Corsica — one trip, three variants, two renderers

→ Open the trip

The day-by-day carousel is the front page. Three variants of the same island:

	Day view	Long reference page
Van, 16 days	/	van.html
Car + hotels, 16 days	/?v=hotels	hotels.html
Unrushed, 27 days	/?v=paced	paced.html

A single-page trip-planning site built from the road-trip film Corsica | An Incredible 10-Day Road Trip, re-paced from the film's 10 days to 15 of content over 16 days away — keeping every stop, and adding a buffer day before the ferry.

No build step and no framework. Open index.html in any browser, or use corsica-van-trip.html — the same site inlined into one 316 KB file you can email or open offline. Either way the interactive map needs internet access for Leaflet and the map tiles; everything else works offline.

Two renderers

Renderer	File	For
Day view (landing)	index.html + days.js	one day at a time — swipe or arrow through a carousel
Long reference page	van.html, hotels.html, paced.html + app.js	reading the whole trip, planning, the reference tables

The day view is the low-clutter reading mode: a left rail of days, and a carousel where each slide is one day — hero photo, a mini map of just that day with start ⚑ and finish 🏁 flags (a single flag when the day is a loop), the running order, a photo strip, and stops / eating folded into <details>.

Mini maps are deliberately non-interactive (no drag, no wheel zoom) so they never fight the swipe gesture, and they initialise lazily — only the current day and its two neighbours, rather than 16 Leaflet instances at once.

It serves all three datasets — /, /?v=hotels, /?v=paced — and the rail's Reference links point at whichever long page matches.

days.html remains as a redirect (it was the URL before the day view was promoted), and index.html remaps the long page's old anchors — #day7 becomes the carousel's #d7, and #camping and friends go to van.html or hotels.html. Slides carry id="dN", so hash deep links scroll the deck natively.

Three variants of the same trip

The same stops, coordinates, photos and routing engine, re-planned around a different constraint — a demonstration that changing how you sleep changes the whole shape of a trip, not just one line of it.

	Van	Car + hotels	Unrushed
File	van.html	hotels.html	paced.html
Days	16	16	27
Sleeps	13 campsites + 2 hotels	7 hotel bases	20 campsites
Shape	one continuous loop	6 days are loops from a base	one loop, cut by optimiser
Driving	1,448 km / 34 h	1,582 km / 37 h	1,340 km / 31 h
Stops	76	65	51
Mean day	12.6 h	12.0 h	7.7 h
Days over 10 h	12	10	0
Cost, 2 people	≈ €2,605	≈ €3,907	not costed
Cut by	the film's route	the film's route	evidence + a solver

The first two are the film's trip at the film's pace, in two sleeping styles. The third is the same island re-cut so it can actually be followed — see Pacing below for how, and why it costs 11 more days.

What's in it

Section	Contents
Map	Leaflet map, 16 selectable day-routes drawn on real road geometry, 104 clustered pins across 12 categories, terrain/street basemap switch
Itinerary	16 days, each with a running order, photo gallery, stops, hikes, a van hazard, vegan food and the campsite it sleeps at
Eating vegan	Per-day vegan-friendly places, evidence-labelled, plus a strategy section — Corsica is genuinely hard for vegans and the page says so
The day, roughly	A guide-style running order per day — soft blocks ("Morning", "After lunch") with slack, and hard clock times only where something won't wait
Timing	Every day carries a "be at your bed by" time; day 16 is a deliberate buffer before the ferry
Photos	88 Creative-Commons photos from Wikimedia Commons, 6 per day, with a keyboard/swipe lightbox that credits every photographer
Hotels	5 mid-budget hotels (+1 mountain bonus) in the van version, 7 bases in the hotel version
Camping	31 van-friendly campsites with coordinates, price bands, check-in windows and websites, plus the Corsican wild-camping law
Hikes	17 hikes and walks with distance, ascent, time and grade
Beaches	21 beaches and freshwater swim spots
Van logistics	Ferries, real driving speeds, the roads a van should respect, water/waste/gas, when to go
Budget	Itemised estimate for two people over 16 days, arithmetic self-checked

