fejkdata/todo.md

Plan

Ordered as the releases that ship it.

v0.1.0

Comprehension, until the panel scores 7.0

Nothing else ships while the score is under 7.0 — no feature, no category, no data — bar what AGENTS.md excepts, per README goal 2. Nine rounds of the nine-seat findings held the score between 5.7 and 6.1, so this round is a restructure the maintainer approved on 2026-09-29: a PR per item, then the nine-seat run again. The run is the nine-seat one, since the four-seat scoring run names too little to steer an architectural change.

Scores, newest last: 5.9 on 2026-09-20 (Navigation 7.0, Locality 5.2, Shape 5.9, Self-sufficiency 5.7); 5.8 on 2026-09-22 (Navigation 6.7, Locality 5.0, Shape 5.9, Self-sufficiency 5.6); 5.7 on 2026-09-26 (Navigation 6.7, Locality 5.0, Shape 5.8, Self-sufficiency 5.6), every seat capped by Locality, and all nine naming drawWalk.walk among the hardest; 5.7 on 2026-09-28 (Navigation 6.7, Locality 5.1, Shape 5.8, Self-sufficiency 5.7), every seat capped by Locality, all nine naming drawWalk.walk and the four path walks among the hardest; 5.8 on 2026-09-29 (Navigation 6.8, Locality 5.3, Shape 5.8, Self-sufficiency 5.7), seven seats capped by Locality, all nine naming drawWalk.walk hardest and eight the unit they would least modify; 5.7 on 2026-09-29 again (Navigation 6.6, Locality 5.2, Shape 5.9, Self-sufficiency 5.8), four seats capped by Locality, six of the seven unit seats ranking readSurvey.walk hardest and all seven the unit they would least modify; 6.1 on 2026-09-29 a third time (Navigation 7.0, Locality 5.6, Shape 6.3, Self-sufficiency 6.1), every seat's lowest dimension Locality, all seven unit seats ranking the readSurvey walk or check hardest and the unit they would least modify; 5.8 on 2026-09-29 a fourth time (Navigation 6.7, Locality 5.0, Shape 5.7, Self-sufficiency 5.7), every seat's lowest dimension Locality, alone or tied, all seven unit seats ranking readSurvey.walk and edge hardest and the unit they would least modify; 5.8 on 2026-09-29 a fifth time (Navigation 6.9, Locality 5.2, Shape 6.1, Self-sufficiency 5.8), every seat's lowest dimension Locality, alone or tied, all seven unit seats ranking readSurvey.walk hardest and the unit they would least modify.

This round, in order:

• Run the nine-seat panel as a checkpoint, and file its scores above: where Locality has not risen from 5.2, stop and bring the scores to the maintainer before the path steps.
• Compile each path once at link into steps holding no pointer, with its leaves and cover, and have renderEdges, heldNodes, valueProof and tableReadOf read them, deleting pathCover, coverPath and pathLeaves. Prototyped green, allocations identical.
• Draw every path through one interpreter over those steps, drawSteps, with descend probing into a stack buffer, deleting pathDraw, and replace the decision "The four path walks are four loops": the prototype held perf_test.go once a step held no pointer, the probe returned its slice and the interpreter took levels.
• Fold checkNoOverlap and checkNoRepeatedRead into heldCheck, so compileFormat compiles and nothing more.
• Fix at compile how readField reads an arm, arm.kind, deleting its two panics and its held lookup per read.
• Switch disagreement on the fixed item's kind, with the other item inside each arm, and rename its proven local, which shadows the proven type: seven seats read the seven-arm switch as a truth table without the table.
• Say at checkNodeFences that its walks recurse unguarded, so a caller runs checkNoCycles first: five seats reconstructed the order, and the inline binding runs no cycle check.
• Give tableSite an isFormat beside isCell, so renderRecord and traceRead stop comparing row with formatRow alone, which holds only because a template outside a table has row 0: four seats found the invariant unstated.
• Move valueProof.checkField to valueproof.go, and rename checkPath, which differs from the pathCheck type it wraps by word order alone: both architects placed them wrong.
• Correct the claims readers followed wrong: IsTemplate's doc says an argument holding a bracket that is not valid JSON errors, where misc.territory[SE] is a path; the README says rows are indexed on the first draw, where New builds the key map; and loadDir's says it builds the tree deterministically, where it reports an unnamed rows file in map order.
• Delete the comments the panel found restating their code or arranging the file: node's second sentence, op's "compile builds these so render never re-scans the format", checkCalc's last sentence, calcVar.eval's four lines, checkNoCycles' grey and black, and the docs on expand, readRows, splitOutside, reservedList and writesRecords.
• Delete doc.go's "The draw fences are …" sentence, which misses checkDrawGroup, checkOwnFamily and checkNestedDrawGroup; the README's Layout indexes the fences by file. Say "fences" in the pin entry then, which names the draw fences too.
• Run the nine-seat panel again and file what it names here.

Data

Shape: T = table, t = template, c = choice. Read every factual list from a register by a data-import/ script, as README goal 12 asks.

• Fetch NVDB through source.fetch in data-import/geo-se.py: nvdb_segments calls urllib itself, with no retry and no User-Agent.
• Test that every geo tree holds the five table names and the address columns the README says port across countries: nothing checks it, and a new country breaks it silently.
• Strip the whitespace data-import/territory.py copies from its register: misc.territory[CW].capital renders Willemstad, with a leading space.
• Place xlsx cells by their r reference in data-import/xlsx.py: a sheet omitting an empty cell shifts every later column left.
• Spell misc.creditcard's digit runs {digits(n)}, and refuse a repeat over a lone {digits(1)} at New, naming that spelling: both render the same run.
• Add the remaining misc tables and templates, one row of the table below per chunk.
• Read misc.emoji from the Unicode emoji-test.txt register, and misc.car from an international make and model source; vPIC and Mobility Sweden are national, so they build en_US.car and sv_SE.car.
• Add the remaining locale categories: company, phone, finance, vehicle, words.
• Give every shipped category building-block columns, misc.uuid → variant among them, and one record shape across the national ids; re-pin the shape.
• Give the address records one column set across countries: region and municipality as columns on geo.SE.address too.
• Fill the 398 Swedish localities weighted 200 from SCB småorter.
• Weight sv_SE.sex and en_US.sex by bearers, as the README's Data section says they are: neither names a weight, so a bare sex draws evenly.
• List title in both locales, and sv_SE's birth-number, in the README's Data list of what each locale carries, which names neither.
• Give url and email a path that draws only domains nobody can register, keeping the wide set as the default, per goal 14: 17 of the 40 distinct ones shipped today sit on .se, .nu, .io and .co, which anyone may register, and only RFC 2606's example.com, .net, .org, .test, .example, .invalid and .localhost provably reach nothing.
• Audit the rest of the shipped set against goal 14 and give each a never-reaching path where it lacks one: phone draws live PTS and NANP ranges — PTS's five fiction series are the Swedish inert set, sourced in research-sources-se.md, and the NANP one still wants a source — and bankgiro, plusgiro and routing draw live prefixes.
• Stop misc.territory[EH].tld rendering .eh, the one shipped TLD misc.tld does not hold: ISO 3166 reserves it for Western Sahara and the root zone has never been delegated it, so no resolver answers for it and goal 3 promises otherwise.
• Draw en_US.phone's exch as a NANP central office code: {int(100,999)} renders a leading 1 in about an eighth of draws, which libphonenumber rejects, and TestShippedUSPhone proves the shape rather than the rule. Assert the rule with it.
• Give a sv_SE.personnummer over 100 the + separator Skatteverket spells, or stop drawing birthdates that reach 100: the format hard-codes -, and the 1930 floor makes the oldest draws invalid from 2031.
• Merge en_US.ip and sv_SE.ip, today byte-identical, into one misc.ip, per goal 4's one spelling per result.
• Accept a middle name, and draw a shipped personnummer inside a selected sex; both want a draw group sharing its family's pins. Stop the conflict error naming a rewrite that returns a different value where the read it conflicts with sits inside another category, and report a struct column's draw conflict with the path spelling a tag takes.
• Let geo.US.locality[1714000].address descend from a selected row into the template beside the family: require the path step to reach a sibling category, and seed every draw group of the render from the outer selector's pins.
• Read email.local's and username's handles from the shipped name tables, which goal 12 asks for and the hand-written list they use today does not meet. Keep the handle shape: username shortens a surname to ahl or sjo, and the US one draws unisex given names, neither of which falls out of the tables unaided.
• Draw email.local from username, so the two share one handle list.
• Rename misc.territory.country to sovereign, per goal 4: it holds a code, and country collides with the misc.country path it replaced.
• Add misc.browser as the parent of misc.useragent, so misc.browser[Chrome].useragent resolves, per goal 3.
• Group data/misc into folders where a group name makes a path easier to guess, per goal 4.

