docs: reference carbontracker.info in climate-nudging research

CHANGELOG.md

@@ -6,6 +6,9 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+### Added — Carbontracker reference in Climate Awareness Nudging
+- Reference [carbontracker.info](https://carbontracker.info/) in `co2-research.md` (new "Measurement tools" subsection + sources entry), `integration.md` (API/proxy layer), and `index.md` ("What makes it hard")
+
 ### Fixed — Project "Last edited" dates on the deployed site
 - GitHub Actions workflows (`deploy.yml`, `verify_build.yml`) now check out the full git history (`fetch-depth: 0`) so the `projectDates` data loader can read the per-project last-commit timestamp. Previously the shallow clone caused every project card to show the same date.
 

docs/projects/climate-nudging/co2-research.md

@@ -30,6 +30,10 @@ The single most important finding: **estimates in the literature disagree by up
 
 Ritchie calculated that 10 ChatGPT queries/day for a UK resident equals ~0.03% of daily electricity use. Abstaining from 50,000 queries (~14 years at 10/day) saves less CO2 than everyday actions like recycling. However, she distinguishes individual footprint from aggregate industry impact — the concern is systemic growth, not individual queries.
 
+### Measurement tools
+
+[**Carbontracker**](https://carbontracker.info/) (Anthony, Kanding & Selvan, 2020) is the most established open-source tool for measuring and predicting the carbon footprint of ML workloads. It samples hardware power draw and combines it with regional grid carbon intensity to produce real-time and predicted-total emissions estimates, with low overhead via separate threads. It supports Intel CPUs, NVIDIA GPUs and Apple silicon, ships as both a CLI and a Python library, and includes log-parsing helpers for third-party integration. Originally created at the University of Copenhagen and now maintained there with EU Horizon Europe support, the original paper has been cited 470+ times. Carbontracker is aimed at training workloads; applying it to inference is straightforward in principle but, as section 5 notes, requires hosting providers to expose the underlying telemetry.
+
 ---
 
 ## 2. Per-Query Estimates by Model
@@ -184,3 +188,4 @@ Google claims 33x reduction in energy per median prompt over 12 months. But tota
 10. "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint" (2025). [arXiv:2505.09598](https://arxiv.org/html/2505.09598v1)
 11. "Why transparency matters for sustainable data centers" (2025). iScience. [Paper](https://www.cell.com/iscience/fulltext/S2589-0042(25)01966-2)
 12. Federation of American Scientists. "Measuring AI's Energy Footprint." [Report](https://fas.org/publication/measuring-and-standardizing-ais-energy-footprint/)
+13. Anthony, L. F. W., Kanding, B. & Selvan, R. (2020). "Carbontracker: Tracking and Predicting the Carbon Footprint of Training Deep Learning Models." [arXiv:2007.03051](https://arxiv.org/abs/2007.03051) · [Project site](https://carbontracker.info/)

docs/projects/climate-nudging/index.md

@@ -55,6 +55,7 @@ See the complete [CO2 Research](co2-research.md) document for detailed findings
 - No major AI provider publishes per-query energy data
 - Reasoning models can use 70x+ more energy than simple models
 - Efficiency improves, but total demand grows faster (Jevons' Paradox)
+- Open-source measurement tools exist for ML training (e.g. [carbontracker](https://carbontracker.info/), University of Copenhagen) but are not adopted by major hosting providers for inference, so per-query numbers remain estimated rather than measured
 
 ### Honest framing
 

docs/projects/climate-nudging/integration.md

@@ -79,6 +79,8 @@ For organizations running their own AI gateway (like AarhusAI), a proxy layer ca
 4. **Apply rate-based nudges** — if a user exceeds a threshold, add a gentle note to the next response
 5. **Detect organizational patterns** — like AI-generated-text-to-AI-summarize chains
 
+For self-hosted inference where the gateway sits on the same machines as the GPUs, [carbontracker](https://carbontracker.info/) is a natural fit for the measurement layer: it samples real hardware power draw and combines it with regional grid carbon intensity, replacing bottom-up token-based estimates with measured energy. Wrap inference calls with the Python library or run the CLI alongside the workload, then attach the resulting per-request energy/CO2 to the response headers proposed in [CO2 Research § 5](co2-research.md#5-guidance-to-hosting-providers). For hosted models (OpenAI, Anthropic, etc.) carbontracker can't reach the hardware, so estimates remain the only option until providers expose telemetry themselves.
+
 ### Shared components
 
 Regardless of frontend, these components are reusable: