About
A one-page treasure hunt for a real park. Gemma 4, an open-weight model, writes clues from its map and recent wildlife sightings. Code checks each one. The finding happens on paper.
What we measured
Gemma 4 on 20 real parks, ages 6-10, 60 runs (53 passes), run 2026-10-10-3 (2026-10-10). Misses stay on the page. Full results
98%of clues quote their source exactly
Mettarget 85% or more
0risky species printed (60 runs)
Mettarget 0, always
$0.00107per pass for the clues (list price; up to $0.00111 if 4 timed-out calls were billed in full)
Missedtarget $0.00100 or less
Grade 2.5reading level (median)
Mettarget 3.5 or lower
90.2%of passes complete
Mettarget 90% or more
9.5 stypical model wait, 25.3 s slow
Missedtarget 10 s / 20 s
1.6%of clues repeat across parks
Mettarget 5% or lower
2754unit tests passing, Oct 10, 2026
Countedall green
In one look
gemma-4-31B-it on DigitalOcean serverless inference (US). One to three calls per pass for the clues, plus one for the trip tips.
No data? We say "No data available" and why. We never pad the list.
The fine print
In short: the real pass builder made passes for 20 real parks with each model, and we counted how often the clues were safe, true to the data, complete, easy to read, quick and cheap. Failures stay in (full results; the run cost $0.09). The last column is the goal we set before the test. Gemma's clues read at grade 2.5, the no-AI template's at 2.5.
| What we measured | Gemma 4 31B (open, the default)60 runs | Llama 4 Maverick (open, a slower alternative you can switch to)20 runs | No-AI template (our code, no model)20 runs | Target |
|---|---|---|---|---|
| Blocked (dangerous) species printedRisky things like snakes or poison ivy. Must be zero. | 0 | 0 | 0 | 0, always |
| Clues quoting their source word for word, before any filterHow often the model quoted its real fact exactly. | 98% (541/552) | 94.4% (201/213) | 100% (141/141) | 85% or more |
| Complete passes (kept at least n-1 items)At most one find missing. | 90.2% (46/51) | 94.1% (16/17) | 17.6% (3/17) | 90% or more |
| Honest empty sectionsNo data? The pass says so instead of filling the gap. | 100% | 100% | 100% | 100% |
| Reading level (Flesch-Kincaid grade, median)3 means a 3rd grader can read them. | 2.5 | 2.5 | 2.5 | 3.5 or lower |
| Clues or hints naming their own answer, before the filterCode removes these before printing. | 3.3% (clue only 2.9%) | 12.7% (clue only 11.7%) | 1.4% (clue only 1.4%) | 5% or lower |
| Model time per call, p50 / p95a usual wait / a slow wait (1 in 20 is slower). | 9.5 s / 25.3 s | 20.6 s / 31.3 s | no model call | 10 s / 20 s |
| Cost per pass (clue calls)At the provider's list price, for the clue calls only; the trip tips (added later) are one more short call. | $0.00107 | $0.00188 | $0 | $0.00100 or less |
| Printed clues repeated across parksShares 5 words in a row with clues for at least 2 other parks. | 1.6% (6/383) | 2.3% (3/130) | 20.7% (19/92) | 5% or lower |
| Printed clues with a wrong countA count must match the map. In brackets: wrong counts code removed first. | 0 of 74 count clues (15 removed) | 0 of 12 count clues (5 removed) | 0 of 3 count clues (1 removed) | 0 |
| LicenceThe terms for using the weights. | Apache-2.0 | Llama 4 Community Licence | MIT (our code) |
No closed model was compared: we chose open models only, and the closed models on our DigitalOcean account answered "403 Forbidden" on 2026-10-05. Switching models is one setting (MODEL_ID).
Run it yourself (measured, a separate small test): We ran the small Gemma 4 E2B (gemma4:e2b-it-qat, Apache-2.0) with Ollama on a Windows laptop CPU (Intel Core Ultra 7 155H, 32 GB RAM, no GPU), on 5 test parks, for $0. With a 70 s limit per model call (the most the app allows; the hosted site gives a first call 30-40 s), 0 of 5 passes were complete: 3 ran out of time and the other 2 came out short. Given more time (an eval-only setting), 4 of 5 were complete, 96.7% of clues quoted their source, reading grade 2.9, 0 risky species printed, at 59.3 s a typical call (hosted Gemma 4 31B: 9.5 s). The app now has a longer clock for a model on your own computer (LOCAL_MODEL_TIMEOUT_MS, off by default): in one browser try, a Celebration Park pass came out with 7 of 8 finds and its map in 96 s (2 model calls). The model used about 5-6 GB of RAM. Setup, every number and how to repeat it: self-host notes.
