architecture.md

Architecture

The design in force for domains.pkgforge.dev. This file wins over every other document. When another document disagrees, that document is the defect.

Evidence tags. MEASURED names the command and the UTC time. READ names the URL or the file. [E] is an estimate and says what would replace it. "Design A" is docs/research/2026-10-27-design-a.md. "Design B" is docs/research/2026-10-27-design-b.md. Both keep their raw results beside them. mirror@669ac72:<path> names a file of https://github.com/pkgforge-dev/mirror-archlinux at commit 669ac725f8081445c164113f2ed092533de112e2; read it at https://raw.githubusercontent.com/pkgforge-dev/mirror-archlinux/669ac725f8081445c164113f2ed092533de112e2/<path>. The rulings of the operator are in TODO/RULES.md §5. This file never asks a ruled question again.

1. What it is

domains.pkgforge.dev publishes certificate transparency (CT) names, newly seen domains and newly registered domains. It also serves reverse IP lookups from a mirror of THC's database. The repository is pkgforge-dev/domains.pkgforge.dev. It is public and licensed 0BSD. Every output is 0BSD. Domain names are public data.

GitHub Actions writes the data into GHCR. Cloudflare Workers read it and answer HTTP. R2 is a cache. A reader that does not trust the Worker reads GHCR directly.

Goals

1. Publish CT names and new domains from many feeds, with a reading of our own beside them.
2. Lose any one source and keep answering from the others. Every answer names the sources used and the sources that failed.
3. Be accurate by definition. seen, claimed_new and registered are separate words (§2).
4. Answer every query type that THC offers: reverse IP to names, subdomains, CNAME targets, bulk downloads and daily CT dumps.
5. Answer finer queries of our own. A window is 1 h, 6 h, 12 h, 1 d or N days. A query takes any suffix or parent domain. Every name has first and last seen per source.
6. Keep the data in GHCR, nothing in R2 but a cache of at most 5 GB, and serve from Workers.
7. Accept a VPS ingester as one more source, if a core source shuts down or a need appears.

Non-goals

• No consumer key, account, quota or rate limit. Cost is bounded by caps on window, page and work, never by refusing a caller (§7.6).
• No VPS that reads whole CT logs. All CT logs hold about 510 GB a day (READ docs/research/web-pkgforge-2026-10-27/ct.md §4).
• No heavy regular expressions. Normalisation is a byte scan and a Public Suffix List (PSL) trie.
• No active DNS resolution. Forward and reverse DNS data come from mirrored sources only.
• No call to THC's live API from any route (§6).
• No legal or terms analysis. A document says what a source returns, how often, and how it fails.

The operator's rulings, as constraints

Constraint	Where it holds
Most data lives in GHCR. R2 holds at most 5 GB as a cache. Workers serve.	§4, §4.9
Nothing is ever deleted from GHCR. Raw feeds are kept always, and raw segments stay beside every compacted tier.	§3.4, §4.7
Actions writes with its own GITHUB_TOKEN. The first push of each package comes from a workflow. The operator deploys the Worker with a local wrangler login. No credential crosses between GitHub and Cloudflare.	§5.9, §12
Reverse IP and CNAME answers come from THC's data mirrored in this project's GHCR.	§6
The freshness clock is measured. GitHub Actions is the default.	§5.2
No consumer is gated.	§7.6
Outputs carry no licence field. Provenance is publisher, URL and fetch time.	§3.7

2. Definitions

Term	Meaning
name	A host name. Lowercase, punycode, no trailing dot, letters, digits, hyphen, underscore and dots. A wildcard *.x is the name x with the flag wild.
suffix	The ICANN public suffix of a name, from the pinned PSL (np, com.np, co.uk, xyz).
registrable domain (reg)	The suffix plus one label.
site	The registrable domain computed with the PSL private section too. x.github.io has the site x.github.io.
key	The labels of a name in reverse order, joined with dots. api.example.com.np has the key np.com.example.api. Keys compare bytewise.
source	One producer of observations, with an id, a bit number (0 to 63), a family and a precision (§3.5).
family	What a source derives from: ct, registry+ct, isc, whoisds, stamus, registry, nsec-walk, dns (FAMILIES of packages/adapter/src/types.ts). registry+ct is a CT event that its publisher matched to a registry date (oi-nrd-confirmed). nsec-walk is a walk of a zone's NSEC records (zone-walks). Two sources of one family are not independent evidence.
observation	(name, source, t_first, t_last) with the source's precision. A merge takes the minimum of t_first and the maximum of t_last, so a merge is idempotent.
precision	second or day. It belongs to the source id. A day value is never shown as a time of day.
seen	The earliest evidence a source shows for a name or domain. CT sources use the certificate log time. A dump uses its UTC date. A list uses its own date, never the file date.
first_seen, last_seen	The minimum and maximum of seen over the segments we hold, per source and over all sources. last_seen is not a liveness claim.
claimed_new	A source says the domain is new. The claim keeps the source's own meaning. ISC says "new to us".
registered	A claim about a registry. It is stated only on zone, RDAP or OpenINTEL-confirmed evidence. CT evidence never makes it.
new (registrable)	The global first_seen of the domain falls inside the window. The corpus must show no earlier observation from any source.
new (name)	The same rule for a name. It is censored until the corpus holds one certificate renewal cycle (§9).
corpus_since	The first day the corpus covers, per dataset. Every answer states it.
confidence	warming while the corpus is younger than 90 days, else full.
suspect	An oi-nrd event whose certificate is older than 7 days, or a domain whose RDAP creation date is before 2026-01-01 (ruling 19). The default feed hides suspect events.
provisional	A newness verdict made by the hourly filter. The daily job makes it exact (§5.8).
corroborated	Two sources of different families first saw the domain within 48 h of each other. A day value stands for its whole day: two first sightings count when the two farthest instants of their intervals are at most 48 h apart.
segment	The pack one adapter writes for one window. A segment is immutable.

The OpenINTEL newly registered stream means "first seen in a certificate", not a registration date. In design B's RDAP check, 25 of 25 events with a fresh certificate were registered within a day. In the same check, 0 of 25 events with a certificate older than 30 days were new (design B §2.4 item 5). Of the events in a 180 s sample, 38% carried a certificate older than 30 days.

3. Data model

3.1 Normalisation

Every adapter runs the same function of the names package (entry DOM-015).

1. Lowercase ASCII. Convert IDN labels to punycode. The PSL's own IDN rules are converted the same way when the list is loaded, so a name under 公司.cn splits under xn--55qx5d.cn.
2. Remove one trailing dot. Remove a leading *. and set wild.
3. Refuse a name with a byte outside [a-z0-9_.-], a label over 63 bytes, a name over 253 bytes, an empty label, no dot, or a top-level label of digits only (numeric-tld: an address literal such as 192.0.2.1; RFC 1123 section 2.1 makes the highest-level label alphabetic).
4. Refuse a name that is itself a public suffix.
5. Compute suffix, reg and site from the pinned PSL. The ref package holds the pinned list under the tag psl-<sha12>, the first 12 hex digits of its sha256 (DOM-019, refTag of packages/ghcr/src/layout.ts). That manifest's dev.pkgforge.domains.sha256 annotation holds the whole sha256. A root names it in its psl field (§4.6), and a pack in its dev.pkgforge.domains.psl annotation (§4.2). sources.json holds no PSL field.

THC's dump of 2026-10-26 held 38,474,309 lines. Design A counted 0 refused lines under rule 3 as it then read and 4 refused lines under rule 4. DOM-009 counted distinct names of the same dump: 1,516 whose top-level label is all digits, under 256 distinct such labels, now refused under rule 3, and 2 that are themselves a suffix once the list's IDN rules are converted (MEASURED 2026-10-28T11:19Z, docs/research/measure/DOM-009/local-2026-10-26.stat.json). The dump is not bytewise sorted. 0.57% of its lines carry upper case (MEASURED 2026-10-27, design B §2.4 item 1, design A M5). The ingester normalises and sorts the dump itself.

3.2 Keys

All sorted datasets sort by key (reversed labels). A suffix, a parent domain or a whole TLD is then one contiguous key range. np.com is one range. com.example is one range. np is one range. This holds for names and for the new-domain feed. Hash partitioning cannot answer "every name under .np" without reading every partition.

3.3 Datasets

Dataset	Package	Key	Record	Size order
names	names	reversed name	source mask, then per set bit (first, last). Flag wild.	10 B a name [E]
domains	names	reversed registrable domain	source mask, per set bit (first, last), evidence list	14 B a domain [E]
feed	feed	reversed registrable domain, one record a key	first_seen, source mask, basis bits, confidence, flags	about 2 to 7 MB a day [E] for the key copy, and about 4 to 14 MB with its time copy: 0.3 to 0.7 million rows a day (design B section 6.3, from its measurements M4 and M6) at the 6.1 to 9.3 B a row of a day pack's key copy, and a pack with its time copy 1.982 to 2.104 times its key copy (both MEASURED 2026-10-29 on bench packs, docs/format.md section 14.6).
new-name	feed	reversed name, one record a key	the same shape, name level	250 MB a day [E], the key copy alone; about twice that with its time copy, as for feed
rev	rev	IP address (4 or 16 bytes, big endian)	the names that resolve to the address	measured by DOM-043

first and last are varints. The unit is the source's precision. Days since 2026-01-01 for a day source. Seconds since 2026-01-01T00:00:00Z for a second source. A record never stores a count, so merges stay idempotent.

The domains dataset is the ledger. It holds one record per registrable domain, with per-source first and last seen. It decides whether a domain is new. The feed dataset is a view of it, sorted for window queries. A feed or new-name pack holds one record a key, and first_seen is a field of it, not part of the key. The same rows in first_seen order are the pack's optional time copy: a DPK1 pack of its own (dataset 6 for feed, 7 for new-name), keyed by first_seen and then the key, in the layers time.idx and time.part.NNN of the same manifest (docs/format.md section 14, DOM-022). A page in time order reads the time copy from the bound of its window until the page is full, so no Worker sorts a day pack. A feed row stores first_seen in seconds and no precision. A reader reads the precision back from the row's sources (rowPrecision of packages/feed): the row is day-precision when one of its sources is a day source and first_seen is the start of a day, and its first_seen then stands for that whole day (§7.4). A row whose every source is a day source and whose first_seen is not the start of a day is refused.

An evidence entry is (tier, basis, source, at, confidence).

Tier	Basis	Meaning
T1	zone, rdap	A registry zone adds the domain, a zone diff adds it, or an RDAP registration event says so. A time is unsigned from 2026-01-01, so an RDAP creation date before it is not evidence: it sets suspect (ruling 19).
T2	nrd-list, nrd-ct+rdap	A third party claims it. OpenINTEL's confirmed topic passes its own confidence (1 to 10) through unchanged.
T3	ct-first	The first certificate in our corpus. It is a proxy.

3.4 Observations and segments

An adapter emits one segment per window. A segment is a pack (§4.4) with one source and the observation fields. Its OCI annotations name the source, the window, the cursors before and after, the count and the precision. Raw data is kept always (ruling 5). Every source's bytes, as fetched, stay in GHCR beside the segment derived from them.

The raw lane. Where a source is a file up to 1 GB (THC's daily dump, ISC, WhoisDS, the NRD lists), the segment's manifest keeps the file as fetched in its raw layer. Two sources are streams, and their raw bytes go to the raw lane of the obs package, the tags raw-<source>.<window start>.<window end>, written before the segment that names them:

• Kafka. The record batches of each fetch, exactly as the broker returned them, with a members index of topic, partition, base offset, last offset, byte offset and length. The batches of certstream_domains are snappy compressed. In every window read of the three NRD topics, a batch has codec none and one record. docs/sources.md, "The Zonestream broker", names each window and its command. The batches are not stored as JSON: the decoded JSON of certstream_domains is 24 GB a day (MEASURED 2026-10-27, design A §3.1), and its compressed batches are a fraction of it.
• Names tiles. The gzip body of each tile as the log's host sent it (Accept-Encoding: gzip), with a members index of log, tile number, byte offset, length and sha256.

A raw lane object uses the manifest kind raw (§4.2). It is never read by the Worker. It is the copy that lets a later change of an extractor re-read a stream that its publisher has expired, such as the Kafka topic's 26 hours. The size is in §4.8, from DOM-004's reading of the bytes a day as the broker sends them.

3.5 The source registry

sources.json lives in the ref package and in every root. One entry per source id.

Field	Meaning
id, bit	The id and its bit number. A bit is never reused. At most 64 sources.
family	See §2.
datasets	Which datasets the source feeds.
precision	second or day. A source whose precision changes gets a new id.
kind	ct_name, claimed_new, registered_claim, zone_add, list, dns.
basis	The evidence bases it writes, each one of the five of §3.3. A source that writes no newness evidence (oi-measure) lists none.
cadence, band	How often it runs, and the expected volume band for the volume check (§5.8).
urls, fetched	The publisher URLs and the fetch time. This is the provenance.
status, through	through is the time through which the source is complete, written by the promotion job from the segments it lists. status is ok, late or down: the Worker derives it at read time from through, the source's cadence and the clock, so it never goes stale in the root. The hourly status workflow reads the same numbers through the Worker (§5.3).