Where the data came from

• Stops and pacing — the film's timestamped chapter list plus its narration. The transcript was fetched with youtube-transcript-api and is kept in data/transcript.txt. Where the narrator gives his own hike figures (Capo Rosso 8.5 km / 500 m / 3–4 h, Monte Senino 6 km / ~4 h, Bavella ~12 km / ~1,000 m, Figarella 5 km / 300 m, Melo 25 km / 8 h) those numbers are used in preference to guidebook figures.
• Coordinates — Nominatim / OpenStreetMap. Every one of the 111 points is inside the Corsica bounding box; the validator in gen_data.py output checks this.
• Campsites — an Overpass query for tourism=camp_site and tourism=caravan_site across ISO3166-2=FR-20R returned 254 sites (242 named); 30 were curated to match the route. Raw results in data/campsites_raw.json.
• Routes — OSRM driving directions between each day's waypoints, so the lines follow actual roads and the distances and times are real. Geometry simplified with Ramer–Douglas–Peucker (55,249 → 6,067 points) to keep data.js at ~163 KB.
• Photos — Wikimedia Commons, selected per stop by geosearch around its coordinates plus a name search, then scored on licence, resolution, aspect ratio and title relevance. Candidates carrying coordinates outside the Corsica bounding box are rejected outright, which is what caught a Slovak tunnel, a Normandy forest and a Tuscan town hall matching on place-name alone. Each day keeps its 6 best, one per stop before any stop gets a second. Downloaded and re-encoded locally to WebP (1200 px full, 560 px thumb, 16 MB total) so the site serves its own images. Photographer and licence are recorded per photo in data/photos.json and shown in the lightbox, as CC-BY/CC-BY-SA require.
• Restonica access — verified by web search: the D623 has been closed beyond the Pont de Tragone since the November 2023 storms; the trailhead is now the A Frasseta car park, served by the Navetta Restonica C13 shuttle from Corte (2 May – 30 Sep, €4 return, app booking only). This materially changes day 10 and the page says so.

How the variants relate

Both are generated from one stop dataset by two generators:

gen_data.py    ──┐                       ┌── data.js        → van.html
                 ├─ shared POI list ─────┼── data-hotels.js → hotels.html
gen_hotels.py  ──┤  (exec'd from         └── data-paced.js  → paced.html
gen_paced_data ──┘   gen_data.py)                              + all three feed index.html

gen_hotels.py imports the POI list out of gen_data.py (by exec'ing the file up to the DAYS marker, so importing does not trigger a write), then re-groups the stops into 16 different days, re-routes them, overrides the 34 stop notes that assumed a van, and drops the 10 stops with no mid-budget bed within an hour.

gen_paced_data.py goes further: it takes the day partition from the optimiser (see Pacing below) rather than a hand-written day list.

That is the interesting part of this repo: changing a constraint is a re-plan, not a filter. Route topology changes (6 of 16 days become closed loops from a base), coverage changes, and total driving goes up by 135 km because loops backtrack.

The same page renderer (assets/js/app.js) drives both; section builders no-op when a variant omits their section, which is how hotels.html simply has no camping section.

Pacing: the plan was rushed, and the numbers said so

The first 16-day plan read well and could not be followed. Audited against realistic time per stop (audit_load.py), 12 of 16 days were over 10 hours, the mean was 12.6 h, and five days stacked a long drive on top of a day-long hike — day 14 had 4 h 53 of driving and a 4 h hike.

Two things caused it, both self-inflicted:

1. Scope creep. 24 of the 56 core stops are places the source film never visits — 51 hours, 40% of the content — added because they are good, without ever saying the trip had grown.
2. A slower pace. The film's days run ~11 h. Mine assume you do things: 5.5 h at Saleccia because the boat sets the day, 8 h for Melo.

The model is calibrated. Given the film's own stops and allowed to stack the way the film does, the optimiser returns 11 days. The film was 10. That is what makes the rest of its numbers believable.

The optimiser

The route is a fixed geographic sequence, so choosing days is a matter of cutting that sequence into consecutive chunks — a 1-D partition problem, solved exactly by dynamic programming in optimiser_core.py rather than by eye. Objective: fewest days within a budget, then the most even spread. Constraints:

• a day must end where you can actually sleep (accommodation within 25 km, by real drive time)
• at most one effortful activity per day, and when a day has one, driving is capped at 1 h 45

That second rule is the one that matters. It is what stops a three-hour drive and a four-hour hike landing on the same date, and it is why the unrushed variant needs 27 days where the original claimed 16.

