Harpswell Renewables — a place to collect notes and links on climate science and renewable energy, and to share what we find worth reading. We wander from time to time. An experiment in automation using AI with human oversight. No warranties on correctness but we try our best.
What’s here
The recurring feature is a roughly monthly scan of renewable-energy developments in Maine, New Hampshire, Vermont, Québec and New Brunswick. It covers state, provincial and regional matters: legislation and PUC dockets, the grid and the regulated utilities that run it, grid-scale projects at every stage from proposal to commercial operation, major outages, and electrification infrastructure. It also tracks adoption data — how much heat-pump, EV-charging, solar and wind capacity is actually deployed, and how fast that is changing.
The two provinces were added in September 2026. They belong here for a practical reason as much as a geographic one: Québec generates well beyond its own demand and sells the surplus south, and the corridors that carry it into New England are a recurring and often contested story on both sides of the border. New Brunswick sits on the region’s other major intertie.
Figures are reported as two sub-regions — Maine/New Hampshire/Vermont and Québec/New Brunswick — and never added together. American and Canadian statistics come from different agencies with different definitions and vintages, and Québec is large enough that a combined number would mostly be a Québec number wearing a regional label.
Some things are deliberately left out: town, municipal and MRC news, product and vendor coverage, and partisan framing. The test is whether a ratepayer or policymaker elsewhere in the region would have reason to care.
How it’s made
Each scan starts as a research pass run by an AI model with live web search, against a fixed prompt that specifies the region, the time window, what counts as in scope, and the sourcing rules. Every run is saved as a dated file before anything is drawn from it.
That raw output is not what gets published. A person reads it, checks the items that matter, drops what doesn’t hold up, and writes the post. The published piece is a human edit of a machine-assembled starting point.
What this is not
It is not original reporting. Nothing here comes from interviews, records requests, or attending a hearing. Every item is assembled from sources that already exist, and the value on offer is aggregation and context — noticing that something happened, and saying why it might matter — not first-hand discovery.
Treat what you read here as a lead, not a settled fact. If something matters to a decision you’re making, follow the links and read the primary source. They are provided on every claim for exactly that reason.
Sourcing and confidence
The research prompt requires a named source, a direct link, and a publication date for every item, and it forbids including anything whose source could not actually be retrieved. Significant claims — a law passed, a project approved or cancelled — are meant to carry two independent sources; a single-sourced item is flagged as unconfirmed. Enacted decisions are kept distinct from proposals, and press releases are treated as announcements rather than as independent verification.
Two failure modes come up often enough to name. The first is a model inventing a plausible URL, bill number, or quote. The second, and the more common in practice, is a genuinely real source that is simply old — search results routinely surface an article from years ago as though it were current news. Both are caught by opening the page and checking the byline date, which is why publication dates are verified rather than taken from a summary.
Adoption figures lag, often by months or quarters. Where they appear, they carry the period they describe, because a number reported without one reads as current and quietly misleads.
Québec’s authoritative sources — the Régie de l’énergie, the ministry’s planning documents — are published in French, with no English mirror. The prompt treats those French originals as the primary source rather than as a fallback: an English-language article about a decision is not a substitute for the decision. Quotations are translated with the original given alongside, and where a French source has been relied on in translation, the scan file says so.
Corrections
Errors get fixed rather than quietly rewritten. Each monthly scan re-checks the one before it, and where the previous briefing got something wrong — a date, a status, a mischaracterised decision — the correction is stated plainly in the new post instead of being left for a reader to notice. Corrections to a published post are noted in the post itself.