3.6 Per-source granularity

CT names tiles are one source per log operator: ct-le, ct-geomys, ct-ipng. The log is a field of the segment's statistics, not of a record. This keeps the source bits under 64 and lets the coverage table compare logs.

3.7 Provenance

A record's provenance is the source id and the segment tag. A segment's annotations carry the publisher, the URL and the fetch time. No record carries a licence field. No document states a source's terms.

4. Storage

GHCR is the only copy of the truth. Every pack is write-once. The root moves forward only. R2 is a cache. Everything in R2 can be rebuilt from GHCR.

4.1 Packages

The namespace is ghcr.io/pkgforge-dev/domains.pkgforge.dev/. This is the repository's own name, as mirror-archlinux does it. Packages are created by the first push of a workflow. They need no manual step (§4.7, ruling 6).

Package	Holds	Tags
root	The root manifest	latest (moves forward), r-<UTC timestamp> (write-once history)
obs	Segments, and the raw bytes they came from: the raw files, the raw lane of Kafka batches and names-tile bodies (§3.4), and the THC monthly raw chunks	<source>.<window start>.<window end> and <source>.<date>.<sha12> for files, each with -d after it for a segment of domains (segTag and fileTag of packages/ghcr); raw-<source>.<window start>.<window end> for the raw lane
names	Tiers of names and domains	<dataset>-<tier>-<span>, for example names-d-20261026; a rebuild of the span adds -rN, N from 2, for example names-d-20261026-r2 (§4.7 rule 2)
feed	Tiers of feed and new-name	feed-h-2026102608, feed-d-20261026, and -rN after it for a rebuild (§4.7 rule 2)
rev	Packs of the reverse IP index	rev-m-202609, and -rN after it for a rebuild (§4.7 rule 2)
dumps	Bulk downloads for people and scripts	names-20261026, new-20261026
state	The cursors, the volume history and the packs owed to the root (state.json, docs/format.md section 17), and the provisional filter with its delta (docs/format.md section 15)	latest (moves, under the state guard of §5.4), state-YYYYMMDDtHHMMSSz (write-once: the state of one move, its time the state's updated; stateRunTag of packages/ghcr), state-YYYYMMDD (write-once: the first state moved in that UTC day), ledger-filter-YYYYMMDD (write-once: the filter through the end of that day; ledgerFilterTag of packages/ghcr), ledger-delta (moves, under the state guard; LEDGER_DELTA_TAG of packages/ghcr), ledger-delta-YYYYMMDDtHHMMSSz (write-once: the delta of one move; ledgerDeltaRunTag of packages/ghcr)
ref	The PSL, the CT log list, sources.json	psl-<sha12>, ctlogs-<sha12>, sources-<sha12>

GHCR tags match [A-Za-z0-9_][A-Za-z0-9._-]{0,127} and resolve without regard to case. Every tag is lowercase. A read of a pinned object uses a digest, never a tag. Only root:latest, state:latest, state:ledger-delta and the dated tags are read by tag. A tag that moves (root:latest, state:latest, state:ledger-delta) moves to an object that is first pushed under a write-once tag of its own: the dated root, or a run tag of state. oras tag needs the manifest in the repository first, and scripts/ghcr/push.mjs refuses a tag that holds another manifest (§4.7 rule 1). The package state gains one run tag for each move: about 144 state- tags a day if the state moves on every run of a chain of 10-minute runs, and about 25 ledger-delta- tags, one for each closed hour and one for the day [E]. An unreferenced run tag is harmless, as an unreferenced dated root is.

4.2 Manifests

Every artifact is an OCI manifest with an artifactType of the form application/vnd.pkgforge.domains.<kind>.v1. ARTIFACT_KINDS of packages/ghcr/src/layout.ts lists the seven kinds, and artifactType builds the type.

Kind	Layers (by title)
pack	part.000, part.001, and so on (blocks, at most 1 GiB each), idx (the sparse index), optional txt.gz (a plain sorted bulk file). A pack of feed or new-name may carry its time copy, time.idx and time.part.000 and up (the same rows in first_seen order, docs/format.md section 14), and day packs of feed carry summary.json (new registrable domains per suffix and source).
segment	the same as pack, and optional raw (the file as fetched)
root	root.json, sources.json, coverage.json
state	one layer: state.json (state:latest and its run tags, docs/format.md section 17), filter.dlbf (a ledger-filter- tag) or delta.dlbd (ledger-delta and its run tags, docs/format.md section 15)
rawchunk	chunk.000 and up, each at most 4 GiB, for the THC monthly CSV
raw	the raw lane (§3.4): raw.000 and up, each at most 1 GiB, the bytes as received (Kafka record batches or gzip tile bodies, concatenated), and members.json, one entry per member with its offset, length and sha256 (docs/format.md section 18)
ref	one object of the ref package (§4.1): one layer, the file as fetched or as committed, titled public_suffix_list.dat, log_list.json or sources.json (KINDS of scripts/ghcr/ref-inputs.mjs)

Each layer carries org.opencontainers.image.title, its file name. The push tool scripts/ghcr/push.mjs sets org.opencontainers.image.source and org.opencontainers.image.created on every manifest. created is the end of the window, or a ref file's own time, never the push time. The caller passes every other annotation, each a key under dev.pkgforge.domains..

• Every manifest carries run, the id of the workflow run that pushed it.
• A pack or segment manifest also carries schema, dataset, tier, span.start, span.end, count, bytes, psl, source, writer and sha256. source is one id, or several separated by commas. When it names one id, the manifest also carries precision, day or second (VERIFY_ANNOTATIONS of packages/pack/src/verify.ts). docs/format.md section 8 states the values of dataset, count, bytes, sha256 and writer.
• A segment also carries adapter (the adapter's directory name), parser (the source's registry parser), cursor.before and cursor.after (the adapter's cursors as canonical JSON, each at most 1024 characters), url (the source's registry urls, joined by one space) and fetched (the time of the adapter run). A segment of a stream source that has a raw lane object also carries raw, that object's tag. A value of 1 to 1024 characters with no control character is the one rule of an annotation value (annotationProblem of packages/ghcr/src/layout.ts), which scripts/ghcr/push.mjs and packManifest apply. The writer of scripts/ingest/run.mjs refuses a cursor that breaks it before it writes the segment (WriteFault cursor), and run.mjs refuses a returned cursor that breaks it before it proposes it, so a stream adapter keeps its cursor within it. Whether a Zonestream cursor fits is not measured: the stream adapters' entries (DOM-026, DOM-036) are to measure it. run and fetched differ between two runs of one window, and scripts/ghcr/push.mjs still finds the second present by the same sha256, artifact type and layers (docs/format.md section 8), and names the digest the tag holds, which the proposal then carries.
• A quarantined segment also carries quarantine, with the value 1 (§5.8).
• A raw lane object (§3.4) carries source, span.start, span.end, bytes (all its raw bytes), sha256 (of all its raw bytes in member order), members (their count), adapter, parser, url, fetched and run. Its created is the end of the window. It is pushed before the segment that names it.
• A state manifest carries sha256 (of its one layer) and run. A delta's manifest also carries filter, the ledger-filter-YYYYMMDD tag that the delta pairs with. The created of a state.json object is the state's updated, of a delta the time of its move, and of a filter the end of its day.
• A ref manifest carries sha256, bytes and url (.github/workflows/ref-refresh.yml). sha256 is the sha256 of the file, which is also the digest of its one layer. url names where the bytes came from.

A blob is scoped to its repository, so the root records the repository next to every digest.

4.3 Reading GHCR

A reader sends Authorization: Bearer QQ== first. GHCR accepts that token for a public package. If GHCR answers 401 or 403, the reader asks the anonymous token endpoint for a pull token and tries once more. A blob request answers 307 to a signed URL on pkg-containers.githubusercontent.com. That URL needs no credential and supports Range. READ docs/research/reference/ghcr.ts and mirror@669ac72:docs/worker.md ("GHCR tokens").

Measurement	Value
Manifest GET, 300 of 300 at concurrency 50	median 0.45 s, p95 0.53 s (MEASURED 2026-10-27, design B M7)
Signed URL life	about 7 to 8.7 minutes (design A M11, design B M7)
Range read of 64 KB from 25 MB	0.94 s from Nepal, 206 (design A M11)
Range read through an unoptimised Worker route	1.5 to 3.0 s (design A M11, design B M7)
Pushes	2,000 requests a minute per token, 10 GB a layer, 10 minutes an upload (READ mirror@669ac72:docs/measurements.md)
Push speed from a runner	63.5 to 121.4 MB/s, MB being 10^6 B: four pushes of one layer of random bytes, 1,536 MiB or 4,096 MiB (MEASURED 2026-10-28, DOM-006, runs 40714201294 and 40716636421, docs/research/measure/DOM-006/)

Reads through a Worker are MEASURED 2026-10-30T19:26Z to 19:36Z (DOM-007, docs/research/measure/DOM-007/result.json). The probe Worker's route /range (probes/worker/src/range.ts) reads a manifest by digest, follows the blob's 307, and reads 25,600 B of the signed URL with Range. A cold read makes those 3 subrequests. A warm read finds the block in the Cache API and makes none. The client was the operator's machine in Nepal, and the colo KTM answered every request. Latency is the client's, from the start of a request to its last byte.

Read	Requests a second, for 60 s	Requests	206	429	5xx	p50 ms	p95 ms	max ms
Cold	20	1,200	1,200	0	0	716	1,741	4,322
Cold	50	3,000	3,000	0	0	703	1,103	4,022
Cold	100	6,000	6,000	0	0	699	728	1,872
Warm	20	1,200	1,200	0	0	23	32	62
Warm	50	3,000	3,000	0	0	19	32	323
Warm	100	6,000	6,000	0	0	17	27	324

• At 100 requests a second, each cold read made a manifest GET and a blob GET to GHCR and one range read to the blob store. That is 12,000 GHCR requests and 6,000 range reads in 60 s, and none answered 429 or a 5xx.
• A cold read one at a time is slower than a cold read under load. 30 cold reads of each size from 1 KiB to 1 MiB, one at a time, had a p50 of 1,384 to 1,578 ms and a p95 of 2,256 to 2,914 ms. The route's own time is within 20 ms of the client's, so the time is in the subrequests. The cause is not measured.
• On 2026-10-29, 2 of 30 cold reads of 1 MiB passed the route's cap of 10 s on one subrequest, at the blob store. On 2026-10-30, 0 of 30 did.
• The numbers are of one colo. A GitHub runner's colo is not measured.

4.4 The pack format

A pack stores sorted records in blocks. It has a sparse index. Its identifier is DPK1. docs/format.md is created by DOM-017 and states every byte.

Part	Content
Records	Sorted by key. Each record stores the shared-prefix length with the previous key, the suffix length with flags, the suffix bytes, then the dataset's fields.
Block	4,096 records by default. The index header states the number, so a reader never assumes it. A block is compressed with deflate-raw. The part header names the codec (deflate-raw is codec 1, gzip codec 2), so a measurement that prefers another block size or codec changes a default of the writer and no byte of the layout.
Part	A run of blocks, at most 1 GiB. Parts are disjoint and ordered by key. A pack has one part or many.
idx layer	Two levels. The top level holds the first key of every 64th index page. The second level holds, per block, the first key, the part, the offset and the lengths.

Why these choices.

• Reversed keys: §3.2.

• deflate-raw: Workers decode it natively with DecompressionStream. The measured density is 6.04 bytes a name at 4,096 names a block. Brotli level 9 and zstd level 12 are within 1.5% of it (design B §2.4 item 3, MEASURED 2026-10-27 06:27Z). Front coding alone costs 12.0 bytes a name.

• Decode CPU in a Worker: one block of 4,096 names costs at most 2 ms of CPU at p99, and an answer of 1,000 rows at most 5 ms. Both are under DOM-008's lines of 3 ms and 15 ms, so deflate-raw at 4,096 names a block stays the default. The table below is MEASURED 2026-10-30T18:44Z to 19:18Z (DOM-008, docs/research/measure/DOM-008/result.json). Each row is 200 invocations of the probe Worker on workers.dev, sent one a second. Each invocation makes two ranged R2 reads, inflates the blocks with DecompressionStream("deflate-raw"), checks their CRC-32, scans every record and answers JSON. The CPU time is the platform's count of the whole invocation, in whole milliseconds, read from wrangler tail.

  Read	Names	CPU p50 ms	CPU p99 ms	CPU per 1,000 names, p50 and p99 ms
  1 block of 1,024	1,024	1	1	0.977, 0.977
  1 block of 4,096	4,096	1	2	0.244, 0.488
  1 block of 16,384	16,384	3	8	0.183, 0.488
  16 blocks of 4,096	65,536	12	17	0.183, 0.259
  16 blocks of 16,384	262,144	44	54	0.168, 0.206
  1,000 rows from 1 block of 4,096	4,096	3	5	0.732, 1.221

  An answer that decodes about 262,000 names in one invocation passes 50 ms of CPU at p99. The sample's blocks hold about 0.84 times the raw bytes of a mean block of a whole day. The CPU time under concurrent requests in one isolate is not measured.