misc (locale-neutral)

Category	Shape	Source	Licence
ip v4 documentation ranges, private, CIDR; ipv6 2001:db8::/32; mac locally administered	t	RFC 5737, 1918, 9637	facts
creditcard per network: IIN, length, CVV, expiry	T+t	network rules, {luhn()}	facts
bic, isin, cusip	t	structure rules	facts
ean, upc, gtin, isbn, issn, imei, asin, sku	t	check-digit rules, GS1 demo prefix 952	facts
product name, adjective, material, department, garmentsize	T/c	LLM-written lists	—
airport IATA, ICAO, territory, municipality; airline; aircraft; flight, seat, pnr	T+t	OurAirports, Wikidata	public domain, CC0
element, unit, planet	T	PubChem, BIPM, NASA	public domain, CC BY 4.0
animal by class, plant	T	Wikidata	CC0
bloodtype weighted (global)	c	AABB	facts
color named, hex, rgb, hsl, cmyk	T+t	CSS Color 4, XKCD	W3C, CC0
programminglanguage, os, browser, database column/type/engine, fileext	T/c	Linguist, curated	MIT
commit message, branch, sha	t	LLM-written, {hex(40)}	—
password, hash, jwt, slug, hostname, avatar, placeholderimage	t	shapes	—
book title, author; genre; instrument; dish, ingredient	T	Gutenberg catalog, MusicBrainz, USDA	CC0
lorem, hacker, hipster, catchphrase, buzzword, quote	T/c	lorem ipsum, LLM-written	—
direction, continent, ulid	c/t	—	—

sv_SE (source research)

Category	Shape	Source	Licence
person title, birthdate, age, blood type weighted	t	geblod.nu distribution	facts
organisationsnummer by form, vat	t	Bolagsverket group digits, Luhn, SE…01	facts
company name patterns, legal form weighted	t	Bolagsverket registrations 2025	CC BY 2.5 SE
sni industry codes	T	SCB SNI 2025	"Källa: SCB"
ssyk occupations	T	SCB SSYK 2012	"Källa: SCB"
phone mobile, fixed per area code, E.164	T+t	PTS numbering plan 2024	facts
bankgiro, plusgiro, bankaccount clearing + Luhn/mod-11	T+t	Bankinfrastruktur CSV, Bankgirot rules	facts
licenseplate ABC 123, ABC 12A, blocked combinations	t	Transportstyrelsen	facts
car make, model weighted by registrations	T	Mobility Sweden 2025	cite
word noun, verb, adjective, adverb; sentence	T+t	SALDO	CC BY 4.0
month and weekday names, holiday	t+T	CLDR sv, lag 1989:253	Unicode
territory localised names keyed by alpha2	T	CLDR sv territory names	Unicode

en_US (source research, and part1–part4)

Category	Shape	Source	Licence
ein valid ranges	t	IRS	facts
company names, suffix, naics	t+T	SEC tickers for patterns, NAICS 2022	public domain
phone NANP with valid NPA/NXX	t	NANPA rules	facts
routing ABA with check, bankaccount	t	Fed prefix ranges	facts
licenseplate per state	T	hand-authored patterns	facts
car make, model	T	NHTSA vPIC	public domain
word, sentence, paragraph	T+t	Moby POS or WordNet	public domain / WordNet
month and weekday names, holiday	t+T	CLDR en	Unicode

Builtins the data cannot express

• Add {isin()} (Luhn over letters expanded to digits), {cusip()}, {aba()} (3-7-1 weights) and {vin()} (position 9 over the whole; a sample taking the WMI, since the check sits mid-string).
• Add {base64url(n)} for JWT shapes.
• Add a unix layout to date() once something needs it.