| What | Where it goes | Why |
|---|---|---|
| The place you type (for example "Allen TX") | Our server (never in the web address), then OpenStreetMap's Nominatim. Cached 30 days by the text, not by who typed it. | To find the town or park. |
| "Use my location" | Rounded in your browser to about 1 km, then our server, then OpenStreetMap's Overpass. | To list parks near you. |
| The map area of your search results | Your browser asks the OpenFreeMap tile server (tiles.openfreemap.org) for the streets of that area, so like any website it sees your IP address. No account, no cookies. OpenFreeMap says its normal access logs keep no IP address and are deleted after 7 days; error logs and logs kept after misuse can hold IP addresses, for up to 30 days. Cloudflare delivers the tiles under its own privacy policy. Details: openfreemap.org/privacy. | To draw the street map behind the park pins. |
| The park you pick (a public place) | Our server, then OpenStreetMap, iNaturalist and SerpApi (name and map position only). The pass page, and making a pass, send only the park's position, rounded to about 1 km, to Open-Meteo and (US parks) the National Weather Service. | For the park map, sightings, monarch counts and review counts, the weather card and the trip tips. |
| The age band (for example 6-10) | Our server, then the model on DigitalOcean, in the prompt. | To set how many finds and how easy the words are. |
| Your IP address | Our server. Our storage (Upstash Redis) keeps only a keyed hash, never the address itself, in rate-limit counters that expire within a day. | To stop abuse and keep the model budget fair. |
| Every request (IP address, web address, time) | Our hosting provider's request logs (Vercel), kept about 1 hour. They never show what you typed or where you are. | Running the website. |
| Signing in with GitHub or Google or with "Try as a judge" (grown-ups only) | GitHub or Google send your public profile (account number, name, picture link; for GitHub any public email). We store only a scrambled ID made from the number; the rest is dropped at once, except your first name, which stays in your own encrypted cookie. Sign-in lasts 7 days; the judge demo sign-in stops working after 1 day. | To count your 5 new passes a day, your reports and your ratings. |
| Your free pass (no sign-in) | One cookie in your own browser, set only when you use a free pass: today's date (Dallas time) and how many free passes you used, signed so it can't be changed. No ID. It expires at midnight Dallas time. Our server keeps no copy. | To count your 1 free new pass a day. |
| Your pass ratings (1 to 5 stars and tags, no text; signed-in grown-ups) | Our storage: your latest rating per pass and its day, under an ID made for that park only. Deleted after 90 days. Never shown publicly; only the site owner sees the totals. | To learn which passes work for kids. |
| Your reports (Found it, Didn't find it, Not safe) | Our storage: your latest report per find and its day, under an ID made for that park only. Deleted after 90 days. | To learn what is findable and catch anything unsafe. |
| The finished pass (park, age band, finds, clues, times) | Saved in our storage (Upstash Redis) for 30 days, so the link and print page work. | Nothing in it is about you or your child. |
No names, no photos, no analytics; nothing about the child is asked for. Browsing and printing set no cookie. Using a free pass sets one small signed cookie (the date and a count, no ID), and a grown-up who signs in gets a sign-in cookie. We only keep a scrambled ID to count your new passes, reports and ratings. No email, no name. Your browser keeps only your light or dark choice and the last age band.
The model runs on DigitalOcean servers in the US, so park facts and the age band leave your device.
Our logs record which source or model ran, timing, outcome and pass id; never the prompt, your IP or what you typed.
Anyone can search, open examples and shared links, and print. A NEW pass is a real model call. Without signing in you get 1 free new pass a day (counted by one small signed cookie, and at most 3 signed-out new passes a day per internet connection). After that a grown-up signs in with GitHub or Google (judges: "Try as a judge"): 5 new passes a day each, reset at midnight Dallas time. No password is stored.
Judges: "Try as a judge" is one click, no sign-up. All judges together can make 60 new passes a day with it, at most 3 from one internet connection.
Signed-in grown-ups can report each find (Found it, Didn't find it, Not safe). A find is left off new passes for that park when at least 3 different signed-in visitors didn't find it in the last 30 days and they are more than 60% of those who reported on it, or at once when 2 say it is not safe (we then review it). Each visitor counts once per find. Judge demo reports are only logged, never counted. Reports are deleted after 90 days.
Signed-in grown-ups can also rate a pass: 1 to 5 stars and tags (Too easy, Too hard, Kids loved it, Something was missing, Not safe), no text. One rating per account per pass (the newest counts), deleted after 90 days. Judge demo ratings are only logged, never counted. The totals are never shown publicly.
Weather on the pass page (the card's words are written by code):
Trip tips ("How to make this a great trip", on screen only): written once per pass by the same open model, from facts code writes first: that day's forecast and alerts, the park's mapped fountains, restrooms, shelters, water and paths, and the few sightings worth planning for (poison ivy, ticks, fire ants, mosquitoes, venomous snakes, stinging insects). Code drops any tip not based on one of those facts, or that says to touch, pick, feed, wade, leave the path or take medicine. If the model doesn't answer within 10 seconds, the page shows a short code-written list from the same facts and says so. One more model call per pass.
Any other species whose own description says it is poisonous, toxic, venomous, stings or burns the skin is left off too, and a clue that uses one of those words is removed.
Every Wild Find carries a fixed "look, don't touch" line written by code.
Thank you
Arbor Hills Nature Preserve (trail vista) by Robert Nunnally, CC BY 2.0 (March 2008)
Dallas Texas - HCP - September 21, 2022 - 007 - White Rock Lake by Vulturesong, CC0 1.0 (September 2022)
Sunday morning walk (Celebration Park, Allen, Texas, November 20, 2011) by Robert Nunnally, CC BY 2.0 (November 2011)
Rowlett Creek in Oak Point Park, Plano, Texas, USA by Jackilometresan, CC0 1.0 (April 2021)
Connemara Meadow (May 2, 2015, during the monthly bird walk) by Robert Nunnally, CC BY 2.0 (May 2015)
Hero illustration generated with v0 by Vercel (AI), not a photo. It shows no real child or park.
MODEL_ID is switched to it; passes then show "Built with Llama".