Thirty-eight questions a careful judge would ask

It keeps a segmented baseline of the coast (250 m stretches), applies recorded or forecast conditions to it, and names the storm-impact regime that is plausible at each stretch, hour by hour: Swash, Collision, Overwash or Inundation. Before an event that is a local action picture; after it, a traceable evidence record. Built for County Wexford, with Storm Kathleen (6 to 7 April 2024) replayed on the Kilmore Quay and Courtown coasts.

Real and live: the GIS, the coastline and elevation layers, the Kathleen replay, the census and planning context layers, the satellite scene, the agent runs on the Agents page. Next to build: the forecast feed, file upload and persistence, sign-in, the approval workflow, and the assistant's model connection. The prototype is preconfigured for Wexford. We say which is which so a judge never has to guess.

AI agents discover, parse and check evidence: they find public services, page them in full, reproject, trim, attach provenance and read licences. They stop there. A deterministic engine built on published methods (Stockdon et al. 2006 runup; Sallenger 2000 regimes) calculates every result. The language model cannot change a regime label; it sits behind a policy gate that rejects unknown sources, unsupported numbers, missing caveats and engineering language. A named person approves what is kept and what leaves.

The Agents page renders from a file of real runs with real timestamps and commit references: ten runs on 11 and 12 September, one refusal, one correction, one agent with no run because it had none. They are supervised jobs a person starts and reviews, not a daemon. The pitch deck's agent slide is drawn from the same file.

On the morning of a storm a coastal engineer opens one screen and sees which stretches to look at first, when conditions peak, and which evidence is missing, with the source behind every number. A ranked list of named stretches replaces a whole-coast survey. One storm run over a Wexford-sized coast is two to three analyst-days instead of a programme. Afterwards, the same screen is the evidence record for the grant application, the audit and the next storm.

Ireland has 8,575 km of coastline, 19 coastal local authorities in the Republic, and the same data, methods and procurement regime for all of them. One maintained national baseline, extended authority by authority, with every storm adding a validation point. The physics is published and travels; the evidence discipline is what CoastLens brings. At Sea View the State has approved about €2.5m of protection since 2021 and rebuilt one road four times. The picture that decides where that money goes next is the product.

Eighty-seven named storms since 2015. UCC published the first national exposure map this month. The open data exists, the models exist, and AI can now do the evidence work that a three-person coastal team could never afford to keep current by hand. The missing layer is the operational one: what this storm does to this stretch, on this council's desk, in time to act.

Coastal local authorities. A fixed-scope proof of value at €48,000 for twelve months, then an operating service at €60,000 per authority per year, priced per authority and never per kilometre because procurement, governance and grant evidence all attach to the public body. Six named authorities is €360,000 a year; all nineteen coastal authorities in the Republic is €1.14m, with ports, transport and asset owners on top. Every figure on the pitch slide traces to a source or a stated assumption.

Yes. The State already buys this class of evidence: OPW's coastal aerial photography and LiDAR survey contracts were awarded at €79,850.75 (2024) and €81,244 (2025), from eTenders award notices. CoastLens competes for an existing line and does it continuously rather than once a year from an aeroplane. At Sea View alone, OPW has approved about €2.5m of coastal protection since 2021, and the access road there was built four times.

Through the published quotation route. Under Circular 05/2023 a fixed scope below €50,000 can go by three-quote or eRFQ rather than full tender, which is the difference between a first sale in one quarter and three. We say this openly: it holds only where the authority's genuine aggregate requirement is below the threshold, with no options or renewals that ought to be aggregated. CoastLens is not a way around procurement.

The alternatives today are periodic, not operational. Aerial survey contractors (Bluesky International won OPW's 2024 coastal survey at €79,850.75) fly once a year and deliver imagery, not an event picture. Engineering consultancies produce coastal studies per commission, valid for years. National research (UCC's exposure index, OPW's flood maps) screens the whole coast once. AI-native start-ups tend to generate a confident number rather than report NO_DATA. None of them sits on a council engineer's desk on the morning of a storm with a ranked list of named stretches and the evidence behind each. CoastLens does, continuously, on the council's own segments, and it is built on the open data the others produce, so it complements them rather than competing for the same commission.

Named, each with a documented reason. Year one: Wexford County Council (live Courtown scheme, live OPW Minor Works register) and Cork County Council (a CARO-recorded coastal protection project). Year two: Clare (Kilcredaun flooding response) and Causeway Coast and Glens (237 km of coast, active PEACEPLUS monitoring, a second jurisdiction). Year three: Dún Laoghaire-Rathdown and Wicklow, where Irish Rail has itself named exposed coastal rail cells. Irish Rail and TII are the adjacent buyer with the least tolerance for surprise: service disruption has a cost a council's land loss does not.