• 4,096 names a block: 6.04 B a name, 7,948 blocks for 32.5 million names. 1,024 names cost 6.26 B and 31,792 blocks. 16,384 names cost 5.93 B and 1,987 blocks.

A lookup reads the cached top index, one index page, and one block. A cold lookup costs at most two small range reads per pack. A merge reads packs sequentially, merges by key with minimum and maximum, and writes a new pack. The same inputs give the same bytes. Compression parameters are fixed.

4.5 Tiers and the live set

A tier is a merge level. A pack of a higher tier covers the spans of the packs it replaces.

Tier	Span	Built from	Built by
seg	one adapter window, 10 minutes or more	the source	the adapter
h	one UTC hour	the segments of the hour	the first live run after the hour closes
d	one UTC day	the 24 hour packs, and the THC day	the daily compaction
w	one ISO week inside one calendar month	the 7 day packs	the weekly compaction
m	one calendar month	the week packs, and the day packs of the weeks that cross its first or its last day	the monthly compaction
y	one year	the month packs	the yearly compaction

The live set is the smallest set of packs that covers [corpus_since, as_of] with disjoint spans. The root lists the live set. The one exception is in names, domains and rev: a segment may overlap a pack of other sources, as the THC day segment lies beside the hour packs of the live lanes until the daily compaction. feed and new-name have no exception. A window is read over disjoint packs (planWindow of packages/feed), and a source that arrives as one day segment enters them through the day pack. A root whose live set leaves a hole is refused (§4.6). The live ingest adds the packs that state owes with --partial: the packs that keep the live set are promoted, and each other one waits, still owed, until a pack covers the hole or the item expires (§4.7 rule 2, §5.4). A query reads the live set. A week that crosses the first of a month is never compacted, so the days of a month before its first Monday and the days of its last week stay day packs until the monthly compaction. The worst fan-out is then 6 segments, 24 hours, 12 days, 4 weeks, 11 months and the years: about 57 packs and one a year for the names dataset, by the rule of the table. The year 2026 of names is never built (the verdict wait below), so its four month packs stay live beside them: about 61 packs, and one a year from 2027. Each pack costs at most two range reads. The feed dataset is small, so its fan-out is cheap. Compaction replaces packs in the live set. It never deletes a pack from GHCR.

The unit. packages/compact (DOM-023) builds the tiers d, w, m and y. planTier(root, dataset, tier, start) reads the root's live set and gives one verdict for one unit:

Verdict	When
skip	The unit is a week that crosses the first of a month. Or every day of the unit is a day the root's gaps names: no empty pack is written (ruling 24). Or every pack of a unit of feed or new-name is quarantined: no empty pack is written either, and the unit of the tier above replaces those packs. Or the unit lies wholly before the dataset's corpus_since: no pack covers it, and no pack of it is ever built (Before corpus_since, below).
present	A live pack of the tier, or of a higher tier, covers the unit.
wait	The packs of the tier below (SLOT_TIERS: h and segments for a day, d for a week, w and d for a month, m for a year) do not cover the unit. A span on a day that gaps names for any source is no missing span, whether no pack covers the day or some pack covers it in part, since ruling 24 has the root accept a hole whose every day gaps names (DOM-029 builds that in packages/root, which refuses every hole until then). A quarantined pack of feed or new-name covers its span as a pack of the tier below does. It is no input of any tier, and the output replaces it. So a week replaces the hour packs of a day whose every pack is quarantined, and does not wait on them. For a day, a source of its packs whose live packs end before the day ends is missing too: the hour packs it has yet to add would overlap the day pack. Only the days gaps names for that source excuse its tail. A gap of another source says nothing of this lane's clock. A span before the dataset's corpus_since is missing too, named as such, and a unit that starts before corpus_since and ends after it is never built (Before corpus_since, below). The verdict names each missing span and the packs of other tiers found there.
ready	The inputs, in the root's order, and every manifest the output replaces.

Before corpus_since. A unit that starts before its dataset's corpus_since is never built (the Decision of DOM-023, ruling 20). The root derives corpus_since from the first span start of the dataset's packs (§4.6). So a pack that spanned the whole unit would move corpus_since back to the unit's start, a coverage no source gives (AGENTS.md absolute 11). A pack from corpus_since alone would not span its tier unit, and the live-set rule refuses it (§4.6). Such a unit that ends after corpus_since keeps the verdict wait, and its reason says that it is never built because it starts before corpus_since. A unit that lies wholly before corpus_since is skip.

The cost is bounded: the first partial unit of each tier stays at the tier below. For names, whose corpus starts at 2026-09-01, the year 2026 stays four month packs. The day 2026-09-01 and the month 2026-09 start at corpus_since and are built. The week from 2026-08-31 crosses the first of September, so it is skip, and the month 2026-09 takes its days 2026-09-01 to 2026-09-06. MEASURED 2026-10-30T04:11:42Z, planTier called by node on a synthetic root of the four month packs of names from 2026-09: the year 2026 is wait, with the reason that it is never built, and the year 2025 and that week are skip. The two tests of packages/compact/test/plan.test.ts whose names hold corpus_since hold both verdicts.

The exact join writes the day pack of feed (buildFeedDay of packages/feed, DOM-022). A merge of its hour packs does not, because the verdicts of the hour packs are provisional (section 5.8 step 5). The daily job takes from planTier the verdict and the manifests that the day pack replaces. new-name has no hour pack. It enters only through the exact day pack of DOM-041, whose Decision rules out a filter of names (the Decision of DOM-023, ruling 20), and scripts/compact/hour.mjs writes the hour pack of feed alone (READ its header).

mergeSpan builds every other unit. A live pack that crosses an edge of the unit is refused (straddle): no pack of the unit can replace it. A pack marked quarantined (its entry's quarantine field, ruling 23) is no input of feed or new-name, and the output still replaces it. In names, domains and rev it is an input.

The merge. mergeSpan is a streaming k-way merge (mergeRecords of packages/pack, with the rules of docs/format.md section 5.6). It holds one block of each input and one output block. A second pass, checkMerged, walks the output beside one cursor on each input: the output holds exactly the distinct keys of the inputs, and each record is the merge of the inputs' records of its key. It shares no code with mergeRecords. It calls the dataset codec's codec.merge and codec.key, which mergeRecords calls too, and compares two records by their bytes from codec.encodeRecord, which the writer calls (READ packages/compact/src/check.ts and packages/pack/src/merge.ts). So a defect in the codec's record rule passes both. The record rule is held by its own tests (READ packages/pack/test/codec.test.ts: per source the earliest first and the latest last, the flags ORed, and every merge idempotent, commutative and associative). Any difference refuses the output (count), with both counts and the first key that differs. A range job runs the same pass over its own part (checkRange). The pass reads the inputs' keys up to the key after the part's last. It starts at the part's first key, or at the start of the key space for part 0. Assembly holds each part's first key to the key after the last record of the part before it. So the checks of all parts cover the key space, with no key in two. A key that the merge drops before the first part, or between two parts, fails the check of one part.

Ranges. A large merge runs as an Actions matrix. planRanges cuts the key space at index page keys into ranges of about equal input bytes, the bytes before a cut being the compressed bytes of the input blocks whose first key is below it: the shards of mirror-archlinux, whose SHARD_BYTES is 536,870,912 (READ in the Premise of DOM-023, TODO/storage.md, from that repository's docs/design.md, "Shards"). The target is DEFAULT_RANGE_BYTES, 536,870,912 B, and a merge has at most MAX_RANGES, 256, ranges (section 5.1). The pack must also fit one manifest. A range owns at most one part, so a pack holds a part a key range and its idx, and with a time copy a part a time range and time.idx. scripts/ghcr/push.mjs pushes at most MAX_LAYERS, 256, layers, the largest count both push runs of DOM-006 accepted (READ its comment), and packages/compact holds the same number, which a test of scripts/compact holds equal to the push tool's. So rangeCeiling caps the ranges at 255, or 127 for a feed or new-name pack of tier d to y. The time copy is cut into no more ranges than the key copy, though by bytes it may want one more, since it is 0.982 to 1.104 of the key copy's bytes (docs/format.md section 14.6). planFile, the plan read back, and packManifest each refuse a pack of more layers. The ranged layout keeps the blocks of one writer byte for byte, because the format allows a short block only at the end of a pack (docs/format.md sections 2 and 6.5). A range owns the blocks whose first record falls in it, and writes them as one part of at most 1 GiB. The ranges that own a block take the part numbers 0, 1, 2 in key order. A range cannot know where its first block starts until the ranges before it are counted, since the merge folds the keys that two inputs share. So a matrix runs in two stages. Stage 1 counts the distinct keys of each range (countRange), and stage 2 writes each range's part from every range's count (windowOf, writeRange). assembleRanges builds the index of all parts. It refuses a range that is missing or given twice, parts out of order, a count that is not the count of its index, and two ranges that do not meet: each part's first key is the key after the last record of the part before it. hashParts reads every block back, holds it to its CRC-32, and computes the identity of docs/format.md section 8.

The time copy. A tier pack of feed or new-name carries its time copy (docs/format.md section 14). When no key is in two inputs, the merge moves no row, and the time copy is the k-way merge of the inputs' time copies by time key, by range like the key copy. The key copy's count tells the two cases apart: the merge's records equal the inputs' records exactly when no key is in two inputs. Otherwise sortTimeCopy sorts the merged rows in memory, at most DEFAULT_SORT_ROWS, 10,000,000 rows [E], and refuses more before it writes a byte. Both forms are held to the key copy by timeBindFault of packages/pack.

The manifest and the tag. packManifest builds the OCI manifest: its layers part.NNN, idx, for a time copy time.part.NNN and time.idx, then its extra layers, such as summary.json of a feed day pack, and the annotations of section 4.2. It builds the manifest of an hour pack too (tier h). It reads the manifest back through packFromManifest of packages/root, so the promotion's refusals happen before any push. tagDecision reads the tags of the package. A tag of the unit that is there means the unit is present, and a re-run writes nothing. A rebuild takes the next tag of the span, -r2 and up (ruling 25), from nextTierTag of packages/ghcr.

A re-run of a rebuild. A re-run of a rebuild whose push landed first finds the build that holds its bytes, and writes nothing. The caller gives tagDecision the dev.pkgforge.domains.sha256 it is about to write, and the same annotation of each build of the span. A build whose annotation equals it answers present, with that build's tag, the latest such build when two match. Only when no build matches does the rebuild write, under nextTierTag. A build that the tags list and the map of annotations lacks is a usage fault, so a caller cannot skip the read. MEASURED 2026-10-30T04:11:42Z, tagDecision called by node with the builds names-d-20261026 and names-d-20261026-r2: a rebuild of the bytes of -r2 answers present at -r2, and a rebuild of other bytes answers write at -r3. The test of packages/compact/test/manifest.test.ts whose name starts tag: a rebuild whose sha256 a build already holds holds this.

When a rebuild is asked and the span has builds, assemble.mjs, hour.mjs and day.mjs read each build's manifest with no credential, through the one helper decideRebuild of scripts/compact/plan.mjs (READ its buildSha256s). A build whose manifest GHCR does not give, or whose dev.pkgforge.domains.sha256 is not 64 lowercase hex digits, is exit 1, never 4: the package lists the build, so the build is lost, or its package went private since. The tests of scripts/compact/assemble.test.mjs, scripts/compact/hour.test.mjs and scripts/compact/day.test.mjs whose names hold rebuild and exit 1 plant a manifest GHCR does not give and one with no annotation, and the one of assemble.test.mjs plants an annotation in capitals too.

The scripts. scripts/compact/plan.mjs plans a unit from root:latest or a root.json file, and writes the plan and the matrix { "include": [{ "range": I }] }. It exits 3 for wait. scripts/compact/range.mjs runs stage 1 (count), stage 2 (write), or both for every range in one process (all). It reads the inputs from files or from GHCR with no credential, writes each part to a new file with its digest and size, and checks each part with checkRange. scripts/compact/assemble.mjs assembles the parts, sorts the time copy when it must, and writes the index, manifest.json and blobs.json. With --inputs, it also runs checkMerged over the whole pack, which reads every input again. Without --inputs, the checks of the ranges hold the pack. If the package holds a build of the unit, it prints present PACKAGE:TAG, writes nothing and exits 0. Each script pushes nothing, and their headers state their exit codes. The job assemble pushes the layers that blobs.json names with scripts/ghcr/push.mjs, which builds the pushed manifest itself from those layers and the annotations of manifest.json (§5.3). Each checks its output paths before it reads anything: an output that is a directory where a file goes, a file where a directory goes, or under a directory that is not there, exits 2. A blob that the validated root, or the plan made from it, names and that GHCR does not give with no credential exits 1, never 4: the root was promoted only after each package read publicly (section 4.7 rule 7), so the pack is lost or its package went private since. Exit 4 is left to the root's own read. A test gives range.mjs a merge that drops the smallest key, and range.mjs exits 1 with FAIL count. A test gives assemble.mjs --inputs a part written again without one key, and it exits 1 the same way. Without --inputs, assemble.mjs holds the count both ways: a pack of exactly the inputs' records when a range says a key is in two inputs, or of more records than the inputs when none says so, exits 1 (READ scripts/compact/range.test.mjs and scripts/compact/assemble.test.mjs, the tests whose names hold FAIL count, drops one input key and held both ways). MEASURED 2026-10-30T09:30:53Z by measure-compact.yml (run 41032308505, docs/research/measure/DOM-023/result.json): one range of a names week, 7 synthetic day packs of 92,402,051 records in 577,765,791 B read from local files, took 82.71 s to count and 382.11 s to write on an ubuntu-24.04 runner of 4 CPUs, at a peak RSS of 300,572 KiB and about 1.2 CPUs, and used 522,960,896 B of disk. The read from GHCR by range, a time copy and the sort of sortTimeCopy are not in it.