Library and CLI

• Report every mistake a load finds in one error, as the decision "A load reports every mistake at once" states: New and NewTemplate stop at the first, so data holding two mistakes takes two runs to fix. Split it into sub-items before starting.
• Rename newRand to newGeneratorState and Generator.rand to state: both name half of a generatorState, hiding its {seq()} counters, and New's local rng shadows the rng interface.
• Compare calcParser.binary's operator as a rune: byte(p.rs[p.pos]) reads U+012B ī as +, so {calc(a ī b)} compiles as a + b.
• Stop the columnKinds suggestion naming a field's own kind or a narrower one: an int64 or float64 field unproven in range is told to become itself, and a uint64 one to become int64.
• Say in the README's Linked tables that the first table a render reads decides whose weights govern the family: geo.SE.address reads street first, so it draws by street segments and the region, kommun and postort populations never apply.
• Parse defaultTable's path with the library's grammar: the CLI counts brackets of its own, so a change to the selector grammar desyncs the default --table silently.
• Step up from a row to its parent as the decision "A link column after a row steps up to the parent row, and a path may not end on it" states, refusing a step back down: city[Oslo].country.name is a load error today, and city[Oslo].country renders the cell NO, as locality.municipality does, which List stops advertising. Replace the README's city[Oslo].country example with fejkdata -d ./mydata 'city[Oslo].country.name', rendering Norway.
• Key every table as the decision "A table selects by one key column, a code, never a free-form name" states: geo.SE.municipality[Lund] and misc.territory[Sweden] select a row today.
  • Have a selector that misses name the key of every row whose cell spells it, matched without case, else say the table selects by its key: misc.territory[Sweden] and misc.territory[se] name [SE], misc.protocol[tcp] its key, and misc.loglevel[Error] [err].
  • Delete the name option and the load checks it carries, a name spelling another row's key and a name repeating inside its parent, and have New refuse a table still carrying name, naming key.
  • Key misc.loglevel on keyword, and geo.SE.locality on the integer data-import/geo-se.py assigns as the decision states, linking street and postal-code by it.
  • Drop key from last-name and sv_SE.title, and re-pin testdata/shipped_shape.txt.
  • Leave no selection by name in README.md, CHANGELOG.md, docs/ or --help: no selector spelling a free-form name, geo.SE.locality[Lund], misc.currency[Euro], first-name[Kim] and the rest, no "key or name", "by name" or "either" wording about selectors, and no name option in an example.
• Render selected rows as the decision "A selected row is fixed, not drawn, and an unnamed read beside it is refused" states: New refuses {/misc.territory[SE].capital} / {/misc.territory[FI].capital} today, and a choice whose items select different rows. Have the refusal of an unnamed read name the rewrite through each selection: beside {/geo.US.locality[1714000].name}, {/geo.US.municipality.name} names none today, where {/geo.US.locality[1714000].municipality.name} holds once the step up above lands. Scope the README's "A reference path into a category is held" to a path that selects no row.
• Refuse a struct tag's datatype naming the Go type that sets it before proving its values, as the README's Library section promises: datatype: boolean over {int(1,2)} on an int field answers prints an integer, not a boolean.