Effort is marked explicitly (pacing.S(..., hard=True)). An early version keyed the rule off raw duration and so treated three hours on a beach like a three-hour climb, giving Rondinara and Santa Giulia a whole day each. Two hikes in a day is a bad day; two beaches is a good one.

Pace is a parameter, and the trade-off is computed rather than guessed:

Scope	Relaxed 7 h	Steady 8 h	Standard 9 h	Brisk 10.5 h
Whole island	31	27	25	22
North + west	22	21	18	16
South + centre	14	11	11	10

Three bugs, all from trusting a router

• OSRM snaps lakes, summits and boat-only beaches to distant roads. The Melo trailhead came back as 149 min from a point 5 km away. Non-road stops now carry an explicit road access point (pacing.ACCESS) — driving fell from 42.8 h to 28.8 h.
• Campsites were matched to stops by straight-line distance, which paired Bonifatu with a site 8 km away across an unroadable mountain range. Bases are now chosen on real drive time, weighing tomorrow's departure as well as tonight's arrival.
• Legs between stops are sanity-capped against straight-line distance; 16 of 53 were implausible.

Corroboration: how well attested is each stop?

The whole plan rested on one film plus two dozen stops added on nothing but opinion. corroborate.py fetches ten independent Corsica guides and itineraries and matches every stop against them; the count is stored in the data and shown on each stop as a chip.

Of 76 stops: 20 scored 7+/10, 25 scored 4–6, 31 scored 0–3.

Two findings, pointing opposite ways:

• Coverage is good. Across ten guides, essentially nothing is missing. Propriano (8/10) was the only real gap, and has been added.
• The evidence contradicted several of my own choices. Stops cut from the unrushed variant as "padding" turned out to be among the best attested — Piana 9/10, Porto-Vecchio 9/10, Palombaggia 8/10, Rondinara 7/10 — while Évisa (1/10), Scala di Santa Regina (1/10), Bodri and Galéria (2/10) were kept. The cut list is now evidence-led and those four are restored.

Six stops score 0/10 and come only from the film: Monte Senino, Bocca di Vezzu, Canyon di A Ruda, the Bridal Veil falls, Ghisoni's D69 and Cristinacce. That is not a verdict — an unlisted place a filmmaker raved about is a find, not a dud — so the chip reads "the film's own find" rather than a bare zero. But the unrushed plan gives Monte Senino a whole day, and it is worth knowing no guidebook we checked mentions it.

Files

index.html                 the day-by-day carousel (landing page)
van.html, hotels.html      the two long reference pages
days.html                  redirect, kept so older links still resolve
assets/css/style.css       styles
assets/js/data.js          generated: pois, camps, stays, days (+route geometry, +photos)
assets/photos/             88 CC photos, WebP, full + thumb
assets/js/app.js           map, filters and section renderers (long pages)
assets/js/days.js          the day-by-day carousel view
assets/css/days.css        carousel styles

pacing.py                  minutes per stop, core/option, hard=effort, road access points
optimiser_core.py          the DP day-partition solver
audit_load.py              how full is each day, really
optimise_days.py           pace sweep over the whole route
scope_options.py           pace x scope trade-off table
video_only.py              calibration: does the model reproduce the film's 10 days?
build_paced.py             re-solve with the evidence-led cut list
gen_paced.py               bases + routing for the unrushed variant
gen_paced_data.py          emit data-paced.js
corroborate.py             score every stop against 10 independent guides
candidates.py              places the guides rate that the plan lacks
eats.py                    vegan places + per-day eating notes
flows_van.py               per-day running order, van version
flows_hotel.py             per-day running order, car + hotel version
gen_data.py                regenerates assets/js/data.js (van)
gen_hotels.py              regenerates assets/js/data-hotels.js (car + hotels)
build_routes_hotels.py     OSRM routing for the hotel version's day loops
data/                      raw research: routes, geocodes, campsites, transcript
shots/                     rendered screenshots from the verification run
verify.py, shots.py        headless-Chromium checks and screenshots
verify_days.py             checks for the carousel view (run: `.venv/bin/python verify_days.py`)