4.6 The root manifest

root:latest is an OCI manifest of kind root, at most 64 KiB, the Worker's cap. Its layers are root.json, sources.json and coverage.json, each application/json. Its layer root.json is the machine contract (§11). validateRoot of packages/root (DOM-020) holds its schema, and refuses a format it does not know. It holds:

• format (domains.root.v1), root_seq (a whole number from 1 that only grows), as_of (the latest span end of a live pack, or null when the root names none), created (when the promotion built it, to the second; the dated tag r-<created> of rootTag names it), and previous (the manifest digest of the root it replaced; null for the first root alone, root_seq 1).
• psl: the sha256 of the list pinned at promotion, the dev.pkgforge.domains.sha256 annotation of the ref package's psl-<sha12> (held equal to that tag's one layer).
• psl_live: the sorted sha256 of every list a live pack was built under, read from each pack's dev.pkgforge.domains.psl (ruling 22). psl is in psl_live only once a pack is built under it. No compaction recomputes a pack under a new list: the plan steps of the compaction workflows hold a ready unit whose keys depend on the list and whose inputs carry more than one PSL hash (§5.3, "PSL lists"). So psl_live keeps naming the older lists. The recompute under one list is an open duty of DOM-023.
• corpus_since per dataset (the first span start of its packs), gaps per source (the days it is known to lack, carried forward, for example the THC days 2026-10-04 and 2026-10-09), through per source (the latest span end of a live pack that carries it), and counts per dataset (packs and records).
• datasets, with the live packs of each, in span order. Each pack has repo, tag, tier (seg for a segment, or h to y), span (start, end), manifest (digest), pack (sha256, the decoded identity, and bytes, from its annotations), idx and each of parts (digest and size), count, key_min and key_max (from its index head; null for a pack of no record), sources and psl. A pack of feed or new-name that carries a time copy (docs/format.md section 14) also has time: { "idx": { digest, size }, "parts": [{ digest, size }, ...] }, its layers time.idx and time.part.000 and up. A pack with no time copy has no time field, never a null one, and no pack of another dataset has one. packFromManifest takes the two layers both or neither, and holds the time index head to the pack: its dataset byte (6 for feed, 7 for new-name), its record count (the key index's), each part's size (the layer's), and the first first_seen of its key bounds no earlier than the span's start and the last before its end, the half of the bind readFeed holds before it reads a block of the copy. validateRoot holds the shape, each blob at least the header of its kind (64 B for the index, 16 B for a part), and planWindow of packages/feed reads it. A writer's validateRoot also refuses a pack of feed or new-name of tier d, w, m or y that holds records and lists no time copy, since readFeed would refuse every page over it; a reader's leaves that to readFeed, so one bad pack fails the pages over it and not the whole root. pack.bytes counts the key copy alone, as the annotations do.
• withdrawn: the manifests a rollback took out of the live set, each { "manifest": "sha256:...", "at": ... }, where at is the created of the root that withdrew it, sorted by manifest, and [] in the first root. A rollback's new root lists the live root's entries, each with its own at, plus each manifest that the live root or a root between it and the target names and the target does not (at its own created), minus each manifest the target names. So a pack a compaction replaced after the target was live, which no live root names any more, is listed too. A root of that chain created more than 31 days before the rollback adds nothing: the run that first added a pack it names started before that root, past GitHub's 30 days of re-runs, and a newer root that names the pack adds it. Every root drops the entries whose at is more than 31 days before its created: GitHub re-runs a workflow run up to 30 days after its initial run (READ 2026-10-29T11:19:40Z, https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs), and the day more is a margin. validateRoot holds its shape and order, each at no later than created and at most 31 days before it, and no manifest it lists among the live packs. A writer refuses a root without it.

Every derived field (as_of, psl_live, corpus_since, through, counts) is computed again from the packs by validateRoot and must be equal. A writer refuses a field or a dataset it does not know. A reader ignores them (§11), and then holds only the per-dataset fields of the datasets it knows.

The live-set rule of §4.5, as validateRoot holds it for each dataset. A compacted pack (h to y) spans exactly its tier unit, aligned, under the tag tierTag gives it, or that tag with -rN for a rebuild of the span (§4.7 rule 2, ruling 25). Two compacted packs never overlap. Two packs that share a source never overlap, so a merge that left a pack it replaced in the set is refused. In names, domains and rev a segment may overlap a pack of other sources: the THC day segment lies beside the hour packs of the live lanes until the daily compaction. That is the one exception to the disjoint spans of §4.5. In feed and new-name no two packs overlap, a segment neither (SEGMENT_OVERLAP of packages/root). The union of the spans is one interval from corpus_since. A hole is accepted only when every day of it is listed in gaps, and no empty pack is written for such a day (ruling 24); a hole that gaps does not name is refused as a missed promotion. Until DOM-029 changes validateRoot to hold that, every hole is refused. No manifest digest appears twice in a root.

sources.json (format domains.root-sources.v1) is the registry, config/sources.json, with each source's through, held equal to root.json, and its fetched time, given or carried from the previous root (§3.5). A live pack may carry only a source the registry holds. coverage.json is the coverage table of DOM-035. Until DOM-035 writes it, the layer holds the table's shape with nothing measured: as_of and day null, no pair, no gap.

The first promotion finds no root:latest: the root is root_seq 1, previous null, as_of null, with no dataset unless it adds packs, and psl the list it is given to pin. layerBytes writes each layer as compact JSON with one final newline: machines read it, whatever its spacing. root.json then holds about 0.73 KB for each pack entry of one part and three sources, about 97 B for each further part and about 115 B for each withdrawn entry: about 37 KB (36,742 B) for 50 packs of one dataset, fewer than the worst fan-out §4.5 gives for names, and about 182 KB (181,552 B) for 50 packs in each of the five datasets [E]. A time copy of one part adds 218 B to its pack's entry [E] (50 feed hour packs with one, less 50 without). Each is measured on synthetic entries with node scripts/root/size.mjs at 2026-10-29T14:34:49Z, and its test holds these figures; the first real roots replace them. The 64 KiB cap is the manifest's, not the layer's: scripts/root/lib.mjs reads a layer of up to 8 MiB. A Worker reads it once, then reads blobs by digest.

4.7 Write rules

⛔ Nothing is ever deleted from GHCR (ruling 5). No script calls the packages DELETE API, oras manifest delete or crane delete. A static check refuses them (entry DOM-014).

1. Write-once tags. A tag that exists is never written again. A script reads the tag first. The same digest means skip. Another digest is a fault that stops the job. A file source that republishes other bytes gets a new tag with a new sha12.
2. Forward-only root. root:latest moves only after everything it names exists on GHCR, by an anonymous HEAD of each pack's manifest by digest. Verification (§5.8) is the duty of the job that wrote each pack, before it calls the promotion; scripts/root/promote.mjs holds each pack's manifest and index head to each other and does not check that the verification ran. It refuses a pack whose manifest carries dev.pkgforge.domains.quarantine: no field of the root would keep a quarantined segment out of new. Ruling 23 serves such a pack, never as new: DOM-029 adds a field to the pack entry that marks it quarantined and lifts the refusal. It builds the next root with root_seq the live one plus 1 and previous the live digest, and each replacement spans every pack it replaces and carries each of their sources. It reads latest with the job's token just before it pushes the root under its dated tag r-<created> (write-once, with scripts/ghcr/push.mjs), and again just before it moves latest with oras tag. Either read refuses the move, exit 1, when latest is not the root the run planned on, when the new root_seq is not above the live one, when the new root was created no later than the live one, or, outside a rollback, when the new root shrinks the live one: an as_of below it, a dataset's corpus_since later or gone, or a source's through earlier or gone. as_of may stay equal: a compaction replaces packs and does not extend the corpus. A rollback (scripts/root/rollback.mjs) is a new root with a higher root_seq that names the packs of an older root, so it may shrink the root, never lower root_seq. Its target must have been live once: it is on the chain of previous from latest, so a dated root that never became latest is refused. A rollback lists in withdrawn each pack it takes out, and each pack a root since the target named and a compaction replaced (§4.6), and every later root refuses an addition or a replacement whose manifest withdrawn lists, so a re-run of the promotion that first added such a pack cannot put it back. A withdrawn pack comes back through a rollback to a root that names it, or under its own digest once its entry leaves the list (§4.6). buildRoot does not refuse a pack of the same span under another manifest digest. The withdrawn pack keeps the tag its span gives, write-once (rule 1). So a rebuild of the span takes that tag with the suffix -rN, such as names-d-20261026-r2, then -r3 (ruling 25). The first build carries no suffix. N is a whole number from 2 to 2^53 - 1 (Number.MAX_SAFE_INTEGER), with no leading zero. The same holds for any tier pack found bad after its push. In packages/ghcr/src/layout.ts, tierTag builds the suffix, parseTierTag reads it back, and nextTierTag gives the next free tag of a span from the package's tags. checkTag refuses a value that is not a string, and the suffix on a tag that is not a tier tag. validatePack of packages/root accepts a tier pack's tag only as its span's tag or that tag with -rN. nextTierTag names a tag for bytes that no build of the span holds. So a re-run of a rebuild whose push landed first finds the build that holds its bytes, by its annotation dev.pkgforge.domains.sha256, and writes nothing: tagDecision of packages/compact compares that annotation of each build with the value the run is about to write (§4.5, "A re-run of a rebuild"). A test of packages/root/test/build.test.ts withdraws names-d-20261026 by a rollback, and a later promotion adds names-d-20261026-r2 for the same span (MEASURED 2026-10-30T04:46:24Z, node scripts/test.mjs packages/root exits 0 with 55 tests). A refusal after the dated push leaves that dated root unreferenced, which is harmless. Then latest is read back as a consumer reads it, with no credential: its digest is the pushed one, and its root.json is the root built. A root that the live-set rule refuses (§4.6) is refused whole, exit 1, unless the promotion runs with --partial, the input partial of root-promote.yml. The job promote of ingest-live.yml passes it (§5.3, the Decision of DOM-024). With it, promote.mjs builds the replacements first, whole, and then adds the additions that the rule allows. It tries each addition with those kept so far, in passes until a pass keeps none, since one pack can close the hole another needs. No addition left out can join the kept ones. It prints each other addition as waiting PKG:TAG: <the rule's message>. When every addition waits and nothing else changes, it ends unchanged and writes nothing. Only the rule's refusal of an addition waits: a replacement the rule refuses, a missing, quarantined or withdrawn pack, the guard and every other fault exit as without it. A test of scripts/root/promote.test.mjs whose name starts promote: --partial adds holds this.
3. One writer. Every move of root:latest runs in .github/workflows/root-promote.yml, whose concurrency group is root-writer with queue: max and no cancel (ruling 18). A workflow whose job only calls it holds no group of its own, since a caller in root-writer would wait on the workflow it calls. That is assumed: no page read states it, and nothing has measured it. That two calls queue in the called file's group is MEASURED (2026-10-29T15:50Z, DOM-020): a caller of it with no group, dispatched twice, made runs 40905625954 and 40905643334, created 6 s apart, and the second call's promote job started at 15:50:58Z, 2 s after the first's ended at 15:50:56Z. Both ended unchanged. The caller was removed when DOM-020 closed. Rule 13 of scripts/check-workflows.mjs holds both. Every other job that moves state runs in root-writer itself. Rule 13 reads a run of scripts/state/move.mjs or scripts/state/hours.mjs as a move of state, since each pushes and tags inside its own process. The job reads the live root, applies its additions and replacements, and pushes. Two writers never move the root at once. Jobs that run in parallel, such as the ingest lanes, write segments and cursor proposals only. In ingest-live.yml the job push advances the cursors, builds the hour packs and moves state:latest, in root-writer. The job promote after it only calls root-promote.yml with the packs that state owes, and partial: true (rule 2, §5.3, §5.4). In compact-daily.yml the job push pushes the day's filter and moves state:ledger-delta, in root-writer.
4. Re-run guard. A re-run of an older run (GITHUB_RUN_ATTEMPT above 1) reads the live root first. It drops every addition and replacement that the live root already names, or supersedes with a pack of a higher tier whose span and sources cover it. It writes nothing when nothing is left to change. A re-run of a rollback writes nothing: it ends unchanged when the live root already names the target's datasets and gaps, and is refused otherwise, since promotions may have landed after the first attempt. A rollback is dispatched again, never re-run. A re-run after a rollback is refused by withdrawn (rule 2): a re-run whose addition or replacement a rollback took out exits 1 and writes nothing. A GITHUB_RUN_ATTEMPT that is not a whole number from 1, in digits alone, is exit 2 before anything is read.
5. Rate limits. A 429 from GHCR costs no attempt. The script pauses 20 to 60 seconds, with a total cap, then tries again.
6. First push from a workflow. A package first pushed from outside is not linked to the repository, and the workflow token can never write to it. Each package's first push comes from a workflow (READ mirror@669ac72:docs/design.md, "Running outside GitHub Actions").
7. Anonymous-read check. After each push of a new package, a job reads its manifest with no credential. A 401, 403 or 404 opens the issue visibility-<package>. mirror-archlinux has no visibility- issue in its whole history (MEASURED 2026-10-27, gh issue list --repo pkgforge-dev/mirror-archlinux --state all --search visibility returned none). A package created by GITHUB_TOKEN from a public repository reads anonymously (READ mirror@669ac72:docs/measurements.md, GHCR, 2026-10-14).

