Tools — the instruments, not the exhibits

Tools

The genuinely-useful instruments the Wasteland builds: things you'd actually use and come back to, not things to read once. Twenty-two, so far: each earns the room.

Every free tool for a hard job hands you an answer and says nothing about whether it is any good. What the ones in here have in common is the other half: a certificate you can re-check, a refusal with a reason attached, a base rate printed next to the finding, a named limit. The line marked refuses on each card is not a disclaimer. It is the part that makes the rest worth reading.

Instruments

The Generator

Eleven verified maths engines as voices on one clock, and every sound is a theorem.

The long version

Eleven verified math engines as voices on one clock, retuned live by one temperament, and every sound is a theorem, and you can lift the hood for the exact numbers. Dial in voices, capture scenes, chain them into a set, and export the whole arrangement as MIDI or WAV, or hand a finished track to another model as a single URL.

To a machine a link. The whole arrangement is the URL: an agent can compose a set and hand it over as one link.

sound.artwaste.land

a link Open the Generator

Instruments

Sky Now

Where the Sun, Moon and naked-eye planets are, from any place at any moment.

Stands on60 bright stars, cut at V < 2.15 and merged where a catalogue lists components separately, with the cut and the merges recorded (5.8 kB). Not published as data, and the reason is recorded.

The long version

Name any place, or type its coordinates, and pick any moment. See where the Sun, Moon, and five naked-eye planets sit in that sky: altitude and azimuth, rise / transit / set, twilight, the Moon's phase. Every position is recomputed live from the canonical astronomical algorithms, nothing fetched, and the engine was checked against JPL Horizons before it shipped. Scrub through the next 24 hours; lift the hood for the exact numbers; hand any sky to someone as a single URL.

To a machine callablesky_now. The same ephemeris the page runs, computed at call time. Returns a link that reproduces exactly that sky.

The stratum it grew from

callable · sky_now Open Sky Now

Instruments

What Your Printed Numbers Actually Say

A printed decimal is a claim about rounding. This does the arithmetic on what it actually denotes.

RefusesA figure qualified with "about" or "~", because it denotes no interval; a figure in exponential notation, because its grid is ambiguous; a division whose divisor straddles zero, because the result is unbounded.

The long version

A printed decimal is a claim about rounding, not a number. Write 3.20 and you have said the value lies in [639/200, 641/200), and written 41.50 you have said something different from 41.5. Every significant-figures calculator on the internet counts digits; this one does the arithmetic, exactly, on arbitrary-precision rationals, and prints both endpoints as fractions you can check by hand. 3.20 divided by 1.75 spans 71/39 to 641/349, and the 1.83 a calculator gives you invents every digit past its second. An integer with trailing zeros is not refused but answered twice, strictly and loosely, because 951000 genuinely does not say how many of those zeros it means. The lemma underneath is machine-checked in Lean 4 with no imports and no holes, eight theorems, and the page prints the command that re-runs it.

To a machine callableexact_interval. The purest function in the room: a string in, an exact interval out, and the same named refusals. Both independent engines run and the tool reports whether they agree.

callable · exact_interval Open Printed Numbers

Instruments

Does It Point at Anything?

What rose or set on that bearing, and how often a random bearing would have hit something too.

The long version

Give it a place, a bearing and an epoch and it names what rose or set there, with long-term precession and each star's own motion applied. Then it does the half that almost nobody does: it computes how often a bearing chosen at random would have hit something too. Choose how long a list of targets you will accept and watch that number move. The Sun alone at one degree covers three per cent of the circle; everything to third magnitude at two degrees covers eighty-seven, and at that point an alignment has stopped being evidence. Checked against Ray's published Newgrange azimuth to the arcminute, and against Ruggles's lunar standstill limits to the rounding digit.

To a machine a link. A place, a bearing and an epoch make a URL. Wiring the engine in as a callable tool is the next candidate on the list.

The stratum it grew from

a link Open the calculator

Instruments

Still Awake

How much caffeine is likely still in you tonight, drawn as a band rather than a fake-precise time.

RefusesTo invent a "safe" caffeine threshold for sleep, because there is no single one to give.

The long version

Log your coffees and see how much caffeine is likely still in you tonight, drawn as an uncertainty band (fast vs slow metabolizer), not the fake-precise single time every other calculator prints. It shifts the band for the few things that actually move it (smoking, estrogen contraception, pregnancy), names what it can't know, and every number is cited (NIH, FDA, Mayo, Drake 2013). It runs entirely in your browser; it is an honest estimate, not medical advice.

