding-pal); the other two run on demand until a project adds them to its audit schedule. They work on any project, not only this one, and use the limits a project declares (size, scope) instead of their defaults.
| Agent | What it verifies | Output |
|---|---|---|
Token Auditor (token-audit) | The harness itself: over-broad scopes, oversized or duplicated instructions, skills, agents, and prompts, loops without a retry cap, LLM steps a script could do. Runs weekly on coding-pal, the source of the harness, so a fix reaches every project | Findings with measured size and estimated tokens |
Accessibility Auditor (a11y-audit) | WCAG 2.2 AA defects with file, line, and rule | Findings plus a manual-check-needed list |
Web Footprint Auditor (web-footprint-audit) | Bytes per page, requests, third-party origins, fonts, images, caching, against the declared budgets | Measured values and budget breaches |
A finding is closed in one of two ways: the control is added, or the statement is reworded to what is true. The audits measure the project; a person checks the page against the results.
Indicators
| Indicator | Direction | Source |
|---|---|---|
| Tokens and cost per merged change | Down | Cost Management |
| Fix-loop iterations and first-pass green rate in Code | Loops down, green rate up | Persistent Context |
| Instruction count and size | Stable | Token Auditor |
| Share of Forge assets reused instead of generated | Up | Forge |
| Accessibility defects open | Down | Accessibility Auditor |
| Page weight and third-party requests vs budget | Within budget | Web Footprint Auditor |
What the factory does not claim
- No carbon or kWh figure: tokens and CI minutes are reported, not converted into emissions.
- No claim that AI-assisted delivery is greener than manual delivery. It is designed to waste less than ad-hoc AI usage, which is a different statement.
- No accessibility conformance statement for a product until its audit and manual checks are complete.
Trade-offs
- Clean CI re-runs repeat checks the agent already ran. They cost compute and are kept because the agent cannot influence their result.
- Frontier reasoning in Think costs more per call. It is kept because one sound decision saves many coding tokens.
- Preview environments consume resources while they exist and are tied to the pull request they validate.
Each trade-off is reviewed with the weekly cost and audit data, and is changed when the evidence says so.