4.8 Sizes

Item	Value
THC daily raw file	319.9 MB for 2026-10-26 (MEASURED 2026-10-27, design A M3)
THC daily names pack	about 200 MB, from 32,554,702 distinct names at 6.04 B (design B M10)
THC monthly CSV, September	37,046,639,482 B (MEASURED 2026-10-27T08:21:52Z, HEAD of the Hugging Face URL)
THC monthly CSV, August	36,029,717,282 B (MEASURED 2026-10-27T08:25:12Z, bucket listing)
Raw lane, Kafka certstream_domains	10.6 GB a day on the wire, 28.3 GB once decoded, a snappy ratio of 2.68 (MEASURED 2026-10-28T15:16Z, DOM-004, run 40714194235: 64,001,338 B for 536,533 messages, scaled to the 88,461,905 messages of the 24 hours before the run; docs/research/measure/DOM-004/result.json). The figure moves with the topic's 24-hour count.
Raw lane, the three NRD topics	0.19 GB a day on the wire: 82.5 MB, 71.5 MB and 37.8 MB for newly_registered_domain, newly_registered_fqdn and confirmed_newly_registered_domain, one record a batch (MEASURED 2026-10-28T15:16Z, DOM-004, run 40714194235).
Raw lane, names tiles	about 1.4 GB a day (MEASURED, design A M10: 26.4 B an entry for 51.7 million entries a day)
GHCR growth a year	derived tiers with every tier kept: 1 to 2 TB [E]. The raw lane adds about 4.4 TB [E]: its two measured parts, 10.7 GB a day of Kafka batches (DOM-004) and 1.4 GB a day of names tiles (design A M10), are 12.1 GB a day, and the file sources add less than 1 GB a day. DOM-023 measures the derived growth. Raw data is kept always (ruling 5), so the total is about 5.4 to 6.4 TB a year [E].
mirror-archlinux, for scale	368.5 GiB on one organisation (READ mirror@669ac72:docs/measurements.md)

4.9 R2

R2 holds a cache and a hot copy. It holds nothing else. R2 has no per-bucket size cap, so the Worker enforces the 5 GB limit itself.

Prefix	Purpose	Budget
root/	copies of the root and sources.json	0.05 GB
idx/	indexes of live packs	0.5 GB
hot/	the newest segments, hour packs and feed packs, copied from GHCR after each promotion	0.8 GB
blk/	blocks by digest/offset-length, immutable, read through	2.7 GB
headroom		0.45 GB

The planned total is 4.5 GB. The Worker keeps a size counter and stops writing blk/ objects at 4.3 GB. A scheduled job recomputes the counter from a listing. A lifecycle rule expires blk/ after 14 days and hot/ after 7 days. The Cache API sits in front of R2. R2 writes stay under 25,000 a day (class A) [E]. Only Workers write R2. Actions holds no Cloudflare credential.

5. Ingestion

5.1 Where each step runs

Step	Runs on	Why
Adapters: THC dump, Kafka, tiles, feeds	GitHub Actions	Free for public repositories. A runner has 4 vCPU, 16 GB RAM and a real network. Kafka needs a client with offsets.
Compaction and verification	Actions, with a matrix when a pack is large	A job lasts up to 6 h. A matrix has up to 256 jobs and 20 run at once for the organisation.
Root and state promotion	Actions: the job push moves state, in the group root-writer; the job promote calls root-promote.yml, which holds the group	One writer moves the root forward (§4.7).
Serving, caching, R2 housekeeping	Workers	They read only. They hold no secret.
Watchdog, status, keepalive	Actions, then the Worker shows the age	A Worker cannot dispatch a workflow without a token, and no credential crosses (ruling 6).

READ for the Actions limits: standard runners are free on public repositories with 4 vCPU, 16 GB RAM and 14 GB SSD. A job lasts at most 6 h and a matrix has at most 256 jobs. The free plan runs 20 jobs at once for the whole organisation. GITHUB_TOKEN is limited to 1,000 REST requests an hour for each repository (design A §2.2, design B §2.2). A runner holds about 87 GB free in practice and 120 GB after the mirror's debloat-runner script (READ mirror@669ac72:docs/measurements.md).

5.2 The clock

The clock is the thing that starts an ingest run. Ruling 8 requires a measurement. The design has three candidates.

Clock	How it works	Evidence so far
A. schedule	A cron line in a workflow.	GitHub delays and drops scheduled runs (READ events documentation). On mirror-archlinux, the */15 workflow ran 40 times in its last 300 runs, with a median gap of 233 min (p90 318, max 377). A re-measurement at 2026-10-27T09:14Z gave 67 runs, median 229 min, p90 314, max 443. The 17 */3 workflow had a median gap of 274 min with 10 dispatches among 37 runs. The hourly status workflow reached a 53 min median only with 90 dispatches. MEASURED 2026-10-27T06:33Z by the operator's agent. Design B measured the same: 66 scheduled runs in 10.1 days against about 970 expected, median 227 min, max 443 min (design B M8). On this repository, the */5 workflow measure-clock-schedule.yml fails the rule over 48 hours (DOM-001; the numbers are in docs/clock.md).
B. Self-dispatch chain	Each run starts the next with gh workflow run, using GITHUB_TOKEN. A concurrency group makes the next run wait until this one ends.	Once a run exists, its runner starts fast: median 2 s, p90 3 s, max 7 s over 250 jobs (design B M8). Dispatch runs queue like scheduled runs. On this repository, a chain of 5 minutes passes the rule over 48 hours (DOM-001; the numbers are in docs/clock.md).
C. Cloudflare cron Worker	A Worker with a cron trigger reads the sources, writes a 72 h hot tier to R2, and the Actions jobs write the durable tiers.	The probe Worker of DOM-061 ran a minute cron for 48 hours and passes the rule (DOM-002; the numbers are in docs/clock.md). A tick records no colo.

The decision rule (DOM-003) reads 48 hours of gaps from each clock. The two thresholds below, a p90 of 15 minutes and a maximum of 60, were proposed from the freshness target, and the operator confirmed them on 2026-10-28 (ruling 17).

• Take the first of A, then B, then C whose p90 gap is at most 15 minutes and whose maximum gap is at most 60 minutes. If all three fail, the verdict is none: no clock is written, and the operator is asked.
• A verdict of C builds C (entry DOM-025). Actions then carries only the durable work: the THC daily job, compaction, verification and the root.
• If B passes, the watchdog (DOM-034) still restarts a broken chain, as mirror-archlinux does (§5.9).

scripts/measure/clock-verdict.mjs (DOM-003) applies the rule to the three results, config/clock.json records the verdict, and docs/clock.md gives it with the numbers it rests on.

The chain is the job chain of ingest-live.yml. It needs no other job, so it starts with the run, and its first step that runs a script runs scripts/ingest/chain.mjs start. That script is the one reader of config/clock.json, which DOM-003 writes. Under the clock actions-chain, on main and on the first attempt of a run, it dispatches the next run with gh workflow run ingest-live.yml --ref main and the job's token (actions: write), at most 3 tries 15 s apart. It dispatches none when another run of the workflow waits, so a live chain keeps one waiting successor and the cron backstop never starts a second chain. The successor waits in the group ingest-live until this run ends. The last step, chain.mjs sleep, sleeps to the next boundary of the interval: the optional whole number interval_min of the file, 1 to 12, default 10. With no file, or with another clock, both steps print chain off and exit 0: only the cron 7,22,37,52 * * * * and the dispatches start runs.

GitHub Actions is the default. C is the alternative a design built only on Actions would not need. A hot tier written by a Worker lets a 1 h window stay fresh when Actions is late. Under C, the hot tier carries what needs no ledger: OpenINTEL claims, ISC rows, and CT names with their log times. Newness verdicts stay provisional until the daily job (§5.8). Kafka over a Worker connect() passed DOM-005 on 2026-10-30 from colo KTM, the colo that answered its HTTP requests: a call holds 16 MiB of max_bytes under a 30 s cap. The read rate of a cron Worker, from the colo its invocations run in, is not measured. DOM-025 is closed on ruling 8 with no code, because an Actions clock passes the rule first (DOM-003), so the Kafka lane runs on Actions.

5.3 Schedules

All times are UTC. Minutes are off the top of the hour because GitHub delays the start of every hour. Durations are [E] until DOM-001 to DOM-012 replace them.

Workflow	Trigger	Does	Run time [E]
ingest-live.yml	the chain (§5.2) under the clock actions-chain, a backstop cron 7,22,37,52 * * * *, and dispatch with sources, from and to	The adapters of the three lanes of config/lanes.json: Kafka topics, tile slices, conditional GETs of ISC and the lists. Pushes segments and raw lane objects. Moves the cursors. Builds the hour pack of each closed hour. Promotes the packs that state owes the root.	3 to 6 min
ingest-thc.yml	20 0 * * *, then 20 1-6 * * * until the day's file exists	Download, normalise, sort, encode, verify, push the day's segment	10 to 30 min
compact-daily.yml	35 1 * * *, and dispatch with day, rebuild and first-ledger	Plan the day dispatched, or the last 7 days. Merge the hour packs and the segments of names, the THC day among them, into d. Build the domains day pack, the feed day pack by the exact join, and the ledger filter (scripts/compact/day.mjs). Verify, push, promote. The new-name day pack (DOM-041), the dumps (DOM-040) and the coverage table (DOM-035) come later.	20 to 40 min
compact-weekly.yml	50 3 * * 1, and dispatch with span and rebuild	Plan the week dispatched, or the last 5 weeks. Merge 7 day packs into w for names, domains, feed and new-name. Skip a week that crosses the first of a month.	matrix
compact-monthly.yml	50 4 1 * *, and dispatch with span and rebuild	Plan the month dispatched, or the last 3 months. Merge the week packs, and the day packs of the weeks that cross the month's edges, into m.	matrix
compact-yearly.yml	50 5 1 1 *, and dispatch with span and rebuild	Plan the year dispatched, or the last 2 years. Merge month packs into y.	matrix
rev-monthly.yml	30 6 * * *	Compare THC's release listing with the last month built. Build when a new month exists.	2 to 6 h
ref-refresh.yml	13 5 * * *	Refresh the PSL and the CT log list. Pin by sha256.	1 min
watchdog.yml	41 7 * * *	Staleness, schedules, anomaly issues	2 min
status.yml	47 * * * *	Start overdue workflows. Read the Worker's status and the sources' upstreams, and raise or clear status-down and worker-down. It publishes no file and pushes no branch: the Worker computes the status from the root.	2 min
ci.yml	push, pull request	The gate	5 min
worker.yml	push to main that changes the Worker	The Worker's tests. It never deploys.	3 min

Total runner time is 100 to 420 minutes a day [E] (design B §10: 105 min; design A §5.2: 6 to 7 h). It is free on a public repository. It shares the organisation's 20 concurrent jobs with mirror-archlinux and every other repository.

The live ingest (DOM-024). READ .github/workflows/ingest-live.yml, whose header names each job. Every run waits in the group ingest-live, with queue: max and no cancel.

Job	Does	Grants
plan	Runs on main alone. scripts/ingest/plan.mjs reads config/lanes.json and the registry, and writes one leg for each lane that holds an adapter: with sources, the lanes' adapters that hold one of those sources; with from and to, a backfill window, which needs sources. When no lane holds an adapter, only plan and chain run, and the run ends green.	contents: read
ingest	One leg a lane, kafka, tiles and feeds, on main alone, at most 3 at once, with fail-fast: false. Runs scripts/ingest/run.mjs for each adapter of its lane: reads the adapter's cursor and the list the root pins with no credential, verifies every segment before it pushes any, pushes the raw lane objects and then the segments to obs, and writes a cursor proposal. A failed adapter fails the leg after the other adapters of the lane ran. Uploads the proposals as the artifact proposals-<lane>. Never writes state.	contents: read, packages: write
push	After the legs, also after a failed one, on main alone, in the group root-writer. Advances the cursors (scripts/state/advance.mjs), builds the closed hours (scripts/state/hours.mjs: each feed-h-HOUR, then the move of state:ledger-delta), and last moves state:latest (scripts/state/move.mjs). Its output is every pack that the moved state owes.	contents: read, packages: write
promote	Only calls root-promote.yml with the owed packs as additions and partial: true, when the job push succeeded and the state owes a pack. The packs that keep the root's live set are promoted, and each other one prints waiting and stays owed (§4.7 rule 2). Holds no group (ruling 18).	contents: read, packages: write
chain	Needs no job, on main alone: dispatches the successor at the start of the run and sleeps to the next boundary (§5.2).	actions: write, contents: read

A failed source fails its own leg. The other legs, push and promote still run, and promote what exists. Owed segments whose spans leave a hole, such as a day source's last day and a stream's first window hours later, hold back no other pack: each prints waiting, the state keeps it owed, and every later run adds it again, until a pack covers the hole or the item expires after 7 days. An hour pack waits (hours.mjs exit 3) when a source of the hour is behind its clock: the hours before it are built, and the run stays green. With no ledger-filter- tag in state, no hour pack is built: hour.mjs needs a filter, and the first filter comes from compact-daily.yml. No job reads a Cloudflare credential. issues: write joins the legs with the issue script of DOM-034, for anomaly-<source> and visibility-<package>.