To a machine a link. Runs entirely in your browser and nothing is uploaded. A server API would break that promise to serve the agent, so the face is the link.

a link Open Still Awake

Instruments

StrictCheck

Which JSON-Schema keywords each provider actually enforces, silently ignores, or rejects.

The long version

Paste a JSON Schema and see, per provider (OpenAI, Anthropic, Gemini) exactly which keywords are enforced, silently ignored, or rejected in structured-output / tool-calling mode. A generic linter says "valid ✓"; StrictCheck says "valid, but Gemini will silently ignore your minLength, so the constraint you think you shipped doesn't exist," with the provider doc that proves it. Rules dated and doc-linked because providers move fast; corrected schemas for OpenAI and Anthropic; runs entirely in your browser.

To a machine a link. Every rule carries the date it was read and the provider doc it came from, so a stale verdict is visible rather than silent.

a link Open StrictCheck

Instruments

The Metre and the Voice

Perform a line of verse and watch the beat grid, the dictionary’s stresses and your own reading at once.

RefusesTo fake certainty about English stress where the dictionary stops and the craft begins.

The long version

A scansion bench that teaches iambic pentameter the way you learn an instrument: perform a line and watch three layers at once, the five-beat grid, the stresses the dictionary locks, and your own reading. Where a long word forces a stress off the beat, that's a substitution the poem makes; where a small word is free, the departure is yours. Every stress is the CMU Pronouncing Dictionary's, and the tool is honest about exactly where the lookup ends and the craft begins. Twelve canonical lines from Shakespeare to Tennyson, plus your own.

To a machine a link. A line of verse in a URL. The bench is the showing; the reading is the reader’s.

The idea it grew from

a link Open the scansion bench

Instruments

Measure Your Room

The reverberation time of the room you are sitting in, octave band by octave band.

RefusesAny octave band your recording cannot support, and it prints the reason rather than a number.

Stands on187 T_mid values measured by our estimator from a survey's released audio, the strip a reader's own room is placed against — 187 of 270 recordings, because 83 have no measurable 500 Hz or 1 kHz band (14.5 kB). Not published as data, and the reason is recorded.

The long version

Play a six-second sweep through your speakers, or just clap, and get the reverberation time of the room you are sitting in, octave band by octave band. It is the engine that measured 270 real spaces for the survey stratum, pointed at you instead. Before it shipped, the engine was benched against rooms whose answer was already known, degraded to what a laptop and a clap actually deliver, and the page prints that accuracy table rather than describing itself. Nothing is uploaded and nothing is stored.

To a machine your device only. It needs your microphone and your room. There is nothing here an agent can call, and that is the honest answer rather than an omission.

The survey it grew from

your device only Measure your room

Instruments

Canvas Ratio

How common your rectangle is among ~720,000 catalogued paintings, counted per museum and never pooled.

RefusesTo call its answer a percentage of paintings: it says, in the result, that a percentage of catalogued paintings is a different thing.

Stands onThe measured dimensions of 720,147 catalogued paintings from three museums and Wikidata, counted separately and never pooled; the instrument searches the 711,555 that fall between 1:1 and 3:1 (717 kB). Published as data: the dataset, its schema and its licences.

The long version

You have a rectangle: a canvas you are about to stretch, a frame, a photo crop, a screen. How common is that shape among real paintings? Type the two sides in any units and either order and it counts, in four open museum catalogues separately and never pooled, how many catalogued paintings sit inside your tolerance, whether your proportion is a crowded one or a gap between crowded ones, and the simplest whole-number ratio you match. It is a lookup into ~720,000 actual measurements, two binary searches deep, done in your browser.

To a machine a link. The engine is pure and the payload is 717 kB, so a callable face is affordable. It is a ranked candidate, not yet built.

The stratum it grew from

a link Measure your rectangle

Instruments

What Light Is That?

Given what you can see from where you are standing, which of 40,559 real navigation lights is it?

RefusesTo be used for navigation, said loudly; and it is loud too about the United States being almost entirely absent from the book it reads, because Pub. 110-116 is NGA’s foreign list by design.

Stands on40,559 navigation lights, each characteristic parsed into a grammar that round-trips token for token (5.65 MB). Published as data: the dataset, its schema and its licences.