Regenerating

python3 fetch_photos.py                   # re-curate photo picks from Commons
python3 download_photos.py                # download + re-encode them
python3 corroborate.py                    # re-score stops against the 10 guides
python3 gen_data.py                       # rewrite assets/js/data.js
python3 gen_hotels.py                     # rewrite assets/js/data-hotels.js

# the unrushed variant, in order
python3 audit_load.py                     # how bad is it?
python3 build_paced.py                    # solve the day partition
python3 gen_paced.py                      # pick bases, route, re-check real loads
python3 gen_paced_data.py                 # emit assets/js/data-paced.js
python3 -m http.server 8765               # then open http://127.0.0.1:8765
.venv/bin/python verify.py                # DOM/console/interaction checks + screenshots

verify_days.py covers the carousel across all three variants plus a narrow-viewport pass. verify.py asserts the day list, chips, itinerary, tables, pins and popups all render, that day- and category-filtering actually changes the visible pin count, and that the page throws no console errors or failed requests. verify_gal.py exercises the gallery and lightbox; verify_live.py runs the same checks against the deployed GitHub Pages URL rather than localhost.

Driving the UI rather than reading the code caught bugs that review would not have — most notably that selecting a day fitted the map to the driven route only, so stops reached on foot or by boat (Lac de Melo, Saleccia, the Lavezzi) fell outside the viewport.

The running order

Each day carries a "The day, roughly" panel: a narrated order of play in the voice of someone who has done it — wake time, what to see before the heat, where lunch happens, what to skip if you are behind.

It is deliberately not a minute-by-minute schedule. Blocks are soft ("Early", "Late morning", "After lunch") and carry slack, because this is a holiday. Across both versions there are 212 blocks and only 34 fixed times — and those are exactly the things that will not wait for you: boat sailings, the Restonica shuttle, campsite receptions, sunset, and the ferry. They are marked FIXED and everything else is a suggestion.

Flows live in flows_van.py and flows_hotel.py, separate from the stop data, so the narrative can be rewritten without touching coordinates or routing.

Eating vegan

Corsican cuisine is charcuterie, brocciu, wild boar and fish, and there are exactly three fully plant-based or all-vegetarian kitchens on the island — VG in Bastia, and Green Farmer's and A Cantali in Ajaccio. The site does not pretend otherwise.

Every day carries an "Eating vegan today" block. 32 places are named, all of them existing in OpenStreetMap or HappyCow at the coordinates given — none invented — and each carries an evidence badge, because the strength of the source varies a lot:

Badge	Means
Fully vegan	plant-based kitchen
Vegan options	vegan dishes confirmed (OSM diet:vegan=yes, or a HappyCow listing)
Veg-friendly — ask	OSM diet:vegetarian=yes; vegan is a conversation
Shop / market	organic supermarket, greengrocer or market

On the days where the honest answer is "carry your own" — Saleccia, Piana and Arone, the Lavezzi, the day-15 east coast — the block says that instead of inventing a restaurant. The #vegan section covers the strategy: pizza marinara as a floor, self-catering, saying végétalien rather than végétarien, and phoning mountain auberges in the morning.

Sourced from an Overpass query for diet:vegan / diet:vegetarian across Corsica (data/vegan_raw.json, 72 results) plus HappyCow. Data lives in eats.py.

Why 16 days for 15 days of content

Both versions are 15 days of plan plus a buffer day, and night 15 is spent 20 minutes from the ferry gate. Ferries, mountain weather and hire cars will take a day off you sooner or later; without a buffer that day comes out of the crossing you already paid for.

The same realism drives the per-day timing. Corsican campsite receptions run roughly 08:00–12:00 and 15:00–19:30 (earlier at altitude, earlier out of season) and hotels check in from 15:00 — so every day states a target arrival, generally 18:00, and the driving is sized to hit it. Days whose payoff is a sunset (the Calanches, the Sanguinaires, Nonza) check in first and go back out. Van day 10 sleeps in Corte a second time for the same reason: an eight-hour hike and a 19:30 reception deadline cannot share a day.

These windows are typical, not per-property facts, and the page says so.

Caveats

Coordinates, road routes and campsite existence are machine-verified. Prices, opening hours, shuttle timetables and road closures are not — they change. Confirm the Restonica situation locally. Wild camping is illegal in Corsica.