⚠ What is not done or not measured:

• No adapter exists until DOM-026, and every lane of config/lanes.json is empty, so a run plans no leg. The first run of ingest-live.yml on GitHub, run 41142944289 (schedule, 2026-10-30T23:11:29Z), came before config/clock.json: its job chain printed chain off, plan passed, and ingest, push and promote were skipped.
• No run under the clock actions-chain is measured, and no successor has been dispatched on GitHub.
• That GitHub runs the job-level group root-writer of the job push inside the workflow-level group ingest-live as described is assumed, as compact-daily.yml assumes it for compact. That actions/download-artifact with a pattern that matches no artifact ends green is assumed too.
• The steps of the four jobs that have steps run under bash in scripts/workflows/ingest-live.test.mjs, with doubles of the scripts they call. scripts/workflows/ingest-live-scripts.test.mjs runs the steps of plan, ingest and push over the real scripts, with the step "Move root:latest" of root-promote.yml (partial true) and the two ledger steps of the job push of compact-daily.yml, against the GHCR double of scripts/root/fixture.mjs and the fixture adapters of scripts/ingest/fixtures/: three runs, each reading what the run before it pushed. In the second and third runs both owed days of the file fixture print waiting across the hole before the stream's first window, the promotion ends unchanged, and the state still owes them. No run of that test promotes some owed packs while others wait; the tests of scripts/root/promote.test.mjs whose names start promote: --partial do.

The compaction workflows (DOM-023). The four files run the same jobs (READ .github/workflows/compact-daily.yml and compact-weekly.yml, whose headers name each job).

Job	Does	Grants
plan	Runs on main alone. Reads root:latest once with no credential, and pins its digest and its PSL. Runs scripts/compact/plan.mjs for each dataset and each unit. Holds a ready unit of domains or feed whose inputs were built under more than one PSL list (in compact-daily.yml, a domains day with the feed of the day): it prints held and a warning, and gives the unit no leg. Writes the matrices: one leg per range, at most 256 legs in a run.	contents: read
count, compact	Run the two stages of scripts/compact/range.mjs, at most 6 jobs at once. Read the inputs from GHCR with no credential. Push nothing.	contents: read
day (daily)	For each day whose domains unit is ready, or live with a feed unit to build: reads the live ledger of the package state with no credential, then runs scripts/compact/day.mjs --day DAY on the pinned root. A state that reads and holds no ledger-filter- tag is the first ledger. A newest filter with no ledger-delta beside it fails the leg: the first job push of ingest-live.yml that builds an hour writes the delta (scripts/state/hours.mjs), and no such job runs before an adapter exists (DOM-026). A state that does not read with no credential fails the leg, unless the run was dispatched with first-ledger. Pushes nothing.	contents: read
assemble	One job a unit, on main alone. Builds the pack (scripts/compact/assemble.mjs), or takes the day pack that day.mjs wrote. A names or new-name unit whose inputs were built under several PSL lists takes the list the root pins (--psl). Verifies the local files with scripts/verify-pack.mjs --require dataset,length,count,sha256,psl,sources,precision, and time for a feed or new-name pack that holds records. Pushes the pack with scripts/ghcr/push.mjs under its write-once tag. Writes the replacement, each replaced pack named in its own package.	contents: read, packages: write
push (daily)	Runs after assemble, in the group root-writer, on main alone. Pushes a day's ledger-filter-YYYYMMDD to state, written once, when the day's domains pack verified and pushed in the run or was live before it. Then moves state:ledger-delta under the guard (scripts/state/move.mjs ledger-delta --onto): it reads the live delta inside root-writer and rebases it onto the newest filter it pushed, so the hours that ingest-live.yml added to the delta after the day leg read it stay in it. A refusal keeps the live delta and fails the job.	contents: read, packages: write
promote	Only calls root-promote.yml with the replacements and additions of the units that pushed or were found present. Holds no group (ruling 18).	contents: read, packages: write

A scheduled run plans the last units of its tier that have closed, so a unit of a dropped run is taken by a later run. A unit that is not whole waits, and the run stays green. On the empty corpus every unit waits or is skipped. MEASURED on the operator's machine, each plan step of the tree run alone under bash in its scheduled form, with the real scripts, against root:latest sha256:bd69a39c1810eaa1c3b21dfa9de4c097a806a2372cb21ad648e9ef34a80eaeac (root_seq 1, no pack, pinned PSL 1f96bae40adfc22ca2474d7c0099bbcfcb25aefd3e2547b5c2147c0dd2e3ebd9). The step of compact-weekly.yml, 2026-10-30T02:45:46Z to 02:46:39Z, printed 16 wait and 4 skip. The step of compact-daily.yml, 02:46:47Z to 02:47:47Z, printed 21 wait. The step of compact-monthly.yml, 02:47:47Z to 02:48:23Z, printed 12 wait. The step of compact-yearly.yml, 02:48:23Z to 02:48:51Z, printed 8 wait. Each then printed nothing to compact: no unit is ready and exited 0.

Late sources. A day waits only on a source of its own packs whose live packs end before the day ends. planTier cannot see a source that has no pack in the day. So a day whose every hour pack comes from one lane and whose THC segment is not promoted yet is ready at 01:35, while ingest-thc.yml retries until 06:20, and the run builds the day's packs without THC. MEASURED 2026-10-30T04:02:48Z with planTier called by node on a synthetic root of the 24 hour packs of one day of names from ct-le alone, which hours of packages/root/test/helpers.ts builds: the day is ready, with 24 inputs. A rule of expected sources, under which a day waits on THC until the last retry of ingest-thc.yml, is an open duty of DOM-029. No code holds it.

A THC segment promoted after its day pack stays beside that pack in names and domains until the week, or the month for a week that crosses the first of a month, merges it. It never enters feed: the feed day pack is built once, by the join, and the week merges feed day packs. MEASURED 2026-10-30T02:44:50Z with planTier called by node on synthetic roots, whose packs pack of packages/root/test/helpers.ts builds: the day with its names day pack and a later THC segment is present, and the week takes the segment as an input. The month of October 2026 takes a later THC segment of 2026-10-27, a day of a week that crosses the month's end, as an input too. The workflow cannot build a live day again: planTier gives a live unit present, so the plan step keeps no unit for rebuild to take.

PSL lists (ruling 22). The keys of names and new-name do not depend on the PSL, so a unit of them whose inputs were built under several lists takes the list the root pins. The keys of domains and feed do depend on it, so assemble.mjs and day.mjs refuse their inputs built under two lists with exit 1 (CompactError psl): the recompute under one list is not built. So the plan step of the weekly, monthly and yearly workflows holds a ready unit of domains or feed whose inputs carry more than one PSL hash: it prints held TAG and a warning, gives the unit no leg, and exits 0 when no other unit faulted. The plan step of compact-daily.yml holds such a domains day, and the feed of the day with it, since day.mjs builds the feed day pack by the join under the one list of the domains inputs. Every unit above a held unit waits, as a week waits on a missing day pack. MEASURED 2026-10-30T02:44:51Z: planTier called by node on 7 synthetic day packs of names, three built under one list and four under another, gives ready with two lists in psl.

The four workflows share one concurrency group, compact, with queue: max and no cancel. So a monthly run plans against the root that the weekly run of the same morning promoted, and never replaces day packs that a week pack has replaced.

⚠ What is not done or not measured:

• scripts/compact/day.mjs writes ledger/delta.dlbd, and no workflow pushes it: the job push rebases the live delta instead.
• The recompute of domains and feed under one PSL list is an open duty of DOM-023. Until it is built, a move of the pinned list holds each domains and feed unit whose inputs mix two lists, and every unit above it (§4.6, §5.4).
• scripts/ghcr/push.mjs builds the manifest it pushes from the same layers and annotations. So the pushed digest is not the digest that assemble.mjs prints.
• scripts/ghcr/push.mjs reads each pushed pack back with no credential and exits 4 when it does not read, which fails the leg (§4.7 rule 7). Nothing raises the issue visibility-<package> until DOM-034.
• A year pack is read from the disk of one runner, and scripts/ghcr/push.mjs pushes all its layers in one copy of at most 600 s an attempt, 3 attempts in all (ATTEMPT_S of scripts/ghcr/rate.mjs; READ the header of scripts/ghcr/push.mjs). At the 63.5 to 121.4 MB/s that DOM-006 measured (§4.3), 600 s moves about 38 to 73 GB. The first yearly run, on 2027-01-01, builds nothing: the year 2026 of each dataset that has a corpus starts before its corpus_since (§4.5, "Before corpus_since"), and a dataset with no pack, such as new-name until DOM-041 builds it, has no corpus_since and waits on its month packs. The first year pack is 2027's, built on 2028-01-01. A year of THC names alone is about 72 to 84 GB at 6 to 7 GB a month [E] (design B section 6.4, in the Premise of DOM-023), above what the slowest push measured moves in one attempt. Whether a later attempt skips the blobs an earlier one uploaded is not measured.
• measure-compact.yml measured one range of about 0.58 GB of input from local files: about 7.8 minutes and under 300 MiB (§4.5, "The scripts"). At that rate a year of THC names, 72 to 84 GB [E], is about 125 to 145 ranges, 16 to 19 runner-hours, about 3 hours at a max-parallel of 6 [E]. The read from GHCR by range and the time copy are not measured.
• The guards of the four workflows' shell run under bash in scripts/compact/workflows.test.mjs, with doubles for node and oras. No compaction workflow has run on GitHub.

5.4 Idempotence, cursors and dedupe

• Segment identity is (source, window, cursor_before, cursor_after). A re-run with the same inputs writes the same bytes, because records are sorted and compression is fixed. The tag exists, with that digest or, from another run, with a manifest of the same sha256 whose run and fetched differ: the push is a no-op and names the digest the tag holds, so the proposal names what a reader of the tag finds.
• Cursors live in state:latest (state.json, docs/format.md section 17), one for each adapter, keyed by the adapter's directory name, since one run() of a module returns one cursor and one directory may hold several sources (§5.6): the Kafka offset of each topic, the next tile index of each log with the tree size at the last read, the last THC day, and the ETags. The lanes of one run (Kafka, tiles, feeds) run in parallel, as the legs of the job ingest of ingest-live.yml. A leg pushes its segments and raw lane objects, then writes one cursor proposal for each adapter run: the cursor it read, the cursor it reached, and the tags it pushed with their digests. It never writes state. One job, push, in the concurrency group root-writer (§4.7 rule 3), advances the cursors with scripts/state/advance.mjs. It keeps a proposal only if each tag it names holds the digest it names, and its before is the live cursor. It drops any other with a warning, and two proposals of one adapter both. It builds the closed hours with scripts/state/hours.mjs. Last in the job, before the job promote, it pushes state once with scripts/state/move.mjs, and only when the live state is the one it read. If the state push fails, or a proposal is dropped, the next run reads the old cursor, regenerates the same segment, skips the existing tag, and advances.
• The guard. Every move of state:latest and state:ledger-delta reads the tag with the job's token and compares its digest with the one the run read. It then pushes the new object under its write-once run tag (§4.1), reads the tag again, and only then moves it with oras tag. A digest that changed refuses the move, exit 1: the live object is kept, and the step warns and fails its job. The tag is then read back with no credential.
• Owed packs. state.json lists in owed each segment and hour pack that the moved state is past and the root does not hold yet. The job promote adds every owed pack, the earlier runs' among them, so a promotion that fails after the state moved loses no segment: the next run adds it again. It adds them with partial: true, so an owed pack that the root's live-set rule refuses waits, still owed, and the others are promoted (§4.7 rule 2). Before a run advances, it settles the list against the live root: an owed pack that the root names (held), an owed hour pack that a live feed pack of a higher tier spans (covered), one whose digest the root withdrew (withdrawn), and one owed for more than 7 days (expired) leave the list. A quarantined segment is never owed (§4.7 rule 2).
• A backfill is a dispatch with sources, from and to. run.mjs calls the adapter's backfill export for that window in place of run, and the proposal keeps the live cursor before and after. So a backfill moves no cursor, and its segments are owed and promoted.
• Dedupe is a property of the merge: minimum of first, maximum of last, union of sources. Order and retries do not change the result.
• Within a source, a name appears once per segment. A 5-minute Kafka window held 194,303 messages and 239,171 distinct names (MEASURED 2026-10-27, design A M8). Tiles hold about 1.3 names an entry (design A M10). Segment size follows distinct names.
• Reference data is pinned by sha256. A pack keeps the PSL it was built under, in its dev.pkgforge.domains.psl, and no code recomputes suffix, reg and site under a new list. The plan steps of the compaction workflows hold a ready unit whose keys depend on the list and whose inputs carry more than one PSL hash, and exit 0 when no other unit faulted (§5.3, "PSL lists"). The recompute under one list is an open duty of DOM-023.