• Name the spelling to use in checkNameHeld's path refusal and checkNoOverlap's, as the README's hold section promises: both answer "name the fields you want instead", and {n} reaches "p" by a second route names neither the read the route ends on, {/row.p.last}, nor the spelling to write, {p.last}. Where an operand route's value is the held name, offer it as the operand: lowercase operand "y" renders "/misc.territory" ends at "stop", where {lowercase(/misc.territory)} loads.
• Split the README's linked-tables paragraph opening "A path descends from a row", 23 lines holding five rules: move its load errors that are no overlap to the table section, and drop its bare {/city} refusal, which the hold section states.
• Offer move {Y} into a field with a drawGroup in overlapError where both reads share one field: "a":"{/cat} {/cat.p.first}" ends at "name the fields you want instead", while the same pair in a format offers the move, which loads there too.
• Call templateError a compile failure: its doc says render failure, and it wraps what NewTemplate and NewRecordTemplate refuse.
• Name the node a selector follows in stepInto's refusal: sv_SE.person[1].first answers 1 is not a table.
• Report the same error every load for a table with two bad options, and for a folder with two unnamed rows files: readTableOptions and loadDir return on the first in Go's map order.
• Refuse a category naming a hidden rows file, "rows": ".x.tsv", as a rows file that is not there: loadDir indexes hidden TSVs and skips them in two places, so one loads where its comment says a hidden file is never data.
• Print --help's repeat range from fejkdata.MaxRepeat: the usage spells 1..1048576 a second time.
• Report a CRLF rows file holding only its header as having no rows: parseRows answers "line 2 is empty".
• Reword prove's refusal of a typed column reading a row, {/region}: say the format is the table's own (region's format "{name}"), name the read to write, {/region.<column>} with its columns, and stop calling a one-column format, {code}, composed text.
• Keep one copy of the selector characters in data-import/: httpmethod.py and protocol.py each copy inSelector, and tsv.py's write enforces neither.
• Refuse two whole reads of one table family in one render, which goal 5 promises is a load error: {/sel}|{/sel} panics out of Fake where the two draws land on rows whose cells select different rows of another table, and {/geo.SE.locality}|{/geo.SE.municipality} renders a locality outside the municipality beside it. Each read draws its own row, so cells of two rows render together, where the fence counts them alternatives and compares neither. Have the error name the path spelling that holds, {/geo.SE.locality.name}|{/geo.SE.municipality.name}, as the README's rule on a rejected spelling asks.
• Name the file a link error comes from beside its category path, sub.x (/d/a/sub/x.json): …: with several --data-path layers the author has to work out which directory won, while a parse error on the same file names it.
• Name the row a refused read sits in: where both reads come from TSV cells the overlap error names the category and the two columns but no row, so the author greps a register-sized file for the cell — and "name the fields you want instead" rewrites a cell that is not the one to fix.
• Spell a table one way across the family errors: drawsApart names drawn.segment, pinRow names t.path and mustRow panics with t.category, so one table is territory and misc.territory, and the short spelling names no file where two folders hold that name.
• Let a table column carry a datatype, so --format json writes "safe": true and --format sql a boolean rather than the text 'true'. Today only a JSON field takes one, so misc.httpmethod's booleans are typed in Go and text everywhere else. Typing a shipped column changes what json and sql write, so the capability is a minor and applying it to misc.httpmethod is a major.