Not the engine. Stockdon and Sallenger are peer-reviewed papers and the data is open; a competent coastal engineer could rebuild the equations in a fortnight, and we chose published physics precisely so a council can audit it. The moat is what compounds: the validated-storm history (every event where a pre-inspection list is compared with field findings), the evidence corpus with its provenance chain including the dead ends, the maintained datum-disciplined baseline extended authority by authority, and position in the council's own record once briefing packs are cited in storm response and grant applications. And one unusual one: a system that provably will not over-claim is easier to buy than one that might.

We do not claim it, even as a goal. Nothing links a coastal outlook to insured loss, and coastal land loss is not a claims category. Storm costs are real (Éowyn: more than €301m in claims; Chandra: €13.49m of Wexford road repair) but they are wind and flood damage, not what CoastLens does. Insurers are a later conversation, once a pilot has numbers.

The physics travels for free: Stockdon and Sallenger apply to any sandy or dune coast. What does not travel automatically is the evidence work, which is exactly what the agents are for. Ireland first, because the national data is open and the 19 authorities share one procurement regime and one set of conveners.

One role each. Source Scout finds a public service or page and pulls it in full, paging past caps (the national coastline service stops at 2,000 records; the first pull had 1,047 features, the paged pull 3,845). Evidence Parser reprojects, trims and attaches provenance; it converts units and records the step, never changes a value. Provenance Validator reads licence, source and datum before anything is kept, and refuses when it must. Gap Resolver matches a missing input to a source or records NO_DATA; it never bridges a gap. Layer Builder wires accepted data into the GIS and verifies it headless. Marine Monitor refreshes wave and water feeds; it had no run this window and says so.

A person. They are supervised jobs with a written brief and a bounded scope, run in a coding agent under review, not a daemon polling the internet. Every run ends with a human decision: keep, correct, refuse or release. In production that person is a named council officer for anything that leaves the system, and a CoastLens analyst for anything that enters it.

Three things. The brief forbids it and the run is reviewed. The engine will not accept an input without source, date, coordinate and datum. And the policy gate downstream rejects unknown source IDs and unsupported numbers. When the Gap Resolver cannot find an input, the segment stays NO_DATA with rank: null; missingness alone never creates a council action.

One correction this week: the commercial slide carried a national revenue figure built on a headcount nobody had sourced; the audit found the published count of coastal local authorities (19) and the figure was replaced. One refusal: OPW flood extents, on licence. Both are on the Agents page because a record of refusals and corrections is worth more to a buyer than a clean feed.

Compute is not the constraint. Storage and hosting for a county come to roughly €50 to €60 a month. The cost line that matters is human review: segment review runs at 75 to 100 segments per analyst-day, so one storm run over a Wexford-sized authority is two to three analyst-days. Everything the agents automate moves that number, which is why the product is a ranked list of stretches rather than a report on the whole coast.

Tailte Éireann high water mark (3,845 features, CC BY 4.0). OPW LiDAR 2011 (98 of 245 km of the Wexford coast, CC BY 4.0). Marine Institute M5 buoy record. Copernicus Sentinel-2 (open). CSO Census 2022 small areas and DHLGH planning applications (CC BY 4.0). OpenStreetMap (ODbL). Esri World Imagery (display only, attributed, undated). Every layer carries its source, licence and date.

OPW coastal flood extents (CFRAM 2021): downloadable, but licensed CC BY-NC-ND and floodinfo.ie forbids adaptation and commercial use. Found, read, left out, recorded as a licence gap, not a data gap. UCC's Coastal Exposure Index (Walsh et al. 2026): its web map allows non-commercial research and educational use with attribution, so it appears on one slide and as a labelled, not-for-release layer presentation licence; product use needs the lead author's agreement, which we will ask for.

The national high water mark has no feature at all for 955 m north of the harbour, exactly where the rock armour is. CoastLens shows that stretch as a source gap rather than draw a line that is not in the data. The retained Courtown stretches are the cliff south of the harbour and 1.6 km of beach north of the gap, the latter with no engine result yet.

UCC's DETECT group, whose exposure index tells us which stretches to baseline first and who could use our event results as ground truth. OPW, who hold the LiDAR, the flood maps and the Minor Works funding that pays for protection at places like Sea View. The Marine Institute, whose wave model is the forecast feed. Met Éireann, whose warnings are the attributed official input a CoastLens outlook sits beside, never replaces. The Climate Action Regional Offices, who already convene the 19 coastal authorities. And Wexford County Council, the first customer, whose engineers know this coast better than any dataset.