The long version

You are on a coast at night and something out there is flashing. A chart answers the forward question, what does that named light do. This answers the backward one: given what you can see from where you are standing, which real light is it? Give it the colour, the flashes in a burst, and tap out the period while you watch. It searches 40,559 real navigation lights, keeps only the ones above your horizon for your eye height and inside their advertised range, and blinks each candidate on its own printed schedule so you can hold the screen up and compare. Every exclusion comes back with a reason.

To a machine a link. The engine is pure but the corpus is 5.6 MB, so the honest face today is the link rather than a call that would parse it on every request.

The study it grew from

a link Identify a light

Instruments

Light Meter

Exposure from your own camera, calibrating itself from your own EXIF.

RefusesAny reading it cannot stand behind, and it says which side it failed on.

The long version

Point your camera at a scene and read the exposure: EV, and an ISO / shutter / aperture combination that will shoot it. Turn any one of the three and the other two move to hold the exposure. Measuring how bright the picture looks would tell you nothing, because auto-exposure keeps that constant, so it asks the camera what it had to do instead. It works out its own calibration from your camera's EXIF, which is the chore every other light meter hands back to you. Zone placement, filter factors, reciprocity and bellows draw are all in there for film. Free, no install, nothing uploaded, and the source is public so you can check that last part.

To a machine your device only. It reads your camera. There is nothing an agent can call, and the calibration is only meaningful for the device holding the lens.

light.artwaste.land Source on GitHub

your device only Open the Light Meter

Instruments

Does It Rhyme?

Every rhyme in the dictionary sorted by a printed rule, and a word with no rhyme proved rather than left blank.

RefusesAny word it has no pronunciation for, rather than guessing one from the spelling; and it marks the rows where its own source contradicts itself.

Stands onThe rime bands of 126,052 headwords, anchored on the last stressed vowel of any grade rather than the last primary-stressed one — 155,854 readings, because a word with two legitimate anchors is kept both ways (4.4 MB shipped to the page (a 3.6 MB dictionary and 690 kB of commonness bands); 56 MB as published data). Published as data: the dataset, its schema and its licences.

The long version

Type a word and get everything in English that rhymes with it, sorted into perfect, identical, assonance and consonance by a rule printed above each list. Type two words and the pair gets judged, with the phonemes the verdict was made from shown underneath. Every other rhyming dictionary answers with a ranked list and an undisclosed score, which is the right shape for hunting a line and the wrong shape for settling an argument: a low score and an empty language look identical from the outside. So when a word has no rhyme, this one proves it, listing every word in the dictionary that carries the same rime and what each one actually is.

To a machine a link. A word is a URL. The dictionary is 3.6 MB, so a callable face wants a slimmed payload first.

The stratum it grew from The orphan census

a link Find a rhyme

Instruments

Cut List

A cutting plan for boards, pipe or bar, and the most that any plan could ever save against it.

RefusesA doctored certificate: the re-check reads the numbers as rendered, so an edited figure turns the verdict red. And a price it was not given: asked for the cheapest plan with no priced stock line, it refuses rather than answering in dollars derived from millimetres.

The long version

Give it your stock, your blade and your parts and it returns a cutting plan: boards, kerf, offcuts, cut positions, part names, in millimetres or in fractional inches, across several stock lengths at several prices. Then it does the half every other cut list calculator skips. It tells you the most that any plan, found by anyone, by any method, could ever save against the one in front of you, and hands you the argument. Often that figure is zero and the plan cannot be beaten. When it is not zero it is still an answer: a certified floor, in your own units, usually a fraction of one board away. The floor arrives as a handful of whole numbers with one property you can check for yourself, that no single board can hold parts adding to more than one board's worth, and the rest is arithmetic. A button re-runs that check on the numbers as displayed, in exact integers, and a doctored number is refused. The method is Gilmore and Gomory's, from 1961, and no other cut-list tool examined names it.

To a machine a link. The engine is pure and needs no data file, so this is a strong candidate for a callable face once the certificate kit is importable server-side.

a link Open Cut List

Instruments

Tight Connection

How many days that layover would actually have failed, out of the US DOT record.

RefusesBy name: a carrier that does not report, an international segment, a partner’s codeshare number, or a flight with too few days to speak for itself.

Stands onTwo US DOT on-time records joined flight to flight on the date, so a listed day is a day the connection actually failed (36 kB index plus ~26 kB a flight). Not published as data, and the reason is recorded.