5.5 Backfill

Source	Reach
THC daily	54 days from the Hugging Face bucket IP-THC-ORG/Daily-Cerstream-Dumps: 35,765,972,194 B (MEASURED 2026-10-27T08:25:12Z). The days 2026-10-04 and 2026-10-09 are missing. A file's path is <year>/<month without a leading zero>/<date>.txt.gz: 2026/9/ holds September only and October is 2026/10/, so the adapter takes each path from the listing. DOM-029 runs a matrix of one day each and records the gaps.
THC monthly	August and September 2026 (DOM-042)
Kafka certstream_domains	about 25 to 27 hours (design A M6, design B M4)
Kafka NRD topics	about 11 to 13 days
Names tiles	any time since a shard began, by binary search on entry timestamps. 21 probes found a 5-minute window (design A M10). The first backfill is 7 days.
ISC	5 dated files
WhoisDS, cenk-nrd, hagezi-nrd	the window each file carries

The corpus horizon is the earliest backfilled day, 2026-09-01. Answers carry confidence: warming for 90 days after it.

5.6 The adapter contract

An adapter is a directory adapters/<dir>/. The ingest entries create the tree, one directory each (AGENTS.md section 6). packages/adapter (DOM-016) holds the contract. A directory may hold several sources: DOM-026 puts three in adapters/oi-nrd/. The directory's name keeps the id grammar and need not be a source id. The directory holds:

1. source.json: { "format": "domains.source.v1", "sources": [...] }, a list of one or more sources. Each source holds exactly id, bit, family, datasets, precision, kind, basis, cadence, band, urls, retention, parser and failure_modes (SOURCE_FIELDS of packages/adapter/src/validate.ts). validateSourceFile refuses any other field. It refuses fetched, status and through with the code runtime-field: a run writes them, or a reader derives them (§3.5).
2. A module with run(cursor, ctx), which returns {segments, cursor, stats, errors}.
3. Fixtures of recorded answers and a test that runs the module against them.

Rules.

• An adapter writes segments. It never writes the root, state or another source's tag.
• A segment is deterministic and passes the verifier (§5.8) before it counts.
• A failed adapter fails alone. It never blocks another source or the root.
• The registry row in sources.json and the row in docs/sources.md are held equal by a test.
• To add a source: allocate a bit, write the adapter, add its failure-mode row. Nothing downstream changes.

5.7 A VPS later

A VPS ingester is one more adapter, for example vps-ct. It runs the same encoder and verifier and pushes segments. GHCR accepts only classic tokens for pushes from outside Actions, and such a token can write every package of its owner (READ mirror@669ac72:docs/design.md, "Running outside GitHub Actions"). This project holds no token (ruling 6), so the VPS does not push. It is a pull model:

1. The VPS writes its segments to its own disk and serves them over HTTPS: a static directory with an index.json of tags, sizes and sha256, and no credential in either direction.
2. An Actions job fetches the segments it does not hold, checks every sha256, runs the verifier (§5.8), and pushes them with its own GITHUB_TOKEN, as it pushes a segment of any other source.
3. The first push of each package comes from a workflow, so each package is linked to the repository. The promote job lists the pulled segments in the root, and validates them again before they count.

The VPS therefore holds nothing that can write GHCR, and a compromised VPS can offer a segment that the verifier refuses and cannot move the root or state. Nothing in this project holds a secret (entry DOM-057).

5.8 Verification before promotion

This step adapts the mirror's verify step. A job fails when any check fails.

1. Decode the whole pack. Check bytewise key order, uniqueness, charset, label length, no upper case, and that counts match the annotations. The verifier (DOM-018) prints SKIP for a check with nothing to compare, so the job that writes a pack names the checks it needs with --require (with time for a pack that carries a time copy) before it calls the promotion, which does not check that it ran (section 4.7, rule 2).
2. Re-find 1,000 random keys with the reader module the Worker uses. One TypeScript module runs in Node and in workerd.
3. Compare volume with the median of the last 7 days that were not quarantined. A THC day under 40% of the median is quarantined: it is pushed with the annotation dev.pkgforge.domains.quarantine set to 1, excluded from new computations, and an issue anomaly-<source> opens. Its names are still served in names, domains and rev (ruling 23, built by DOM-029). THC's volume fell to about a third in early October with no stated reason: 843 to 986 MB a day in September, 984 and 957 MB on 10-01 and 10-02, 445 MB on 10-03, and 77 to 340 MB a day from 10-05 (MEASURED 2026-10-27T09:34:52Z, curl -sS 'https://huggingface.co/api/buckets/IP-THC-ORG/Daily-Cerstream-Dumps/tree?recursive=true', sizes summed per file). A baseline that never moves would quarantine every day after a lasting drop: all 21 observed days from 10-05 sit below the floor of 379 MB (40% of 947 MB). So the rule has a step change: 3 consecutive observed days below the floor reset the baseline to the median of those 3 days, and the third is accepted. A missing day neither counts nor breaks the run. The backfill, run in date order, quarantines 10-05 and 10-06, accepts 10-07 with a baseline of 294 MB, and quarantines 10-15 (77 MB against a median of 276 MB). checkBand (entry DOM-016) holds the rule.
4. Check the PSL hash and the source's precision.
5. For the hourly filter's provisional verdicts, the daily job recomputes first_seen exactly against the domains table and writes the corrections. Hot answers carry provisional: true until then.

A verification uses fixed charset and length rules. It uses no content regular expressions.

5.9 Workflow rules

Each rule is checked by DOM-014.

• Every workflow starts with permissions: {}. Each job grants the least it needs. A job that writes GHCR has packages: write. A job that opens an issue has issues: write. A job that dispatches or enables a workflow has actions: write. Only the jobs named in one table of the workflow check may hold one of the three (scripts/check-workflows.mjs, created by entry DOM-014, which lists the names). No job has contents: write: no workflow pushes a branch, so there is no status branch, and status is read through the Worker.
• A job that holds packages: write runs on main alone: its if: holds github.ref == 'refs/heads/main' as one operand of a top-level &&, with no top-level || (rule 3 of scripts/check-workflows.mjs, onMainOnly). A job that only calls root-promote.yml is exempt, since the called file's own job holds the grant and that test.
• Every uses: pins a commit. The runner is ubuntu-24.04.
• No workflow holds a Cloudflare credential. No workflow calls the packages DELETE API.
• Concurrency groups: root-writer (queue max, no cancel) for any job that moves the root or state; ingest-live, the one group of ingest-live.yml (queue max, no cancel, §5.3); ingest-<source>; compact, the one group of the four compaction workflows (§5.3); ci- with a pull request's number, or the run id for any other event (READ .github/workflows/ci.yml).
• The workflows start each other when a run is overdue, with dispatch-overdue, and re-enable disabled workflows, with keep-schedules-enabled. Both are copied from mirror-archlinux. GitHub disables schedules after 60 days with no activity (READ events documentation).
• A watchdog raises schedules-stopped when a schedule stops, and stale-<source> when a source is late.

6. THC's data, mirrored

THC publishes two things. Both are mirrored. No route calls THC's live API.

• The daily CT dumps. cs2.ip.thc.org lists <date>.txt.gz from 2026-10-01. A file appears about 00:04 UTC the next day. The 2026-10-26 file was 319,917,410 B. The Hugging Face bucket holds the same bytes plus September. The adapter thc-ct reads cs2 first, then the bucket (DOM-029).
• The monthly database. THC publishes it at each month's end. September's was released 2026-10-04 with 6.36 billion records (READ https://ip.thc.org/docs/bulk-data-access, 2026-10-27; the page answers only a request with a browser User-Agent: curl with its default one got 404, MEASURED 2026-10-27T09:38:43Z). Gzipped Parquet needs full decompression before its footer can be read. Gzipped CSV streams on a runner. This project mirrors the CSV.

Where the monthly files live (READ the page above, MEASURED 2026-10-27).

Host	URL	Notes
Hugging Face (fast)	https://huggingface.co/buckets/IP-THC-ORG/Datasets/resolve/2026/10/rdns-sept.csv.gz?download=true	302 to a signed CDN URL. 206 on Range. First 1 MiB in 2.96 s.
dns.team-teso.net (slow)	https://dns.team-teso.net/2026/rdns-sept.csv.gz	A 1 MB Range read did not answer within 120 s at 2026-10-27T08:03Z.

Measured columns, from the first 1,048,576 compressed bytes of the September CSV (MEASURED 2026-10-27T08:21:59Z, curl -r 0-1048575 then gzip -dc):

• 72,834 lines, every line with exactly two comma-separated fields: ip,name. There is no header.
• Both IPv4 and IPv6 addresses appear. The file is not sorted by address. One address carries several names (41,973 distinct addresses in 72,834 lines).
• No line has a third column, a record type or a CNAME target. Names that contain the text cname are host names that resolve to an address.
• The sample's average is 41.2 bytes a line. The page's 252 GB for 6.36 billion records gives 39.6 bytes a record, so the whole file has the same two-column shape.

So the monthly database holds reverse IP to names and holds no CNAME targets. DOM-010 confirmed it over the whole stream: all 6,364,111,723 records of the September file have exactly two fields and an address first (MEASURED 2026-10-28T15:53Z on a runner, run 40714223764, docs/research/measure/DOM-010/result.json). Until a mirrored source holds CNAME targets, GET /v1/cnames/{target} answers 501 with the reason unsupported-by-source (§7.2). DOM-046 takes the first of them that DOM-058 (OpenINTEL's measurement topic) or DOM-047 (its zone Parquet) shows to carry CNAME records (ruling 10, TODO/RULES.md §5).

The mirror keeps the CSV as fetched: chunks of at most 4 GiB in package obs, each with a sha256, written once per month. A job builds the rev packs from it (DOM-043): the keys are the address bytes, the values are the names. A query reads the newest month by default.

7. The API

One Worker serves domains.pkgforge.dev. The hostname is one level below the pkgforge.dev zone, so the zone's Universal certificate covers it. The Worker reads and never writes GHCR. Its name is domains-pkgforge-dev. Its code lands in apps/api/ with DOM-030.

7.1 Common rules

• Every route is GET unless it says otherwise. HEAD answers like GET with no body.
• JSON is the default. List routes accept format=ndjson|csv|txt.
• Every JSON answer has the shape below. Every error is application/problem+json with status, title, detail, reason and see.
• CORS: access-control-allow-origin: *. No credentials are ever allowed.

{
  "data": [
    {
      "domain": "example.com.np",
      "first_seen": "2026-10-27T04:12:09Z",
      "first_seen_precision": "second",
      "last_seen": "2026-10-27T05:30:00Z",
      "registered": { "basis": "nrd-ct+rdap", "confidence": 3, "sources": ["oi-nrd-confirmed"] },
      "sources": {
        "oi-nrd": { "first": "2026-10-27T04:12:09Z", "last": "2026-10-27T05:30:00Z" },
        "thc-ct": { "first": "2026-10-27", "last": "2026-10-27" }
      },
      "flags": ["provisional"]
    }
  ],
  "meta": {
    "as_of": "2026-10-27T05:50:00Z",
    "root": "sha256:<digest>",
    "corpus_since": "2026-09-01",
    "confidence": "warming",
    "sources_used": ["oi-nrd", "thc-ct"],
    "sources_failed": [{ "source": "ct-geomys", "since": "2026-10-27T04:00:00Z", "error": "429" }],
    "tiers_read": ["h-2026102705", "d-20261026", "w-2026w43"],
    "partial": false,
    "next_cursor": null
  }
}

A day-precision value prints as a date. A second-precision value prints as a time.

Provenance headers say what answered, and nothing the path did not use.

Header	Value
x-domains-root	the digest of the root the answer was computed from
x-domains-as-of	the root's as_of
x-domains-tiers-read	the number of packs read
x-domains-sources-used	the source ids that contributed
x-domains-sources-failed	the source ids that failed; absent when none
x-domains-cache	hit (Cache API), r2, ghcr or mixed
x-domains-body	derived, or verbatim for a dump served as stored
x-domains-degraded	1, only when an answer used stale data or lacked a source

7.2 Routes

THC parity.

Route	THC counterpart	Source
GET /v1/ip/{ip-or-cidr}	POST /lookup	rev packs. Rows {domain, apex_domain, tld}. No organisation, country or ASN.
GET /v1/subdomains/{domain}	POST /lookup/subdomains	names tiers. Rows {subdomain, last_seen_on, first_seen_on, sources}.
GET /v1/cnames/{target}	POST /lookup/cnames	none yet. Answers 501 unsupported-by-source with the list of mirrored sources it checked.
GET /v1/ip/..., /v1/subdomains/... with format=csv	GET /download, /subdomains/download	the same, streamed. THC caps a file at 50,000 rows. Ours caps at 5,000,000.
GET /v1/dumps/names/{date}.txt.gz	cs2.ip.thc.org/<date>.txt.gz	the dumps package. 302 to the signed blob URL. Sorted, lowercase, wildcards kept.
GET /v1/dumps/rdns/{month}	the monthly files	302 to the mirrored raw chunks' listing.