Release

• Read the tag through git/ref/tags/{tag} in publish_release.py: GitHub answers tags/{tag} with 404 whatever exists, so the burnt-version guard never fires and a release lands on a tag already pointing at another commit.
• Document NewRecordTemplate, ErrNoColumns, ErrNoData and MaxRepeat in the README's Library section, which names none of them.
• Build on a manual run of test.yml: workflow_dispatch leaves github.event.before empty, so the diff compares HEAD with itself and skips docker build.
• Rewrite CHANGELOG.md's [Unreleased] as what v0.1.0 holds: nothing has shipped, so "no longer paths", "where it used to fail" and "where it used to be an empty string" describe versions no reader can have installed.
• Ship prebuilt binaries so the CLI needs no Go: GoReleaser attaching them to the Gitea release the tag workflow publishes, a container image, Homebrew and Scoop. A checkout build prints devel for --version; the binaries carry the stamped tag.

v0.2.0

• Promise in Generator's godoc that its renders run one at a time, so a caller wanting parallel throughput makes one generator per goroutine.
• Ship the full registers as packs, each a Go module with its own embed.FS and a zip for --data-path: every US place of 10,000 and more streets per locality, built by --min-population and --streets-per-locality. The default embed stays under ~1 MB per country and New under ~50 ms.
• Add the locales nb_NO, da_DK, fi_FI, de_DE, en_GB, nl_NL, fr_FR and es_ES, with person, address, phone, national id, company and date names each, and the geo/ trees NO, DK, FI, NL, FR, AU, CA, ES, GB and DE their address reads by reference; AU and CA get a tree with no locale of their own.
• Add {btc()} and {eth()}, which need sha256 and keccak for Base58Check and EIP-55.
• Pair a street with its exact postnummer, which needs an application to Lantmäteriet; today a street goes to the nearest postal code centroid. It rewrites shipped rows, so it is a major once 1.0 is cut and a minor before that.

v0.3.0

Names no items yet, so planning it is a chunk of its own: it takes the additive work that lands after v0.2.0, and it has to add something, since a minor only adds. Its job is to be the release with no breaking change that v1.0.0 waits on — v0.2.0 cannot be, because pairing a street with its exact postnummer rewrites shipped rows.

v1.0.0

• Cut it once the shipped data is in its record shape and one full minor has shipped with no breaking change, per the README's Versioning table: v0.1.0 settles the record shape and v0.3.0 is the quiet minor.
• Announce it where a developer choosing a fake-data tool already reads, which is goal 1's second half and waits for the tag: complete, stable, installable without Go. Write the README's first screen for someone deciding in a minute.

Not release-bound

• Give every entry in docs/decisions.md the goal it serves, its date, who made it and the premise it rests on; propose a goal for an entry that serves none.
• Hold every change the charter says owes a CHANGELOG.md entry to one in CI: the check covers data and testdata/shipped_shape.txt alone, where AGENTS.md adds a flag, an exit code, an exported name, a fence, a builtin and the lowest Go.
• Measure the allocation gate over the shipped geo trees too: pinSet.inline [8] is sized for them, and perf_test.go checks a synthetic five-deep tree only.
• Name goal 2's four criteria with the axes the panel scores, so a recorded score maps back to the clause it came from: the goal spells them out in prose while todo.md records Navigation, Locality, Shape and Self-sufficiency.
• Cut the README's Layout block to the lines that say what a file name cannot: calc.go, doc.go and cmd/fejkdata/ restate their names.
• Cut AGENTS.md to the rules only it states: the round procedure stands word for word in todo.md and the gh rule opens with the README's GitHub decision, so each is a second copy that drifts. Split the comprehension rule while there, which packs the gate, its exceptions, the round and the at-or-above-7.0 regime into one line.
• Scope AGENTS.md's Hard tabs. to the Go source, or drop it: gofmt already gates Go at Dockerfile:16, and the shipped JSON under data/ is two-space, so the rule as written is one no Go file can break and every data file does.
• Make gitea.larvit.se/larvit/fejkdata a pull mirror of GitHub: Gitea converts a repository to a mirror only by re-creating it, so until that runs the copy there is a second owner of one history and goes stale from this commit.
• Stop /cache collecting root-owned files, which today every command documented without --user leaves for the next --user one to trip over, and which the README answers with a chown a reader has to repeat. Put --user on all of them, or give the x-go anchor a user: — that one needs a mechanism, since Compose interpolates from the process environment and neither UID nor GID is exported there, so user: "${UID}:${GID}" resolves to ":".
• Spell the latest supported Go once: Dockerfile's ARG GO_VERSION and compose.yaml's golang:${GO_VERSION:-…} pin one version twice, and a bump that moved only the first left docker compose run --rm fmt formatting to a gofmt the gate rejects. A renovate.json custom manager reads the second today; drop it with the duplication.
• Publish a homepage with an in-browser generator: the library compiled to WebAssembly, so visitors generate as much data as they like in their own browser.