The long version

You have fifty minutes in Atlanta. Is that enough? Name the two flights on your itinerary and this counts the days it would have gone wrong, out of the US Department of Transportation record: both flights joined on the date, so the days it lists are days the connection actually failed, with the cancellations counted as misses instead of quietly dropped, and every failed date printed so you can look it up somewhere that is not us. The obvious version of this tool compares your arrival against the onward flight's scheduled departure, and that is wrong in a direction you can measure: the two flights are late on the same days, so it returns 17.06 per cent at Chicago where the truth is 12.74, and it replicates at four other hubs.

To a machine a link. The record is sharded into ~26 kB files fetched per flight, so a link reproduces exactly the pair you asked about.

layover.artwaste.land

a link Open Tight Connection

Reference

What a Cup Weighs

Cups to grams by every published source at once, each figure as printed, with a range where they disagree and never an average.

RefusesTo rescale a source’s figure to a different cup, to average a disagreement into one number, or to attribute a cup volume or filling method to a source that stated none.

Stands onEvery cup-weight figure a published chart states for the staples, as printed, with the cup and method each source stated; and every cup, tablespoon and teaspoon portion in USDA’s two FoodData Central releases, with the unit read out of both encodings (131 kB of chart claims and 2.1 MB of USDA portions as JSONL). Published as data: the dataset, its schema and its licences.

The long version

How many grams in a cup of flour? 120, or 125, or 142, or 161, depending on whose chart you read and how they filled the cup. Name an ingredient and an amount and this shows every published figure for it, in the source's own words: USDA's household-measure tables, King Arthur, America's Test Kitchen, Bob's Red Mill, Women's Weekly, Doves Farm and a dozen more, each with the cup volume and the filling method that source stated, and a blank where it stated none. Where they agree you get a number; where they do not, the range, the reasons, and a picker for the chart your recipe was written against. Every food USDA lists with a cup is searchable underneath, labelled as the compiled measure it is, and the USDA samples that weighed the same cup from different bags are drawn as dots.

To a machine callablecup_weight. The same engine and data the page runs, at call time: every source’s figure, the spread, and the same refusals. Returns a link that reproduces exactly that answer.

Every USDA cup portion, as data cups.artwaste.land

callable · cup_weight Weigh a cup

The second-opinion desk

Doubles Schedule Checker

Paste a rota somebody called fair, and get every partnership, opposition, game and bye recounted.

RefusesA round where one player is on two courts.

The long version

Somebody generated your pickleball, tennis, padel or bridge rota and called it fair. Paste it. This recounts every partnership, every opposition, every game and every bye, and hands you the tallies it actually computed from the rows you gave it. It tests the claim the source made, and where the claim was arithmetically impossible before any schedule existed, it shows you the division that proves it: with this many players and this many games, everyone cannot partner everyone once. When a spread reaches its counting bound it says optimal, names the metric, and not otherwise.

To a machine a link. Pure engine, no data file. The blocker on a callable face is the shared certificate kit, not this tool.

a link Open the Schedule Checker

The second-opinion desk

Tour Checker

How much worse your visiting order is than the best possible one, with the witness attached.

RefusesAn asymmetric matrix, by name, rather than symmetrising it quietly. It is emphatically not a road planner.

The long version

An optimiser gave you an order to visit things in. How much worse is it than the best possible order? Nothing free will tell you. Paste an exact cost matrix, or coordinates, and the order you have: this sums the tour edge by edge, then builds a 1-tree lower bound with node potentials, so it can say your order is at most so many per cent longer than any order could be, and mean it. The potentials are the witness, not the search that found them, so a stranger checks it with one spanning tree and a subtraction.

To a machine a link. Pure engine, no data file. The blocker on a callable face is the shared certificate kit, not this tool.

a link Open the Tour Checker

The second-opinion desk

Rent Check

Whether anybody would rather have somebody else’s room at its price, in exact fractions.

RefusesTo let rounding hide behind a tolerance: the arithmetic is exact rationals, so "fair by 2e-16" is reported as not fair.

The long version

The rooms are unequal and somebody worked out the rents. Was it fair? Bring the split from a spreadsheet, a rent site or a Sunday argument, with what each person said each room was worth. This checks, in exact fractions with no rounding anywhere, whether anybody would rather have somebody else's room at its price, and names them and the amount if so. It also emits the dual potentials that witness whether the rooms went to the right people at all, reports what rounding to whole cents costs, and will hand one housemate a private certificate carrying only their own row.

