Heft

A daily puzzle. A hanging mobile with empty hooks and a rack of weights — fill every hook so the whole thing hangs level. You never learn what anything weighs.

Live: heftdaily.com · Build: 1.8.0 · the daily · Boards: 144, all six hooks

How it works

Each arm's angle is directly proportional to its fractional torque imbalance — the tilt is the arithmetic. The mobile hangs from the rule under the header, which is its ceiling and lights when the board comes level. An arm can be level with the wrong weights on it, but then its subtree has the wrong total mass, so the arm above must lean. Nothing in the interface locks, hints, or indicates correctness.

Placing a weight is free. Replacing one costs a swap. Zero is perfect.

The weights are ordered and the rack is sorted to match: diamond, triangle, hexagon, pentagon, house, lightest to heaviest, on every board. Nothing states this — the glyphs share one height and their ink areas climb, so it reads as apparent weight. You still have to hang something to learn the ratios, which are the puzzle.

Five weights on six hooks is the heavy rung, shipped alone for feedback. The four-weight bank it replaced (1.5.12, 96 boards at five hooks and 48 at six) is the easy rung and comes back beside it when the daily is built.

Repo

• index.html — the game. ~33 KB of code, no build step. Vercel serves it as-is.

• boards.js — the 144-board bank, one board per line, loaded before the game script.

• tools/clearance.py — hooks apart and nothing crossing, in every reachable state. Applied by the generator; select.py runs it on any pool that skipped it.

• tools/select.py — picks a bank from a pool by score: forks count, repeats of a shape are discounted, lopsided boards are capped.

• tools/icon.py — draws icon-2.png, the icon a link preview shows.

• docs/ — decision record, project plan, generation spec, publishing notes.

• vercel.json — one header rule: Cache-Control: public, max-age=0, must-revalidate on every path, so a deploy is never served from a stale cache. This is what stops the two-file trap below from biting in practice.

• tools/generate.py — the board generator. Python 3, standard library only, no dependencies.

Earlier docs said generation lived in a separate private repo. It never did — no such repo exists, and the code that produced the shipped 144 was not kept. tools/generate.py was written fresh against docs/heft-generation.md; run python3 tools/generate.py --verify to check its property code against boards.js before trusting it on new boards. A 200k-draw run at six hooks and five weights takes about 35 minutes with the clearance check and keeps roughly one board in 650 draws.

The bank was split out of index.html in 1.4.0. It had been a single 42 KB line — 63% of the file — which made every one-line edit expensive and every diff unreadable. The game reads PUZZLES as a global, so the load order matters and there is no module system involved.

URL parameters

The default experience is the daily: a date-seeded board (No 1 = 15 Sept 2026), saved in localStorage with a streak, that everyone plays in common. The parameters below are testing affordances; none changes the daily, and the pins write and log nothing.

parameter	default	effect
?practice	—	random boards on the live build, with a Next button; writes and logs nothing
?board=N	—	pins puzzle N (1–144); a testing pin, writes nothing
?day=N	—	pins day number N; a testing pin, writes nothing
?tilt=N	1	raises the imbalance to the power N before mapping to angle; 1 is proportional
?bow=N	0	arm curve as a fraction of arm length. Straight since 1.5.5; 0.025 restores the old sag
?ink=N	1	lifts the wire off the background — colour and stroke weight; 0 is the 1.4.7 drawing, 2 is the ceiling
?floor=N	1.25	minimum lean in degrees for an arm that is not level; 0 restores the 1.5.2 drawing
?tiltmax=N	15	the tilt range in degrees, 8–30. Widening reduces invisible swaps but risks the staircase 30° was rejected for

A non-default tilt, bow, ink, floor or tiltmax prints next to the build number, so a screenshot always says which version produced it.

Deploying

Commit to the default branch; Vercel publishes automatically.

Bump the build string in the same commit — it lives in three places:

1. const BUILD in the script
2. data-build on the <html> element
3. an HTML comment near the top of <head>

The last two exist so the version can be read without running the page — by curl, by view-source, by anything that strips scripts. The script checks all three agree and prints · BUILD MISMATCH next to the build number if they don't, so a half-done bump announces itself rather than lying. The build number renders under the buttons. Never reuse one: 1.3.x was skipped deliberately.

Verify in a private tab, since browsers cache hard enough that a stale copy looks like a failed deploy. With two files there is a new version of that trap — a cached boards.js against a fresh index.html looks like a working game with the wrong bank, and nothing on screen says so.