Our own routes.

Route	Purpose
GET /v1/new	New or newly registered registrable domains, or names
GET /v1/names	Names under a suffix or parent with first and last seen in a window
GET /v1/domain/{reg}	The dossier: first and last per source, evidence, recent names
GET /v1/name/{fqdn}	First and last per source
POST /v1/lookup	Up to 500 names or domains in one call
GET /v1/dumps, /v1/dumps/{dataset}/{date}	List and fetch bulk files
GET /v1/suffixes?since=	Counts per suffix for a window, from the daily summaries
GET /v1/sources, /v1/status, /v1/coverage	Source health, freshness, corpus_since and the cross-source table
GET /v1/index/latest	The root digest and the GHCR references, so a client can read with no Worker
GET /healthz	{ok, version, root_age_s}. 503 with ok: false when the root is older than 6 h.
GET /, /llms.txt, /llms-full.txt, /openapi.json, /SKILL.md	Generated documents (§7.7)

7.3 GET /v1/new

Parameter	Meaning
since	1h, 6h, 12h, 1d, Nd (N at most 31), or an ISO 8601 time. Default 1d.
until	An ISO 8601 time. Default: now.
kind	registrable (default) or fqdn
suffix	A public suffix: np, com.np, co.uk, xyz. One contiguous key range. For kind=registrable it keeps the domains whose suffix is exactly it, one label under it: suffix=np keeps example.np and leaves example.com.np, whose suffix is com.np. For kind=fqdn it keeps every name under it, and DOM-041 narrows that by the PSL.
under	Any parent: a TLD, a suffix, a registrable domain or a name. One contiguous key range.
basis	any (default), seen, claimed, registered, zone, rdap, nrd-list, ct. any and seen keep every row: each row of the feed is newly seen. claimed keeps a row with a basis of tier 1 or 2 (zone, rdap, nrd-list, nrd-ct+rdap), and registered one with a registry claim (zone, rdap, nrd-ct+rdap). zone, rdap and nrd-list keep a row with that basis, and ct one with ct-first.
source	One or more source ids. A row is kept when at least one of them carries it.
min_confidence	1 to 10. It keeps the rows that a registry claim backs at least that strongly: a zone or rdap basis, which outranks every OpenINTEL level, or an nrd-ct+rdap basis whose stored OpenINTEL level is 1 to the value, level 1 being the closest to the RDAP date. Every other row is left out, a row of no stated level (0) among them.
suspect	exclude (default), include, only
provisional	include (default) or exclude
order	first_seen.desc (default) or first_seen.asc
limit, cursor	limit defaults to 1,000 and is at most 10,000. The cursor is opaque.

A window is read from the live set. A window of 1 h to 12 h reads segments and hour packs. A window of N days reads day packs, and week packs where a week is whole. meta.tiers_read lists them. A pack of tier d, w, m or y is read through its time copy, from the bound of the window in order, and one without a time copy is refused; a segment or an hour pack may have none and is read from its key copy and a sort (docs/format.md section 14.4). A page of 1,000 rows with no filter read 1 or 2 blocks of a day pack of 500,000 synthetic rows. A filter that drops most of the rows read costs their blocks too: source=oi-nrd on a window from 12:00, ascending, read the 49 blocks of the day's day-precision rows (§7.4) before its first row, 50 in all (MEASURED 2026-10-29T14:42:47Z, docs/format.md section 14.6). The block budget of §7.6 bounds such a page.

7.4 Time

since and until are UTC. first_seen is compared at its own precision: a day-precision row is inside a window when its day overlaps the window. meta.as_of is the root's time. Each source's through time is in /v1/sources. A window that overlaps a gap of a source says so in meta.

7.5 Caching

Object	Cache time
Blocks, indexes and pack manifests by digest	1 year, immutable
root:latest	30 s
Hour and segment packs of the current day	60 s
Closed day, week and month packs	1 year, immutable
Signed blob URLs	until the URL's se minus 120 s, at most 600 s
Query answers	the minimum of the cache times of their inputs

The Cache API is per data centre. Single-flight per key inside one isolate. One query fans out over at most 6 concurrent subrequests.

7.6 Limits, and no gating

No route asks for a key, an account or a quota. No route limits a caller's rate. The limits below bound the work of one request. They are design values. DOM-054 measures them and may change them.

Limit	Value
Window by the API	at most 31 days. Longer ranges use dumps.
Rows in a JSON page	at most 10,000
Rows in a streamed csv, ndjson or txt answer	at most 5,000,000
Batch lookup	at most 500 keys
Packs read in one request	at most 60
Blocks read in one request	at most 2,000. A broader query returns partial: true and a cursor.

Cloudflare's own abuse controls still apply. The Worker never refuses a caller on the caller's identity.

7.7 Root documents

Every service of this family answers its root with generated documents. The Worker does the same. A registry of routes lives in one typed file beside the handlers. A generator renders llms.txt, llms-full.txt, openapi.json and SKILL.md from that registry at build time. The gate regenerates them and fails on a difference. A test fails when a handled route has no registry entry (entry DOM-059). / answers a browser with a small HTML index and any other client with llms.txt.

8. Redundancy and failure

No source is complete. On the same 5 minutes of 2026-10-26, THC's day dump held 39.8% of the names in OpenINTEL's Kafka window. That stream carried no Let's Encrypt, Geomys or IPng log. THC held 97.7% and 98.1% of the names of two Let's Encrypt shards (MEASURED 2026-10-27, design A M8, M9). ISC's new domains of 2026-10-26 were in THC's 2026-10-26 dump at 25.5% and in its 2026-10-25 dump at 54.8% (design B M6). The service answers from the union and publishes the coverage table daily.

What fails	Detection	The service then	Recovery
THC cs2	no file by 06:00 UTC	reads the Hugging Face bucket. If both fail, thc-ct is late; other sources answer.	The hourly retry. Backfill when it returns. Issue stale-thc-ct after 36 h.
THC volume shifts	under 40% of the 7-day median of accepted days	quarantines the day; it does not feed new. Three such days in a row reset the baseline (§5.8).	the operator decides; issue anomaly-thc-ct
OpenINTEL Kafka	no new offsets in 30 min	marks oi-* late; other sources answer	replay from the stored offset (about 11 to 13 days for the NRD topics, 25 to 27 hours for certstream_domains)
A names-tile log	404 or 429	other logs continue; the cursor holds	per-run availability probe; the log list refreshes daily
ISC, WhoisDS	no new file for 6 h	marks the source late	5 days of ISC are re-readable
GHCR slow or down	a subrequest over 10 s	serves Cache API and R2; x-domains-degraded: 1	retry with the 429 rule
R2 down	R2 errors	reads GHCR	none
Cloudflare down	no answer	clients read GHCR through root:latest (§11)	none
Actions stops	the root's age; watchdog	the Worker serves the last root and shows its age. The Worker cannot dispatch a workflow.	dispatch-overdue, keep-schedules-enabled, issue schedules-stopped
A bad segment	the verifier	the segment is excluded and the tier is rebuilt without it	The root's previous digest allows rollback.

A merge never removes a record from a promoted tier. A rollback is a new root that names older packs.

9. Accuracy

• CT-derived rows say seen, never registered.
• Feed rows keep the feed's own date. WhoisDS's file dated D matches ISC's day D-1 at 79.8% for one pair and 66.2% for another, and 0% for the same date (MEASURED 2026-10-27, design B M6). The adapter dates a WhoisDS file as D-1.
• An oi-nrd event with a certificate older than 7 days is suspect. So is a domain whose RDAP creation date is before 2026-01-01 (ruling 19).
• Name-level newness is censored. A daily THC dump holds the names of certificates issued that day. 86.6% of 2026-10-26's names were absent from the union of four earlier days (design B M2). Against every earlier day of THC's 54 daily dumps, each day from 2026-10-16 to 10-26 still carries 25% to 31% names never seen before (MEASURED 2026-10-28, DOM-009's novelty curve, docs/research/measure/DOM-009/result.json). A name recurs only at a renewal. Until the corpus holds one validity period (about 200 days [E]), new for names carries censored: true. Domain-level newness from feeds with their own history is not censored.
• corroborated needs two families. Lists derived from one another (hagezi from Stamus) count as one family.
• Machine-generated spaces hold many names: 25 registrable domains hold 35% of THC's names (design B M2). They stay in the index. A query pages through them.
• CT sees only domains that obtained a certificate. A domain registered and never certified is invisible to every CT-derived source.
• A validation sample runs daily. DOM-050 checks 200 random claimed_new rows against RDAP for the TLDs that have it. 233 of 1,437 root-zone TLDs have no RDAP, including 175 of 248 ccTLDs (READ docs/research/web-pkgforge-2026-10-27/landscape.md). .np has no RDAP and no open zone, so .com.np rows are T2 or T3.
• The coverage table (DOM-035) publishes the share of each source's names found in each other source's same-day set.

10. Cost

Item	Per day	Basis
Actions	$0. 100 to 420 runner-minutes [E]	public repositories are free
GHCR storage and egress	$0. 1.2 to 2 GB net new of derived data [E] and about 12 GB of raw lane (§4.8)	"currently free", with at least a month's notice before a change (READ GitHub's packages billing page via design A §2.2). Raw data is kept always (ruling 5).
Workers	$0 over the $5 monthly plan at 100,000 requests a day [E]	10 million requests and 30 million CPU-ms are included
R2	at most 5 GB at $0.015 per GB-month, about $0.075 a month	the cap
Inbound network on runners	THC 0.32 GB, Kafka 10.7 GB on the wire (MEASURED, DOM-004; 28.4 GB once decoded), tiles 1.4 GB	not billed

11. The consumer contract

web.pkgforge.dev's domain and ip services read this project. They read three things: the names index, the new-domain feed and the reverse lookups. The contract has four parts. It states nothing about the consumer's internals.

1. HTTP. The routes of §7.2 and the parameters of §7.3 are the contract. docs/consumer.md is created by DOM-060. It lists them with an example for each. A route is stable once listed in stable.json (created by DOM-060). A removal or rename needs a new path version.

2. The root. ghcr.io/pkgforge-dev/domains.pkgforge.dev/root:latest is readable anonymously. GET /v1/index/latest answers its digest. root.json (§4.6) names every live pack by digest. A client that reads the root and the packs needs no Worker.

3. The pack format. DPK1 (§4.4). The format document, created by DOM-017, states every byte. DOM-017 publishes test vectors. DOM-060 publishes a reference reader module that runs in Node and in workerd.

4. Versioning.

• Within domains.root.v1, DPK1 and /v1, changes are additive. A reader ignores unknown fields and unknown layers.
• A breaking change gets a new format id and a new path (/v2). The old one stays served for at least 90 days.
• A reader checks format and schema and refuses a value it does not know.
• One logical question reads one root digest. The answer carries it in x-domains-root.

The consumer must honour meta.confidence, censored, provisional, suspect and sources_failed. It must not read a day-precision value as a time of day. It must not assume that last_seen means the domain is live.

12. Cloud resources

Each resource is created by the entry named. The undo of each goes in docs/resources.md, in the same change.

Resource	Created by	Undo
Probe Worker domains-probe (on workers.dev) and R2 bucket domains-probe, temporary	DOM-061	wrangler delete; delete the bucket. DOM-056 deletes both when DOM-002, DOM-005, DOM-007 and DOM-008 have closed.
Drill Worker domains-drill (on workers.dev) and R2 bucket domains-drill, temporary	DOM-053	wrangler delete; delete the bucket. The drill deletes both when it ends.
Worker domains-pkgforge-dev and custom domain domains.pkgforge.dev on the pkgforge.dev zone	DOM-030	wrangler delete; remove the custom domain
R2 bucket domains-pkgforge-cache with a lifecycle rule: the one cache bucket, at most 5 GB	DOM-031	delete the bucket
GHCR packages of §4.1	the first push of a workflow	none. Nothing is ever deleted (ruling 5).

The pkgforge.dev zone's own Cloudflare account hosts the Worker, because a custom domain needs the zone in the same account. mirror-archlinux deploys its Worker the same way, and its worker/wrangler.toml names that account in account_id (READ at commit 669ac725f8081445c164113f2ed092533de112e2). The operator deploys with a local wrangler login. CI deploys nothing.

13. Which entry builds what

Part	Entries
Measurements	DOM-001 to DOM-012, and DOM-061 (the probe Worker they share)
Foundation: workspace, policy, names, adapter contract	DOM-013 to DOM-016
Pack format, verification, GHCR, root, ledger, feed, compaction	DOM-017 to DOM-023
The new-domain feed: adapters, Worker, routes, watchdogs	DOM-024 to DOM-035
Names	DOM-036 to DOM-041
THC mirror and reverse routes	DOM-042 to DOM-046
Authority sources and validation	DOM-047 to DOM-052
Hardening	DOM-053 to DOM-058
Root documents and the consumer contract	DOM-059, DOM-060