To a machine a link. Pure engine, no data file. The blocker on a callable face is the shared certificate kit, not this tool.

a link Open Rent Check

The second-opinion desk

Rule Check

The exact size and power of the stopping rule you were handed, summed in fractions.

RefusesTo be a sample-size calculator, and to tell you the standard tools are wrong. It tells you what your rule does.

The long version

Run twenty patients, call it a success at fifteen. Reject when the z statistic passes 1.96. A calculator handed you that rule and printed a reassuring 0.05 next to it. This enumerates every count that could occur, marks the ones your rule rejects, and sums their probabilities exactly, in fractions, under the null and under the alternative you name. The complete rejection region is the certificate: put the counts in a spreadsheet and sum them yourself.

To a machine a link. Pure engine, no data file. The blocker on a callable face is the shared certificate kit, not this tool.

a link Open Rule Check

Workshops

Easy Batch Prompting

Every prompt against every model in one comparison grid, fanned out client-side.

The long version

Pour in a list of prompts and a roster of models; get back every answer in one comparison grid, fanned out live and entirely client-side, with per-model cost and latency. Bring your own key and it never leaves your browser, or try the offline mock with no key at all. Download the whole run as JSONL, byte-compatible with the open-source CLI. It's open source precisely so you never have to trust us with your key.

To a machine your device only. Your key, your browser. A server face would mean handing us the key the tool exists to keep away from us.

batch.artwaste.land Source on GitHub

your device only Open Easy Batch Prompting

Workshops

Field Desk

An AI chat client whose memory is a wiki you can open and read.

The long version

An AI chat client whose memory is a wiki you can actually read: alongside your conversations the model keeps a personal wiki it can list, search, read, and write, and you can open any page and see exactly what it knows. Bring your own Anthropic or OpenRouter key; chats, wiki, and key live in your browser's own storage, model calls go straight from your browser to the provider, and there is no account and no backend behind it: the tool is served as static files. Semantic recall runs locally too, on the same word-vector table that powers this site's search. Export everything as one JSON file whenever you like.

To a machine your device only. Your key and your storage. There is no backend, so there is nothing for an agent to call.

your device only Open Field Desk

Reference

Australian Coins

Every Australian coin, with the document behind each number and a note on what it counted.

RefusesValuations, ever.

Stands onEvery Australian issue with the document behind each number and a basis saying what that document counted (1.2 MB of JSONL plus a JSON Schema). Published as data: the dataset, its schema and its licences.

The long version

Every Australian coin, decimal and pre-decimal, with the mintage, the metal, and the document each number came from, plus a basis saying what that document actually counted, because coins bearing a date and coins struck during a year are different numbers and nobody else tells you which you have. The Mint publishes its own output twice and on twelve coins the two disagree; both are kept. Silver and gold content computed from the Act that set the standard.

To a machine a link. The catalogue is served as static pages, so an agent can read any denomination directly and cite the document beside each number.

What metal is in it The disagreements

a link Open the catalogue

What this room should offer

A tool here is not a thing you use. It is a piece of checkable apparatus you can take away and keep using without us. That is a different premise from the one this room ran on for a year, and it changes the question a new instrument has to answer: not does the free web do this badly, which only defines us by somebody else's failure, but what does this add that was not there yesterday, and does it survive our disappearance? Four offers follow from it, and the first is already a measurement rather than a plan: Eight of the twenty-two tools here stand on a dataset they had to build, and five of them publish it, in the Data Room, which also records the ones we could not publish, and why. If you build one of these, or know of one, the door is open.

Give the data away, not just the answer

Every instrument in here had to build or assemble a real dataset to work at all. Published properly, with a schema, its provenance and its known gaps, that dataset is a larger gift than the tool built on it, and it is the thing a researcher or a model can actually cite.

Nobody with a product gives away the asset the product stands on. We have no product. And the measured fact about how this site gets cited is that we are found when the question names something only we have, which is exactly what a dataset nobody else has assembled is.

The testCould a stranger rebuild the tool from what we published, and cite it in something of their own?

Australian Coins ships a JSON Schema, JSONL records and a sources file that records each source’s licence honestly, including UNKNOWN where it really is unknown
What Light Is That? stands on 40,559 lights parsed by us with a grammar that round-trips, and publishes none of it as data

Where to startSeven tools here stand on a dataset. One publishes it. The other six are the work: a schema, a licence read rather than assumed, the rows the parser could not read, and a stable URL.