That is the design. Council surveys, drone runs, photographs, inspection reports and access records enter through the same agents and the same checks as national data, keep their provenance, and appear in the evidence rail beside everything else. Upload is the next build; the workflow and the checks it goes through are live.

We do not say, and will not. CoastLens names the regime and when it peaks. Volumes, areas and retreat distances need survey. The "eighteen football pitches" at Courtown is Wexford County Council's figure reported by RTÉ (4 July 2025), not ours.

Never, at any confidence. Hard defences can move erosion next door. Outputs are operational: where to look, when, what evidence to collect. Engineering is the council engineer's call.

It says NO_DATA. Every score exposes its components and its missing state. Nothing is interpolated, borrowed from a neighbouring stretch, or filled with a guess. Storm Kathleen's evidence-grade result at Sea View stays NO_DATA until every mandatory input passes; what the demo shows is a Kathleen-informed prototype scenario, labelled PROTOTYPE / NOT FIELD VALIDATED on the display.

The methods are published and the regime boundaries are unit tested. It is not yet calibrated against field measurement, so it produces no probabilities and no damage estimates. Forcing comes from offshore records (the M5 buoy is tens of kilometres from the sites), so we describe a plausible forcing envelope, never "we recreated the storm".

Covered today: 98 km of the Wexford coast with a public elevation baseline, the full national high water mark (3,845 features), a 93-hour storm record, sixteen pilot stretches with source-derived geometry, and four context layers. The limits, stated on screen rather than in a footnote: sixty per cent of the coast has no public elevation baseline (OPW LiDAR 2011 covers 98 of 245 km), and the baseline that exists is thirteen years older than the storm. Forcing is offshore: the nearest buoy is tens of kilometres away and the surge value is a reanalysis grid. Tide at Sea View is a synthesised semi-diurnal proxy, labelled. The surge reanalysis ends in 2024. Storm Kathleen's evidence-grade result is still NO_DATA. Seven of Courtown's eleven stretches have no engine result yet. All of this is on screen, not in a footnote.

A post-2011 elevation survey of the uncovered coast (OPW's annual aerial programme is the obvious vehicle). A local water-level chain at Sea View with a stated uncertainty. A nearshore wave transformation from the offshore record, which is the engine's known blocker. A second benchmark storm with field findings to validate against. Each is named as REQUIRED or CONTEXTUAL in the research register with the downstream consumer it unblocks.

No. A CoastLens outlook is a proposal for local validation by a council, not an official warning. Met Éireann warnings are a separate, attributed input. Consumer alerts, if they ever exist, would supplement official warnings and never replace them. Nothing on this site has been sent to an external authority.

Connect the Marine Institute wave forecast and run the outlook mode on a real feed. Extend the baseline to the north beach at Courtown once the coastline gap is resolved with Tailte Éireann. Ask UCC and OPW for what we could not take. Run Wexford County Council's first paid proof of value against an agreed baseline, and validate the first pre-inspection list against what the engineers find on the ground. That last one is the moat starting.

They are inventory and screening: the UCC exposure index, OPW's flood maps on a 2021 base, OPW's annual coastal aerial survey (about €80k a year, from eTenders award notices), the Government's 2023 coastal change report. Each is published once, valid for years, and silent on the morning of a storm. CoastLens is operational (per event and per forecast hour), continuously maintained (agents keep source, licence and datum on every item), deterministic with NO_DATA visible, and delivers a decision to a named person. It is built on the open data those studies produced. It complements them.

Built on their open data, not partnered yet. UCC's index says which coasts are exposed in general; CoastLens says what this storm does to this stretch. The natural fit is to use their index to choose which stretches to baseline first and feed event results back as ground truth. The requests to UCC and OPW are the first emails after the pitch.

Same engine, different input: tomorrow's wave and tide forecast in place of Kathleen's record. The Marine Institute's wave model is open and is the feed that unlocks it; it is not connected yet, and the workspace says so ("forecast pending"). A CoastLens outlook would be a proposal for local validation, never an official warning; Met Éireann warnings are a separate, attributed input. We do not infer impact from a storm "looking like Kathleen"; Kathleen is how we explain a result and check the engine, not the method.

Today the Wexford prototype is preconfigured. In production a council would use its organisational sign-in, choose its authority and priority stretches, and connect approved national sources plus its own surveys and observations. One shared national data and methods foundation, council-specific layers, the same versioned engine. Not a rebuild per authority.

Surge reanalysis (GTSM-ERA5) ends in 2024, so Storm Éowyn (2025) and Storm Chandra (2026) cannot be reconstructed the same way. Sixty per cent of the Wexford coast has no public elevation baseline. Forecast forcing is not fed. We would rather say so than have you find it.