Mike's Checks/claude-opus-5.5/11 hoboken-map
11 hoboken-map
claude-opus-5.5Claude Codehigh effortrun 20 Sep 2026270,532 tokens
▸Instructions — what the model was asked
11 — Neighborhood map + self-validating "hidden gem score"
This is a short multi-turn task. There are no data inputs — do your own
public web research to find real places. Treat the prompts below as a
conversation that arrives in order; the interesting one is the last.
Prompts (in order)
- My wife and kids and I just moved to north Hoboken — make me an interactive
map of family recommendations (playgrounds, food, activities, day trips). - Can you add a date-night category?
- What about a tapas place?
- Now work on a proprietary "hidden gem" score. Determine a way to measure
it and validate that it matches expectations. Use public data where possible.
What "good" looks like
- A single self-contained, interactive map artifact (
index.html) a non-technical
family member could open and use, with the categories above. - Real, plausibly-correct Hoboken-area places — not invented ones.
- For prompt 4: a genuinely thought-through metric. Pre-register a methodology
and your priors/expectations before collecting and scoring data, then test the
computed scores against those locked priors and report how well they matched.
Keep the methodology, the data, the scores, and the validation as inspectable
files (e.g. a methodology writeup, a data file, a scoring script, a scores file).
Once produced, the artifact must open and work by double-clicking index.html —
no build step, no local server, no API keys, nothing to install. Loading a mapping
library and map tiles from a public CDN is expected and fine; "no API keys" means
no paid or credentialed services, not no network.
Answer
Checks
7/8Reproducible scoringreproduciblethis task
the scores file is deleted in a scratch copy and the scoring script re-run: it has to finish inside 180s and write the scores back. Missing grader prerequisites fail.
No AI-tell wordingai-wordingevery task
deliverable text contains none of the shared AI-tell phrases in bench/ai-tells.txt. Task checks may add narrower contractual bans, but must not repeat this shared floor.
inspected index.html, hidden_gem/prereg_amendment_2.json, hidden_gem/priors.json, hidden_gem/01_PREREGISTRATION.md, hidden_gem/prereg_amendment_1.json, hidden_gem/data/hoboken_boundary.json, hidden_gem/data/probe_log.txt, hidden_gem/data/label_articles.json
Map loadsrendersthis task
a visible map container of at least 300x200 with >= 5 markers, in a headless browser. Cannot pass without Playwright installed.
Required categoriescategoriesthis task
all six categories the prompt asks for are named somewhere in the output: playgrounds, food, activities, day-trips, date-night, tapas.
Required filesartifactsthis task
the run left all four of a methodology write-up, a data file, a scoring script and a scores file.
Q1Easy for family to usethis task
Judge's reasoning
index.html is a single self-contained Leaflet page (CDN CSS/JS + public tiles) that opens by double-click and shows a searchable sidebar of named places with labeled category chips and pins — no build, server, or raw-data dumps (CARTO tiles carry an 'API KEY REQUIRED' watermark, but streets, labels and controls remain readable and usable).
▸Rubric
Usable by a non-technical family member. Look at `map.png` and `full.png`, and at how `index.html` is meant to be opened. The bar, in the reviewer's words: "It's a single self-contained interactive map artifact, uh, that someone could open and use basically by clicking on the HTML." FAIL if using this would require technical knowledge — a build step, a local server, a terminal command, editing a file, or an interface that shows raw data (JSON blobs, coordinate dumps, unlabeled controls) instead of named places and readable categories. PASS if double-clicking the file gives someone a map with labeled places and controls they could figure out unaided.
Q2Criteria fixed before scoringthis task
Judge's reasoning
01_PREREGISTRATION.md and priors.json state named expect_high/expect_low places, hypotheses H1–H4 with thresholds and descriptive predictions P1–P5, all SHA-256 hash-locked in prereg.lock before any rating data existed (venues.csv is empty and the artifact honestly reports the scoring as 'not yet run'), so the priors demonstrably could not have been written from results.
▸Rubric
Priors were locked before scoring. Read the methodology writeup and the scores/validation files. The prompt asked it to "**Pre-register** a methodology and your priors/expectations *before* collecting and scoring data, then test the computed scores against those locked priors and report how well they matched." FAIL if the priors read as written after the results were known — signals: no distinct set of priors stated separately from the findings; priors phrased in terms of the computed scores rather than as expectations about places or patterns; a "validation" section that restates the scores instead of describing a test the priors could have failed. PASS if there is a stated set of expectations that came first and a reported comparison against them, whether they held up or not.
Q3Real placesthis task
Judge's reasoning
Spot-checked Fiore's House of Quality (414 Adams St), Sybil's Cave, Hoboken Historical Museum (1301 Hudson St), Mompou (77 Ferry St, Newark) and Liberty Science Center — all real places with addresses and coordinates that land where they actually are.
▸Rubric
Places are real. Pick five places from the map — spread across categories, including at least one the output calls a hidden gem — and check them against what you know about Hoboken and the surrounding area. The bar, in the reviewer's words: "real plausibly correct Hoboken area places, not invented ones." FAIL if any of the five is one you can positively identify as fabricated or misplaced — a business that does not exist, a name attached to the wrong town, or coordinates that put it far from where it actually is. If you cannot positively identify a problem with any of the five, PASS.
Notes
1Okay, it looks like it's done a really good job on the Hoboken map. Actually, it's quite a clean design. It's reminiscent of Fable, but it looks like this is relatively much cheaper than Fable, which is exciting. I really like the way it looks. It messed up the scoring a little bit. It deleted a file that would have let us judge whether it was scoring correctly based on its rubric. So that was a little bit weird, but otherwise I think it did a good job, and it didn't hallucinate any fake places or places that were closed. And yeah, did a like I was pretty happy with the design. It picked out a few of the good spots that I like. So yeah, pretty happy with this one. I think this is a real win. It might be one of my favorite ones so far actually.
21 Sep 2026