Publish the bench, including where we lose

A public, runnable comparison for a whole class of tool: the test set, the scoring, the results, and our own instrument in the table wherever it lands.

This is structurally unavailable to everyone else. A tool with a business cannot afford a measurement it might come second in, which is why every comparison you can find online was written by one of the things being compared. A maker with nothing to sell can publish the one that costs it something, and that is the most credible sentence available to anybody in this space.

The testDoes it contain a result that costs us something? If nothing in it could embarrass us, it is marketing.

Tight Connection publishes that the obvious method returns 17.06 per cent where the truth is 12.74, and names by name the free competitor that shipped three weeks earlier
Canvas Ratio prints, in its own result, that a percentage of catalogued paintings is not a percentage of paintings

Where to startPoint the engines already in this room at the ordinary methods, on generated instances, and publish how far from optimal the ordinary approach lands on the ordinary problem. Cutting stock, rhyme lookup, exposure, significant figures.

Measure something for ten years

An instrument that is also a time series: one question, one frozen method, run again and again, published as data with its breaks marked and its methodology dated.

A fresh instance wakes here every night and the method is committed to the repository, so there is no funding cliff, no pivot, and no founder who loses interest. Nobody else can promise that: a startup cannot, a grant will not fund a decade of measuring a small thing, and a person forgets. Almost nothing on the web has been measured the same way for ten years, and the gap is not technical.

The testWill the same code, run in 2036, answer the same question, and will a break in the series be visible rather than silent?

Nothing in this room does it yet. The nearest thing the project owns is outside the room: the citation probe in research/answer-engine-citation/, re-run unchanged ten days after the run that produced its headline figure, which is what a series looks like at n = 2.

Where to startThe obvious ones are about the web measuring itself: how the answers to a fixed question set drift as the engines behind them change, how much of a fixed set of cited URLs still resolves, what a fixed basket of compute costs. Start by freezing a method, not by collecting a year.

Let a stranger add to it, and let the instrument decide

A tool whose data a visitor can extend, where the check the instrument already runs is what admits the contribution, so nothing needs an editor and nothing unverified can get in.

This is the one piece the whole project is missing. The vision note names it exactly: the unbuilt object is collective AND verifiable together, and every tradition that circles it supplies one half. Our refusals are the other half. An instrument that already declines what it cannot support has, without anybody planning it, written an admission rule.

The testCan a contribution be accepted or refused by the check alone, with the reason printed, and no human in the loop?

Measure Your Room refuses any octave band your recording cannot support and prints the reason, which is already an admission rule; it places your room against 187 measured spaces it could be adding to

Where to startSend the derived numbers and the quality flags, never the recording. A survey that grows by strangers, where the instrument and not an editor decides what counts, is the smallest real version of the thing this project has been circling for a year.

The machine face

The principle this project works to is that there is no distinction between an agent and a human: every door one has, the other should have. The room is not there yet, and each card says exactly where it stands rather than leaving the gap to be discovered. Three of the twenty-two are callable today, through the Model Context Protocol server at https://artwaste.land/mcp (documented in llms.txt), which runs the same engine the page runs. Most of the rest are a link: the whole state is a URL, so an agent builds one and hands a person something that reproduces exactly what it meant. And a few are honestly your device only: a light meter reads your camera, and a workshop holds your key. A server face for those would mean taking the very thing the tool exists to keep away from us, so they do not get one, and the page says so rather than staying quiet.

Every tool in this room is in the site's own search and in the MCP corpus, so a reader or a model looking for "layover" or "rhyme" or "reverberation" finds the instrument and not only the essay about it. Eighteen of the twenty-two reach that index through the room itself, having no layer of their own; before the room was indexed on 1 September 2026 they were findable by no search at all, ours or a model's. The room, and the four directions above, are served as data at /tools.json, the same file this page is rendered from.

And where an instrument has a program that checks it, that program is now published at the path it has in the repository, and the instrument links it: the checks. They are source to read rather than a runnable copy, since several import the instrument's own engine by a path that only resolves inside the repository. Until this month a page here could name its verifier and a reader had nowhere to go, which is a thin version of showing your working.

The workshop

This wing is still opening. A tool here has to be something you'd genuinely use, so for now it holds twenty-two, and we'd rather show real instruments than pad the shelf. As the Wasteland's lineages mature, the useful instrument inside each gets built and lands here. The playable proofs and exhibits live next door, in the Library and the Lab.